diff --git a/skills/assemblyai/SKILL.md b/skills/assemblyai/SKILL.md
index 81d7948..63306e2 100644
--- a/skills/assemblyai/SKILL.md
+++ b/skills/assemblyai/SKILL.md
@@ -43,36 +43,35 @@ Authorization: YOUR_API_KEY
 
 ## Speech-to-Text Models
 
+**`universal-3-5-pro` is the model to use across the board.** It is GA for **both** realtime/streaming and pre-recorded/async transcription and is the default on each. Use it everywhere by default; drop to `universal-2` only for cost or for languages outside Universal-3.5 Pro's 18. `universal-3-pro` (async) and `u3-rt-pro` (streaming) have both been **superseded by `universal-3-5-pro`** and removed from their model lists/enums.
+
 ### Pre-Recorded
 
 | Model | Languages | Best For |
 |-------|-----------|----------|
-| **Universal-3.5 Pro** | 18 (auto-falls back to Universal-2 for the other 99) | Latest flagship: best accuracy, native code-switching across its 18 languages, contextual `prompt`, keyterms up to 1,000 words |
-| **Universal-3 Pro** | 6 (en, es, de, fr, pt, it) | Highest accuracy, promptable transcription, keyterms up to 1,000 words |
-| **Universal-2** | 99 | Broadest language coverage, keyterms up to 200 words |
-
-Use `speech_models` as a priority list with fallback: `["universal-3-pro", "universal-2"]` (the default when omitted).
+| **Universal-3.5 Pro** (recommended default) | 18 (auto-falls back to Universal-2 for the other 99) | Flagship: best accuracy, fastest, native code-switching across its 18 languages, contextual `prompt`, keyterms up to 1,000 words |
+| **Universal-2** | 99 | Cost-effective; broadest language coverage; keyterms up to 200 words; fallback for languages outside Universal-3.5 Pro's 18 |
 
-**Universal-3.5 Pro is also available for pre-recorded/async transcription**, not just streaming. Opt in with `speech_models: ["universal-3-5-pro"]` on `POST /v2/transcript`. It supports contextual `prompt` (a plain-language *description* of the audio — domain/scenario/full detail) and `keyterms_prompt` (up to 1,000 terms), does native code-switching, and **auto-falls back to Universal-2** for languages outside its 18. Note: the formal OpenAPI `speech_models` enum still lists only `universal-3-pro`/`universal-2`, but the async API accepts `universal-3-5-pro`.
+On `POST /v2/transcript`, `speech_models` is a priority list with fallback and **defaults to `["universal-3-5-pro", "universal-2"]`** when omitted — Universal-3.5 Pro handles its 18 languages and automatically falls back to Universal-2 for the rest. The async `speech_models` enum now accepts only `universal-3-5-pro` and `universal-2` (`universal-3-pro` has been removed, superseded by `universal-3-5-pro`). Universal-3.5 Pro supports contextual `prompt` (a plain-language *description* of the audio) and `keyterms_prompt` (up to 1,000 terms). Check the `speech_model_used` response field to see which model actually ran.
 
 ### Streaming
 
 | Model | Languages | Best For |
 |-------|-----------|----------|
-| **universal-3-5-pro** | 18 | **Recommended default for new realtime/streaming code** — next-gen flagship: more languages, improved prompting + conversational context |
-| **u3-rt-pro** | 6 | Universal-3 Pro Streaming — punctuation-based turn detection, promptable, `mode` accuracy/latency tradeoff |
+| **universal-3-5-pro** | 18 | **Recommended default for new realtime/streaming code** — next-gen flagship: more languages, native code-switching, improved prompting + conversational context |
 | **universal-streaming-english** | 1 (English) | Voice agents, ~300ms latency |
 | **universal-streaming-multilingual** | 6 | Per-utterance language detection |
+| **u3-rt-pro** | 6 | **Legacy** — Universal-3 Pro Streaming, **removed July 2026** from the model picker and streaming spec `speech_model` enum. Superseded by `universal-3-5-pro`. Still seen in older code |
 | **whisper-rt** | 99+ | **Legacy** — removed from the public model picker (June 2026) and the streaming spec enums, but still functional via `speech_model: whisper-rt` for broadest streaming language coverage, auto-detect only |
 
-For realtime/streaming STT, use `speech_model: "universal-3-5-pro"` by default. The raw API parameter is optional and defaults to `universal-3-5-pro`; set it explicitly when pinning behavior or using an SDK that requires the field. The `mode` connection param (universal-3-5-pro / u3-rt-pro) trades off accuracy vs latency: `min_latency`, `balanced` (default), `max_accuracy`.
+For realtime/streaming STT, use `speech_model: "universal-3-5-pro"` by default. The raw API parameter is optional and defaults to `universal-3-5-pro`; set it explicitly when pinning behavior or using an SDK that requires the field. The streaming spec `speech_model` enum now lists only `universal-3-5-pro`, `universal-streaming-english`, and `universal-streaming-multilingual`. The `mode` connection param (universal-3-5-pro only) trades off accuracy vs latency: `min_latency`, `balanced` (default), `max_accuracy`.
 
 ### Medical Mode (Add-On)
 
 `domain: "medical-v1"` enables Medical Mode — an add-on that improves accuracy for medical terminology (medications, procedures, conditions, dosages). Works with both pre-recorded and streaming models.
 
-- **Pre-recorded:** Universal-3 Pro (`domain: "medical-v1"` in request body), Universal-2
-- **Streaming:** universal-3-5-pro, u3-rt-pro, universal-streaming-english, universal-streaming-multilingual
+- **Pre-recorded:** Universal-3.5 Pro (`domain: "medical-v1"` in request body), Universal-2
+- **Streaming:** universal-3-5-pro, universal-streaming-english, universal-streaming-multilingual
 - **Supported languages:** English, Spanish, German, French (4 languages only)
 - Billed as a separate add-on. If used with an unsupported language, the API ignores `domain` and returns a warning — transcript still completes and you are NOT charged for Medical Mode.
 
@@ -80,7 +79,7 @@ For realtime/streaming STT, use `speech_model: "universal-3-5-pro"` by default.
 
 `prompt` and `keyterms_prompt` are **complementary** — use either, or both together. Neither changes the output format, and both work the same way for **streaming and async** (`POST /v2/transcript`). Transcription behavior (verbatim, punctuation, formatting) is built in and managed by AssemblyAI.
 
-- **`prompt`** (string, ~1500 chars max): a plain-language **description of the audio** — its domain, scenario, or full details. It carries *context*, **not instructions** — formatting/behavioral commands (punctuation rules, "transcribe verbatim", "don't…") are not supported and are ignored. The model stays grounded in the audio: irrelevant or only-partially-applicable context won't make it insert words that weren't spoken, so you can safely send the same description on every session/segment.
+- **`prompt`** (string; length cap varies by API — streaming ≤1750 chars, async up to ~1,500 words, Sync STT ≤4096 chars): a plain-language **description of the audio** — its domain, scenario, or full details. It carries *context*, **not instructions** — formatting/behavioral commands (punctuation rules, "transcribe verbatim", "don't…") are not supported and are ignored. The model stays grounded in the audio: irrelevant or only-partially-applicable context won't make it insert words that weren't spoken, so you can safely send the same description on every session/segment.
 - **`keyterms_prompt`** (string[]): an explicit list of names/brands/domain terms to boost. Streaming: up to **100** terms, ≤50 chars each. Async: up to **1,000** terms.
 
 **Contextual prompt — three levels of specificity** (use the least specific that covers your case):
@@ -129,11 +128,11 @@ See `references/llm-gateway.md` for models, tool calling, structured outputs, an
 | Gotcha | Details |
 |--------|---------|
 | `prompt` + `keyterms_prompt` | **Complementary** for Universal-3.5 Pro — use either or both together. `prompt` is a contextual *description* of the audio; `keyterms_prompt` is an explicit term list. Neither changes output formatting |
-| `summarization` / `auto_chapters` | **Deprecated.** Use LLM Gateway instead (transcribe → send text to LLM) |
+| `summarization` / `auto_chapters` (top-level params) | **Deprecated.** Use **Speech Understanding** `summarization` (chaptered summary w/ timestamps + headlines) or `action_items` — `speech_understanding.request.<feature>` on `POST /v2/transcript` — or the LLM Gateway for fully custom prompts |
 | PII redaction scope | Only redacts words in `text` — other feature outputs (entities, summaries) may still expose sensitive data |
 | Upload key scoping | Files uploaded with one API key project cannot be transcribed with a different project's key |
 | Structured outputs | Supported by OpenAI, Gemini, Claude 4.5+, Qwen, and Kimi — Claude 3.x does NOT support `json_schema` structured outputs |
-| U3 Pro-family turn detection | `universal-3-5-pro` and `u3-rt-pro` use punctuation (`.` `?` `!`), NOT confidence thresholds — `end_of_turn_confidence_threshold` has no effect |
+| universal-3-5-pro turn detection | `universal-3-5-pro` uses punctuation (`.` `?` `!`), NOT confidence thresholds — `end_of_turn_confidence_threshold` has no effect (it applies only to Universal Streaming English/Multilingual) |
 | `prompt` is context, not instructions | Universal-3.5 Pro's `prompt` *describes* the audio (domain/scenario/details). Formatting or behavioral commands (punctuation rules, "transcribe verbatim", negative directives like "don't…") are **not supported** and are ignored — transcription behavior is managed internally |
 | PII audio redaction method | `override_audio_redaction_method: "silence"` replaces PII with silence instead of default beep |
 | Language detection | Requires minimum 15 seconds of spoken audio for reliable results |
@@ -154,9 +153,14 @@ See `references/llm-gateway.md` for models, tool calling, structured outputs, an
 | LLM Gateway `tool_calls` location | `tool_calls` lives at `choices[i].message.tool_calls` (under `message`), NOT at `choices[i].tool_calls` (under `choice`). `content` is `null` when only tool_calls are present |
 | LLM Gateway `finish_reason` is provider-native | Don't branch tool-calling loops on `finish_reason == "tool_calls"` — the Gateway passes the provider's value through, so **Claude returns `tool_use`/`end_turn`** (OpenAI returns `tool_calls`/`stop`). Detect a tool call by the **presence of `message.tool_calls`**, not by `finish_reason` |
 | Transcript `metadata.warnings` | The `Transcript` response now includes an optional `metadata` object. When present, `metadata.warnings` is an array of `{message}` objects describing issues processed during transcription (e.g. Medical Mode skipped due to unsupported language). `metadata` is omitted entirely when there is nothing to report |
-| U3 Pro-family streaming context carryover | On by default — the model carries prior finalized turns forward as context (per-session, ~3 entries, ~1500 chars). Pass your agent's spoken reply via `agent_context` (connection-time query param to seed an opening greeting, or mid-stream via `UpdateConfiguration`) so the model knows the question the user is answering. Use with `universal-3-5-pro` or `u3-rt-pro` |
+| universal-3-5-pro streaming context carryover | On by default — the model carries prior finalized turns forward as context (per-session, up to `previous_context_n_turns` entries, server default `5`, range 0–100). Pass your agent's spoken reply via `agent_context` (connection-time query param to seed an opening greeting, or mid-stream via `UpdateConfiguration`; ≤1750 chars per value) so the model knows the question the user is answering. `universal-3-5-pro` only |
 | Streaming diarization revised labels | With `speaker_labels` enabled, a single `SpeakerRevision` message is emitted right before `Termination` (after you send `Terminate`), containing a `revisions` array of only the turns whose speaker labels changed (matched by `turn_order`). Text and word timestamps never change — only speaker assignments. Adds ~400ms latency at session close. Use it for the final, highest-quality attribution |
 | LLM Gateway `model_region: "global"` | Optional request field (only accepted value `"global"`) routes to the provider's global, non-region endpoints for lower cost. Live for Anthropic Claude now; Google Gemini 3 series coming soon. Omit for default in-region processing. **Effective July 1, 2026, in-region LLM Gateway requests cost 10% more** (provider pass-through, no AssemblyAI upcharge) — global routing keeps current pricing |
+| Streaming language selection | The connection param is **`language_codes`** (plural, a **list** e.g. `["en","es"]`) — there is no singular `language_code` *input* param. It steers output toward those languages while still allowing native code-switching. `universal-3-5-pro` only. (`language_code` appears only as an *output* field when `language_detection: true`.) |
+| Streaming Opus input | `encoding` accepts `opus` (raw packets — one per binary WS message) and `ogg_opus` (Ogg stream — arbitrary chunks; ffmpeg/gstreamer/MediaRecorder output), in addition to `pcm_s16le`/`pcm_mulaw`. For Opus, `sample_rate` is ignored (stream is self-describing) |
+| Voice Agent HTTP tool `headers` | Server-side HTTP tools take `headers` as a **list of `{name, value}` entries** (since June 2026), NOT a `{name: value}` dict. Reads return `{name, last_set_at}` only — values are write-only and never returned (not masked as `"***"`) |
+| Voice Agent BYO LLM (`llm`) | Set `llm` on a **stored agent** (REST create/update, not `session.update`) to run on your own OpenAI-compatible endpoint: `[{base_url, model, api_key}]` — one entry only, must stream, `api_key` write-only. `"llm": []` reverts to the managed model. Point `base_url` at the LLM Gateway to use a frontier model on your AssemblyAI account |
+| Voice Agent webhooks | Subscribe to Voice Agent events via `POST https://agents.assemblyai.com/v1/webhook-subscriptions` (events: `session.started`/`session.completed`/`call.connected`/`call.ended`/`call.failed`; scope with `agent_id` or omit for account-wide). See `references/api-reference.md` |
 
 ## Common Mistakes
 
@@ -164,7 +168,7 @@ See `references/llm-gateway.md` for models, tool calling, structured outputs, an
 |---------|------------|
 | `Authorization: Bearer KEY` | `Authorization: KEY` (no Bearer prefix) — BUT the Voice Agent API (`agents.assemblyai.com`) uses `Authorization: Bearer KEY` |
 | Using LeMUR API | **Deprecated.** Use LLM Gateway instead |
-| Using `summarization` or `auto_chapters` | **Deprecated.** Use LLM Gateway instead (transcribe then summarize via LLM) |
+| Using top-level `summarization` or `auto_chapters` | **Deprecated.** Use Speech Understanding `summarization`/`action_items` (`speech_understanding.request.<feature>`) or the LLM Gateway |
 | LeMUR `transcript_ids` with LLM Gateway | Pass transcript text in messages, not IDs |
 | `anthropic/claude-...` model IDs | No provider prefix: `claude-sonnet-4-5-20250929` not `anthropic/claude-sonnet-4-5-20250929` |
 | `claude-opus-4-20250514` / `claude-sonnet-4-20250514` on LLM Gateway | **Removed June 2026.** Use Claude Opus 4.5/4.6/4.7 or Claude Sonnet 4.5/4.6 |
@@ -172,9 +176,9 @@ See `references/llm-gateway.md` for models, tool calling, structured outputs, an
 | Using Java/Go/C# SDKs | **Discontinued.** Use Python, JS/TS, Ruby, or raw API |
 | `word_boost` on the async REST API | Use `keyterms_prompt` instead. **Exception:** the Sync STT API *does* use `word_boost` (in its `config` part) — that's its documented keyterms param |
 | Hardcoding v2 streaming URL | v3 (`/v3/ws`) is current; v2 still works but is legacy |
-| Assuming streaming defaults to `u3-rt-pro` | Streaming `speech_model` defaults to `universal-3-5-pro` at the raw API layer. Set a different model only when you intentionally need legacy behavior, cost tradeoffs, or broader language coverage |
+| Using `speech_model=u3-rt-pro` for streaming | **Removed July 2026** from the model picker and streaming spec enum — superseded by `universal-3-5-pro` (the streaming default). Set a different model only for cost tradeoffs (`universal-streaming-english`/`-multilingual`) |
 | Python SDK rejects `universal-3-5-pro` | Upgrade to `assemblyai>=0.64.21` for Streaming v3 SDK support. Older SDKs such as `0.64.4` validate `speech_model` against an enum that omits `universal-3-5-pro` |
-| `aai.SpeechModel.universal_3_pro` in Python SDK | Use raw strings: `"universal-3-pro"`, `"universal-2"` — these enum aliases don't exist in the SDK |
+| `aai.SpeechModel.universal_3_5_pro` in Python SDK | Use raw strings: `"universal-3-5-pro"`, `"universal-2"` — these enum aliases don't exist in the SDK |
 | S2S `session.update` without `"session"` key | Must wrap config: `{"type":"session.update","session":{...}}` |
 | S2S tool schema using `{"function":{...}}` nesting | S2S tools are flat: `{"type":"function","name":"...","description":"...","parameters":{...}}` |
 | Voice Agent S2S URL | Correct URL: `wss://agents.assemblyai.com/v1/ws` — not `/v1/voice` (renamed April 2026), `/v1/realtime` (older), or `speech-to-speech.us.assemblyai.com` (very old) |
@@ -197,9 +201,9 @@ Read the relevant reference file based on what the user needs:
 | `references/streaming.md` | Real-time/streaming STT, v3 protocol, temp tokens, error codes |
 | `references/voice-agents.md` | Voice agent integrations: LiveKit, Pipecat, turn detection, latency optimization |
 | `references/llm-gateway.md` | Applying LLMs to transcripts, tool calling, available models |
-| `references/speech-understanding.md` | Translation, speaker identification, custom formatting |
+| `references/speech-understanding.md` | Translation, speaker identification, custom formatting, summarization, action items |
 | `references/audio-intelligence.md` | PII redaction, diarization, summarization, sentiment, chapters |
-| `references/api-reference.md` | Full parameter list, export endpoints, webhooks, upload, PII policies, Sync STT API, Voice Agents REST API (stored agents) |
+| `references/api-reference.md` | Full parameter list, export endpoints, webhooks, upload, PII policies, Sync STT API, Voice Agents REST API (stored agents, HTTP-tool headers, BYO LLM, webhook subscriptions) |
 
 ## API Spec Source of Truth
 
diff --git a/skills/assemblyai/references/api-reference.md b/skills/assemblyai/references/api-reference.md
index ede86f4..b928ae3 100644
--- a/skills/assemblyai/references/api-reference.md
+++ b/skills/assemblyai/references/api-reference.md
@@ -32,7 +32,7 @@ Submit an audio file for transcription. Send a JSON body with the parameters bel
 | Parameter | Type | Description |
 |---|---|---|
 | `audio_url` | string | **Required.** URL of the audio file to transcribe. Can be a public URL or an `upload_url` from the upload endpoint. |
-| `speech_models` | array | **Optional** (as of June 2026). Priority-ordered list of speech models to use (e.g., `["universal-3-5-pro", "universal-2"]`). First model is used if supported; falls back to next. If omitted, defaults to `["universal-3-pro", "universal-2"]`. Universal-3.5 Pro is accepted here (`["universal-3-5-pro"]`). |
+| `speech_models` | array | **Optional.** Priority-ordered list of speech models. First supported model is used; falls back to the next. **If omitted, defaults to `["universal-3-5-pro", "universal-2"]`.** The enum now accepts only `universal-3-5-pro` and `universal-2` — `universal-3-pro` was removed (superseded by `universal-3-5-pro`). The `speech_model_used` response field reports which model actually ran. |
 | `prompt` | string | For Universal-3.5 Pro, a contextual *description* of the audio (domain → scenario → full detail), **not** formatting/behavioral instructions (those are ignored). **Complementary with `keyterms_prompt`** — both can be set together. |
 | `keyterms_prompt` | array | List of key terms/phrases (strings) to boost recognition accuracy — up to **1000** terms for Universal-3.5 Pro, **200** for Universal-2, max **6 words per phrase**. **Complementary with `prompt`** — both can be set together. |
 | `language_code` | string | Language code (e.g., `"en_us"`, `"es"`, `"fr"`). Defaults to `"en_us"`. |
@@ -42,11 +42,11 @@ Submit an audio file for transcription. Send a JSON body with the parameters bel
 | `speaker_labels` | boolean | Enable speaker diarization. Default `false`. |
 | `sentiment_analysis` | boolean | Enable sentiment analysis on each sentence. Default `false`. |
 | `entity_detection` | boolean | Enable entity detection. Default `false`. |
-| `auto_chapters` | boolean | **Deprecated.** Enable auto chapters. Use LLM Gateway instead. |
+| `auto_chapters` | boolean | **Deprecated.** Use Speech Understanding `summarization` (chaptered summary; see `speech-understanding.md`) or the LLM Gateway instead. |
 | `iab_categories` | boolean | Enable IAB content category detection. Default `false`. |
 | `content_safety` | boolean | Enable content safety detection. Default `false`. |
 | `content_safety_confidence` | integer | Minimum confidence threshold (25-100) for content safety labels. |
-| `summarization` | boolean | **Deprecated.** Enable summarization. Use LLM Gateway instead. |
+| `summarization` | boolean | **Deprecated.** Use Speech Understanding `summarization`/`action_items` (see `speech-understanding.md`) or the LLM Gateway instead. |
 | `summary_model` | string | **Deprecated.** Model for summarization. |
 | `summary_type` | string | **Deprecated.** Type of summary. |
 | `redact_pii` | boolean | Enable PII redaction. Default `false`. |
@@ -68,10 +68,10 @@ Submit an audio file for transcription. Send a JSON body with the parameters bel
 | `speech_understanding` | object | Enable Speech Understanding inline. Features nest under `speech_understanding.request` (the `request` wrapper is required): `translation`, `speaker_identification`, and/or `custom_formatting`. See `speech-understanding.md`. |
 | `speakers_expected` | integer | Hint for number of speakers (diarization). Deprecated in favor of `speaker_options`. |
 | `speaker_options` | object | Diarization options: `min_speakers_expected` (int, default 1), `max_speakers_expected` (int). |
-| `temperature` | float | 0–1. Controls randomness. Universal-3 Pro only. |
-| `domain` | string | Domain-specific model variant. `"medical-v1"` enables Medical Mode (EN, ES, DE, FR). Supported on Universal-3 Pro and Universal-2. |
-| `remove_audio_tags` | string | Remove inline annotations from the transcript. `"all"` removes all (audio event markers and speaker cues); `"speaker"` removes only speaker cues while keeping other annotations. Universal-3 Pro only. |
-| `language_codes` | array | List of language codes for code-switching (must include `"en"`). Universal-3 Pro only. |
+| `temperature` | float | 0–1. Controls randomness. Universal-3.5 Pro only. |
+| `domain` | string | Domain-specific model variant. `"medical-v1"` enables Medical Mode (EN, ES, DE, FR). Supported on Universal-3.5 Pro and Universal-2. |
+| `remove_audio_tags` | string | Remove inline annotations from the transcript. `"all"` removes all (audio event markers and speaker cues); `"speaker"` removes only speaker cues while keeping other annotations. Universal-3.5 Pro only. |
+| `language_codes` | array | List of language codes for code-switching (must include `"en"`). Universal-3.5 Pro only. |
 | `audio_start_from` | integer | Start transcription from this time offset, in **milliseconds**. |
 | `audio_end_at` | integer | End transcription at this time offset, in **milliseconds**. |
 | `speech_threshold` | float | Confidence threshold (0-1) for filtering low-confidence speech. Requires at least **30 seconds** of audio. |
@@ -105,7 +105,7 @@ The transcript response may include an optional `metadata` object with additiona
   "metadata": {
     "domain_used": null,
     "warnings": [
-      { "message": "'ja' is not supported in universal-3-pro — transcription is handled by universal-2. To silence this warning, set speech_models: [\"universal-3-pro\", \"universal-2\"]." }
+      { "message": "'ur' is not supported in universal-3-5-pro — transcription is handled by universal-2. To silence this warning, set speech_models: [\"universal-3-5-pro\", \"universal-2\"]." }
     ]
   }
 }
@@ -479,11 +479,11 @@ A REST API for creating **reusable** voice agents. An agent stores its `system_p
 |---------------|-------------|
 | `POST /v1/agents` | Create an agent. Returns `201` with the full record including a generated `id`. |
 | `GET /v1/agents` | List agents (lightweight records, no tools/prompt), newest first. |
-| `GET /v1/agents/{agent_id}` | Retrieve one agent. Tool header **values** are masked as `"***"`. |
+| `GET /v1/agents/{agent_id}` | Retrieve one agent. Read responses return tool-header **names** and `last_set_at` only — header **values are never returned** (write-only, encrypted at rest; they are *not* masked as `"***"`). Likewise a connected LLM (`llm`) comes back as `base_url` + `model` only, never `api_key`. |
 | `PUT /v1/agents/{agent_id}` | Update an agent. Every field optional — send only what you want to change. |
 | `DELETE /v1/agents/{agent_id}` | Delete an agent. Returns `204`, no body. |
 
-**Create body** (`application/json`): required `name`, `system_prompt`, `voice`; optional `greeting`, `input`, `output`, `tools`. Note `voice` is a **top-level** field here (in the WebSocket `session.update` it lives under `output.voice`). `greeting` is spoken straight to TTS on connect — omit it to have the agent listen first.
+**Create body** (`application/json`): required `name`, `system_prompt`, `voice`; optional `greeting`, `input`, `output`, `tools`, `llm`. Note `voice` is a **top-level** field here (in the WebSocket `session.update` it lives under `output.voice`). `greeting` is spoken straight to TTS on connect — omit it to have the agent listen first.
 
 ```bash
 curl -X POST https://agents.assemblyai.com/v1/agents \
@@ -491,3 +491,43 @@ curl -X POST https://agents.assemblyai.com/v1/agents \
   -H 'Content-Type: application/json' \
   -d '{"name":"Support Bot","system_prompt":"You are a concise support agent.","voice":"ivy","greeting":"Hi, how can I help?"}'
 ```
+
+### Server-side HTTP tools (`headers` shape)
+
+A tool with an `http` block (`url`, `http_method`, `headers`) is executed **server-side** — AssemblyAI makes the request on your behalf. As of June 2026 `headers` is a **list of `{name, value}` entries**, NOT a `{name: value}` dict:
+
+```json
+{ "http": { "url": "https://api.example.com/orders", "http_method": "POST",
+  "headers": [ { "name": "Authorization", "value": "Bearer xyz" } ] } }
+```
+
+- **Write (create/update):** each entry is `{ name, value?, remove? }`. Provide `value` to set/rotate; provide name only to keep the stored value unchanged (round-tripping on update); `remove: true` deletes it. Sending both `value` and `remove: true` is rejected.
+- **Read (get/list):** each entry is `{ name, last_set_at }` — values are never returned.
+- Server-side constraints: `https` + public hosts only (private/loopback/CGNAT blocked); redirects not followed; response body capped at 8 KiB; per-call timeout from `timeout_seconds`.
+
+### Bring your own LLM (`llm`)
+
+Point an agent at your own **OpenAI-compatible** chat-completions endpoint instead of AssemblyAI's managed model. Set the `llm` field (a list; **only one entry accepted today**) on create/update:
+
+```json
+{ "llm": [ { "base_url": "https://api.openai.com/v1", "model": "gpt-4o-mini", "api_key": "sk-..." } ] }
+```
+
+- `base_url` (required, `https` + public host — the agent calls `POST {base_url}/chat/completions`), `model` (required), `api_key` (required, **write-only**, never returned).
+- Must support **streamed** chat completions. Send `"llm": []` to switch back to the managed model.
+- To use a frontier model without your own provider account, point `base_url` at the **LLM Gateway** (`https://llm-gateway.assemblyai.com/v1`, EU: `https://llm-gateway.eu.assemblyai.com/v1`) and pass your AssemblyAI key as `api_key`.
+
+## 18. Voice Agent Webhooks API
+
+A REST API (base URL `https://agents.assemblyai.com`, same auth) to subscribe URLs to Voice Agent lifecycle events.
+
+| Method & Path | Description |
+|---------------|-------------|
+| `POST /v1/webhook-subscriptions` | Subscribe a URL to one or more events. Scope to one agent with `agent_id`, or omit for account-wide events. Returns the created subscription. |
+| `GET /v1/webhook-subscriptions` | List subscriptions (paginated). |
+| `GET /v1/webhook-subscriptions/{subscription_id}` | Retrieve one subscription. |
+| `PUT /v1/webhook-subscriptions/{subscription_id}` | Update `url`, `events`, `secret`, or `enabled`. Sending a new `secret` rotates it and increments `secret_version`. |
+| `DELETE /v1/webhook-subscriptions/{subscription_id}` | Delete a subscription. Returns `204`. |
+
+- **Events:** `session.started`, `session.completed`, `call.connected`, `call.ended`, `call.failed`.
+- **Body fields:** `url` (`https` + public host, required), `events` (array, required), `secret` (32–256 printable ASCII chars, no whitespace — signing secret, **write-only**, only its `secret_version` is returned), `agent_id` (optional; omit for account-wide).
diff --git a/skills/assemblyai/references/audio-intelligence.md b/skills/assemblyai/references/audio-intelligence.md
index dfc342d..1254753 100644
--- a/skills/assemblyai/references/audio-intelligence.md
+++ b/skills/assemblyai/references/audio-intelligence.md
@@ -77,7 +77,7 @@ Legacy parameters (still functional but not recommended for new code):
 
 - Enable with `domain: "medical-v1"` in the request body
 - Improves accuracy for medical terminology: medications, procedures, conditions, dosages
-- Supported models: Universal-3 Pro, Universal-2 (pre-recorded); all streaming models
+- Supported models: Universal-3.5 Pro, Universal-2 (pre-recorded); all streaming models
 - Supported languages: English, Spanish, German, French only
 - Billed as a separate add-on
 - If used with an unsupported language, the API ignores `domain` and returns a warning — no charge is applied
diff --git a/skills/assemblyai/references/python-sdk.md b/skills/assemblyai/references/python-sdk.md
index 31d95a1..03655f8 100644
--- a/skills/assemblyai/references/python-sdk.md
+++ b/skills/assemblyai/references/python-sdk.md
@@ -47,7 +47,7 @@ Use `speech_models` to specify a preferred model with automatic fallback:
 
 ```python
 config = aai.TranscriptionConfig(
-    speech_models=["universal-3-pro", "universal-2"]
+    speech_models=["universal-3-5-pro", "universal-2"]
 )
 
 transcriber = aai.Transcriber(config=config)
diff --git a/skills/assemblyai/references/speech-understanding.md b/skills/assemblyai/references/speech-understanding.md
index 8f9ad5b..33aa4cf 100644
--- a/skills/assemblyai/references/speech-understanding.md
+++ b/skills/assemblyai/references/speech-understanding.md
@@ -2,7 +2,9 @@
 
 ## Overview
 
-Speech Understanding provides post-transcription intelligence: **Translation, Speaker Identification, and Custom Formatting**. (Entity Detection, Sentiment Analysis, Key Phrases, and Topic Detection are *not* part of this object — they remain classic boolean transcript params; see `audio-intelligence.md`.)
+Speech Understanding provides post-transcription intelligence: **Translation, Speaker Identification, Custom Formatting, Summarization, and Action Items**. (Entity Detection, Sentiment Analysis, Key Phrases, and Topic Detection are *not* part of this object — they remain classic boolean transcript params; see `audio-intelligence.md`.)
+
+> **Summarization vs the deprecated `summarization`/`auto_chapters` params:** the top-level `summarization` and `auto_chapters` boolean params on `POST /v2/transcript` are **deprecated**. The modern replacements are the Speech Understanding `summarization` (chaptered summary with timestamps + headlines) and `action_items` features below — or the LLM Gateway for fully custom prompts.
 
 Two ways to run it:
 
@@ -29,7 +31,7 @@ Auth header: `Authorization: API_KEY` (no `Bearer` prefix).
 
 Translates the transcript into one or more target languages.
 
-- **Models:** Universal-3 Pro and Universal-2 only. **Regions:** US & EU. Supports 100+ language codes.
+- **Models:** Universal-3.5 Pro and Universal-2 only. **Regions:** US & EU. Supports 100+ language codes.
 
 ### Parameters (`speech_understanding.request.translation`)
 
@@ -181,6 +183,93 @@ The format params are **strings (format patterns), not booleans**:
 
 `speech_understanding.response.custom_formatting` contains `formatted_text`, `formatted_utterances` (only when `format_utterances: true`), and a `mapping` object (original → formatted).
 
+## Summarization
+
+Generates a chaptered summary of the transcript — each chapter has timestamps, a `text` summary, and a `headline`. Replaces the deprecated top-level `summarization`/`auto_chapters` params. **Open beta.** Models: `universal-3-5-pro`, `universal-2`. Regions: US & EU.
+
+### Parameters (`speech_understanding.request.summarization`)
+
+- `summary_type` (string, **required**): `"paragraph"` (longer, more detailed) or `"bullets"` (short, concise).
+- `effort` (string, default `"low"`): `"low"` or `"medium"`. `medium` spends more processing power for higher-quality summaries — use it for high-stakes meetings, multilingual audio, or very long audio (≈1.5h+). Default `low` is sufficient for most cases.
+
+### Example
+
+```json
+{
+  "audio_url": "https://example.com/audio.mp3",
+  "speech_understanding": {
+    "request": {
+      "summarization": { "summary_type": "paragraph", "effort": "low" }
+    }
+  }
+}
+```
+
+### Response
+
+`speech_understanding.response.summarization` → `{ status, summary_type, effort, summary: [{ start, end, text, headline }] }`:
+
+```json
+{
+  "speech_understanding": {
+    "response": {
+      "summarization": {
+        "status": "success",
+        "summary_type": "paragraph",
+        "effort": "low",
+        "summary": [
+          { "start": 240, "end": 37100, "text": "Smoke from hundreds of Canadian wildfires...", "headline": "Smoke from Canadian Wildfires Affects US Air Quality" }
+        ]
+      }
+    }
+  }
+}
+```
+
+## Action Items
+
+Extracts action items from the transcript, each with the source `quote` and a `timestamp`. **Open beta.** Models: `universal-3-5-pro`, `universal-2`. Regions: US & EU.
+
+### Parameters (`speech_understanding.request.action_items`)
+
+- `include_decisions` (boolean, default `false`): when `true`, also captures decisions made in the conversation as action items.
+- `effort` (string, default `"low"`): `"low"` or `"medium"` (see Summarization for guidance on `medium`).
+
+Pass an empty object (`"action_items": {}`) to use defaults.
+
+### Example
+
+```json
+{
+  "audio_url": "https://example.com/audio.mp3",
+  "speech_understanding": {
+    "request": {
+      "action_items": { "include_decisions": true }
+    }
+  }
+}
+```
+
+### Response
+
+`speech_understanding.response.action_items` → `{ status, effort, items: [{ action_item, quote, timestamp }] }`:
+
+```json
+{
+  "speech_understanding": {
+    "response": {
+      "action_items": {
+        "status": "success",
+        "effort": "low",
+        "items": [
+          { "action_item": "Observe the expansion and contraction of the Arctic ice cap throughout the seasons.", "quote": "The Arctic ice cap ... expands in winter and contracts in summer.", "timestamp": 9520 }
+        ]
+      }
+    }
+  }
+}
+```
+
 ## Post-hoc via the Understanding Endpoint
 
 To run Speech Understanding on an **existing** transcript, POST to `/v1/understanding` with a `transcript_id` (same `speech_understanding.request.<feature>` structure):
diff --git a/skills/assemblyai/references/streaming.md b/skills/assemblyai/references/streaming.md
index 5358d52..e42e3c7 100644
--- a/skills/assemblyai/references/streaming.md
+++ b/skills/assemblyai/references/streaming.md
@@ -18,32 +18,32 @@ For new realtime/streaming code, use **`speech_model=universal-3-5-pro`** by def
 
 | Parameter | Description |
 |-----------|-------------|
-| `speech_model` | **Optional at the raw API layer; default `universal-3-5-pro`.** Other models: `u3-rt-pro`, `universal-streaming-english`, `universal-streaming-multilingual`. `whisper-rt` (99+ languages) is **legacy** — removed from the public model picker and the streaming spec enums (June 2026) but still functional via `speech_model=whisper-rt` |
-| `mode` | **universal-3-5-pro / u3-rt-pro only.** Accuracy/latency tradeoff: `min_latency` (fastest time-to-text), `balanced` (**default** — best for voice agents), or `max_accuracy` (highest accuracy, for scribes/post-call). Sets the per-mode defaults for `min_turn_silence`, `max_turn_silence`, `interruption_delay`, `continuous_partials`, and `vad_threshold`. Set at connection time and updatable mid-stream via `UpdateConfiguration`. |
+| `speech_model` | **Optional at the raw API layer; default `universal-3-5-pro`.** Other current models: `universal-streaming-english`, `universal-streaming-multilingual`. The streaming spec `speech_model` enum now lists only these three. Two models are **legacy** — removed from the model picker and the spec enum but still seen in older code: `u3-rt-pro` (Universal-3 Pro Streaming, removed July 2026, superseded by `universal-3-5-pro`) and `whisper-rt` (99+ languages, removed June 2026, still functional via `speech_model=whisper-rt` for broadest language coverage). New integrations should use `universal-3-5-pro`. |
+| `mode` | **universal-3-5-pro only.** Accuracy/latency tradeoff: `min_latency` (fastest time-to-text), `balanced` (**default** — best for voice agents), or `max_accuracy` (highest accuracy, for scribes/post-call). Sets the per-mode defaults for `min_turn_silence`, `max_turn_silence`, `interruption_delay`, `continuous_partials`, and `vad_threshold`. Set at connection time and updatable mid-stream via `UpdateConfiguration`. |
 | `sample_rate` | Audio sample rate in Hz (e.g., 16000) |
-| `encoding` | Audio encoding: `pcm_s16le` or `pcm_mulaw` |
-| `end_of_turn_confidence_threshold` | Confidence threshold for turn detection. Only affects Universal Streaming, not U3 Pro. **Officially deprecated** — tune `min_turn_silence`/`max_turn_silence` instead. |
+| `encoding` | Audio encoding, default `pcm_s16le`. Raw PCM: `pcm_s16le` or `pcm_mulaw`. **Opus (added June 2026):** `opus` (raw Opus packets — each binary WebSocket message must contain exactly one packet) or `ogg_opus` (Ogg-encapsulated Opus stream, as produced by ffmpeg/gstreamer/opusenc/browser MediaRecorder — binary messages can be arbitrary chunks). For both Opus encodings `sample_rate` is ignored (the stream is self-describing). |
+| `end_of_turn_confidence_threshold` | Confidence threshold for turn detection. Only affects Universal Streaming (English/Multilingual), not universal-3-5-pro. **Officially deprecated** — tune `min_turn_silence`/`max_turn_silence` instead. |
 | `format_turns` | Set to `true` to enable formatted final transcripts with punctuation, casing, and inverse text normalization (dates, times, phone numbers). Also activates turn-level keyterm boosting for Universal Streaming models. **Does NOT control digit rendering** — numerals (e.g. "22") are a model behavior, and lexical number output (e.g. "twenty-two") is not supported in streaming. |
-| `prompt` | **universal-3-5-pro / u3-rt-pro only.** Natural-language *context about the audio* (domain, topic, scenario, conversation details) — **NOT** behavioral/formatting instructions. The transcription instruction (verbatim behavior, punctuation, formatting) is built in and managed by AssemblyAI; formatting or behavioral commands placed in `prompt` are not supported. **Complementary with `keyterms_prompt`** — use either or both together. If omitted, a built-in default prompt optimized for turn detection is used automatically. Recommended: test with no prompt first, then add context only for domain vocabulary the model gets wrong. |
-| `keyterms_prompt` | JSON-encoded array of strings (up to 100 terms, max 50 chars each) to bias transcription (universal-3-5-pro, u3-rt-pro, and Universal Streaming). **Complementary with `prompt`** — both can be set together. When passing via URL query param, must be JSON.stringify'd: `keyterms_prompt=["term1","term2"]`. Costs additional $0.04/hr. |
+| `prompt` | **universal-3-5-pro only.** Max **1750 characters.** Natural-language *context about the audio* (domain, topic, scenario, conversation details) — **NOT** behavioral/formatting instructions. The transcription instruction (verbatim behavior, punctuation, formatting) is built in and managed by AssemblyAI; formatting or behavioral commands placed in `prompt` are not supported. **Complementary with `keyterms_prompt`** — use either or both together. If omitted, a built-in default prompt optimized for turn detection is used automatically. Recommended: test with no prompt first, then add context only for domain vocabulary the model gets wrong. |
+| `keyterms_prompt` | JSON-encoded array of strings (up to 100 terms, max 50 chars each) to bias transcription (universal-3-5-pro and Universal Streaming). **Complementary with `prompt`** — both can be set together. When passing via URL query param, must be JSON.stringify'd: `keyterms_prompt=["term1","term2"]`. Costs additional $0.04/hr. |
 | `inactivity_timeout` | Seconds of silence before session auto-closes |
 | `speaker_labels` | Enable diarization (`true`/`false`) |
-| `max_speakers` | Maximum number of speakers for diarization |
+| `max_speakers` | Integer 1–10. A **hard cap** (strict limit, not a hint) on speaker labels — once reached, additional speakers are merged into the closest existing label rather than given a new one. Give a little headroom above the expected count; setting it too high causes over-splitting. Only used when `speaker_labels` is enabled. |
 | `domain` | Set to `"medical-v1"` to enable Medical Mode (improves accuracy for medical terminology). Supported models: all streaming models. Supported languages: en, es, de, fr. |
 | `redact_pii` | Enable real-time PII redaction. Default `false`. Only applies to **final turns**. See Streaming PII Redaction below. |
 | `redact_pii_policies` | PII entity types to redact. Pass a comma-separated string (e.g. `person_name,phone_number`) over the raw WebSocket or an array via the SDK. Default: all. |
 | `redact_pii_sub` | Replacement scheme: `hash` (default — replaces with `#` chars) or `entity_name` (replaces with `[ENTITY_TYPE]`). |
 | `include_partial_turns` | Whether to include partial (non-final) turns. Defaults to `true` normally, but **`false` automatically** when `redact_pii` is `true` so unredacted text never reaches the client. |
 | `filter_profanity` | Filter profanity from transcripts (replaces with `***`). Default `false`. |
-| `interruption_delay` | **universal-3-5-pro / u3-rt-pro only.** Integer milliseconds (0–1000, default `500`). How soon the first partial is emitted — lower = faster TTFT and earlier barge-in but more false interruptions; higher = more confident interruptions but slower partials. The server adds a minimum of ~256–300ms on top (`interruption_delay: 0` → ~256–300ms effective, `500` → ~756–800ms effective). The LiveKit plugin keeps the API default of `500`. |
-| `continuous_partials` | **universal-3-5-pro / u3-rt-pro only.** Boolean — **default `true`** (changed June 2026; previously `false`). Now defaults to `true` on both the API directly and the LiveKit plugin. When `true`, emits additional partial transcripts approximately every ~3 seconds during long turns, each covering the full turn transcript so far. The first early partial (at 750ms / your `interruption_delay`) is unaffected. Set `false` if you only want silence-based partials. |
-| `agent_context` | **universal-3-5-pro / u3-rt-pro only.** String (≤~1500 chars). Your voice agent's most recent spoken reply (TTS text), used as context for the next user turn — see Context Carryover below. Set at connection time to seed an opening greeting, and/or update mid-stream via `UpdateConfiguration`. |
-| `previous_context_n_turns` | **universal-3-5-pro / u3-rt-pro only.** Integer (default `3`). Max number of prior conversation entries (finalized user transcripts plus any `agent_context` values) carried forward as context for each transcription. Set to `0` to disable automatic context carryover entirely. Most integrations leave this at the default — see Context Carryover below. |
-| `vad_threshold` | **universal-3-5-pro / u3-rt-pro only.** Float 0.0–1.0 (default `0.3`). Confidence threshold for classifying audio frames as silence. Increase in noisy environments to reduce false speech detection. |
-| `voice_focus` | **universal-3-5-pro / u3-rt-pro only.** Noise suppression that isolates the primary voice and suppresses background chatter, keyboard clicks, fan hum, and room echo before audio reaches the model. Set to `near-field` (headsets, handsets, close-talking mics) or `far-field` (conference rooms, laptop/drive-thru mics, distant capture). Omit to disable. Set as a connection parameter. |
-| `voice_focus_threshold` | **universal-3-5-pro / u3-rt-pro only.** Optional float `0.0`–`1.0` controlling how aggressively background audio is suppressed when `voice_focus` is set — higher = more aggressive. |
-| `language_code` | **universal-3-5-pro / u3-rt-pro only.** Optional ISO 639-1 code that biases the model toward a single language when you know the session is monolingual (improves language accuracy). Omit to keep default multilingual code-switching. `universal-3-5-pro` supports en, es, de, fr, pt, it, tr, nl, sv, no, da, fi, hi, vi, ar, he, ja, zh; `u3-rt-pro` supports en, es, fr, de, it, pt. |
-| `language_detection` | **universal-3-5-pro / u3-rt-pro only.** Boolean (default `false`). When `true`, each `Turn` message includes the detected `language_code` and `language_confidence`. U3 Pro-family models natively code-switch without this — use it only when you need the per-turn language reported. |
+| `interruption_delay` | **universal-3-5-pro only.** Integer milliseconds (0–1000). Default is **mode-dependent** (set by the `mode` preset), not a fixed value. How soon the first partial is emitted — lower = faster TTFT and earlier barge-in but more false interruptions; higher = more confident interruptions but slower partials. The server adds a **fixed 256ms** on top (`interruption_delay: 0` → 256ms effective, `500` → 756ms effective). |
+| `continuous_partials` | **universal-3-5-pro only.** Boolean — **default `true`** (changed June 2026; previously `false`). Now defaults to `true` on both the API directly and the LiveKit plugin. When `true`, emits additional partial transcripts approximately every ~3 seconds during long turns, each covering the full turn transcript so far. The first early partial (timed by `interruption_delay`) is unaffected. When `speaker_labels` is enabled, continuous partials are disabled by default. Set `false` if you only want silence-based partials. |
+| `agent_context` | **universal-3-5-pro only.** String (≤1750 chars per value). Your voice agent's most recent spoken reply (TTS text), used as context for the next user turn — see Context Carryover below. Set at connection time to seed an opening greeting, and/or update mid-stream via `UpdateConfiguration`. |
+| `previous_context_n_turns` | **universal-3-5-pro only.** Integer, range 0–100 (server default currently `5`). Max number of prior conversation entries (finalized user transcripts plus any `agent_context` values) carried forward as context for each transcription. Set to `0` to disable automatic context carryover entirely. Most integrations leave this unset — see Context Carryover below. |
+| `vad_threshold` | Float 0.0–1.0. Confidence threshold for classifying audio frames as silence — frames below this are considered silent. Increase in noisy environments to reduce false speech detection. Defaults: `0.2` on universal-3-5-pro (`0.5` when `speaker_labels` is enabled), `0.4` on Universal Streaming. |
+| `voice_focus` | **universal-3-5-pro only.** Noise suppression that isolates the primary voice and suppresses background chatter, keyboard clicks, fan hum, and room echo before audio reaches the model. Set to `near-field` (headsets, handsets, close-talking mics) or `far-field` (conference rooms, laptop/drive-thru mics, distant capture). Omit to disable. Set as a connection parameter. |
+| `voice_focus_threshold` | **universal-3-5-pro only.** Optional float `0.0`–`1.0` controlling how aggressively background audio is suppressed when `voice_focus` is set — higher = more aggressive. |
+| `language_codes` | **universal-3-5-pro only.** Optional **list** of ISO 639-1 codes (the connection param is **plural** — there is no singular `language_code` input param). Steers output toward the given languages on a per-token basis while still allowing native code-switching among them. Pass the languages you expect (e.g. `["en", "es"]`), or a single-element list (e.g. `["es"]`) for a monolingual session. When unset, no steering is applied and the model code-switches natively across all its supported languages. Distinct from `language_detection` (which only reports the detected language). Accepted codes: en, es, fr, de, it, pt, tr, nl, sv, no, da, fi, hi, vi, ar, he, ja, zh. |
+| `language_detection` | **universal-3-5-pro and universal-streaming-multilingual only.** Boolean (default `false`). When `true`, each `Turn` message includes the detected `language_code` and `language_confidence`. universal-3-5-pro natively code-switches without this — use it only when you need the per-turn language reported. |
 | `llm_gateway` | JSON-stringified LLM Gateway config — triggers LLM analysis on each completed turn, results delivered as `LLMGatewayResponse` messages |
 
 ### Messages Sent (Client to Server)
@@ -58,7 +58,7 @@ For new realtime/streaming code, use **`speech_model=universal-3-5-pro`** by def
 
 - **Begin:** Session start confirmation, includes session `id`
 - **Turn:** Transcript data with `transcript` text, `end_of_turn` boolean flag, and `words` array
-- **SpeechStarted:** Voice Activity Detection (VAD) event indicating speech has begun (U3 Pro only — use for barge-in detection)
+- **SpeechStarted:** Voice Activity Detection (VAD) event indicating speech has begun (universal-3-5-pro only — use for barge-in detection)
 - **SpeakerRevision:** Revised speaker labels at session close (only when `speaker_labels` is enabled). See Streaming Diarization below.
 - **LLMGatewayResponse:** LLM analysis result for the completed turn (only present when `llm_gateway` connection parameter is set)
 - **Termination:** Session end confirmation
@@ -88,14 +88,14 @@ Streaming is billed on **WebSocket-open duration per session**, and concurrent s
 ### universal-3-5-pro (recommended default)
 
 - Next-generation flagship streaming model; use it by default in new realtime/streaming integrations
-- More languages than u3-rt-pro: EN, ES, DE, FR, PT, IT, TR, NL, SV, NO, DA, FI, HI, VI, AR, HE, JA, ZH
-- Improved prompting and enhanced conversational-context features; supports `mode`, `prompt`, `agent_context`, and language detection like u3-rt-pro
+- 18 languages with native code-switching: EN, ES, DE, FR, PT, IT, TR, NL, SV, NO, DA, FI, HI, VI, AR, HE, JA, ZH
+- Punctuation-based turn detection, promptable, and enhanced conversational-context features; supports `mode`, `prompt`, `keyterms_prompt`, `agent_context`, `language_codes`, and language detection
+- The only U3-Pro-family streaming model still in the spec — supersedes the removed `u3-rt-pro`
 
-### u3-rt-pro
+### u3-rt-pro (legacy)
 
-- Universal-3 Pro Streaming — use when you intentionally need Universal-3 Pro behavior
-- 6 languages (EN, ES, DE, FR, PT, IT) with native code-switching
-- Punctuation-based turn detection, promptable, supports the `mode` accuracy/latency tradeoff
+- Universal-3 Pro Streaming (6 languages: EN, ES, DE, FR, PT, IT)
+- **Removed July 2026** from the streaming docs, model picker, and the `speech_model` spec enum; superseded by `universal-3-5-pro`, which offers more languages and improved prompting/context. New integrations should use `universal-3-5-pro`.
 
 ### universal-streaming-english
 
@@ -118,13 +118,13 @@ Streaming is billed on **WebSocket-open duration per session**, and concurrent s
 
 ## Turn Detection
 
-### U3 Pro
+### Universal-3.5 Pro
 
-Uses **punctuation-based** turn detection (`.` `?` `!`). The `end_of_turn_confidence_threshold` parameter has **NO effect** on U3 Pro models.
+Uses **punctuation-based** turn detection (`.` `?` `!`). The `end_of_turn_confidence_threshold` parameter has **NO effect** on universal-3-5-pro. Turn-end timing is set by the `mode` preset and tuned via `min_turn_silence`/`max_turn_silence` (defaults: `max_turn_silence` 1536ms, or 768ms when `speaker_labels` is enabled).
 
 ### Universal Streaming
 
-Uses **confidence-based** turn detection. The `end_of_turn_confidence_threshold` defaults to `0.4`.
+Uses **confidence-based** turn detection. The `end_of_turn_confidence_threshold` defaults to `0.4` (Universal Streaming English/Multilingual only); `max_turn_silence` defaults to 1280ms.
 
 ### Entity Splitting Caveat
 
@@ -137,7 +137,7 @@ A low `min_turn_silence` value can split entities like phone numbers across turn
 Change settings mid-stream without reconnecting. Fields are model-dependent:
 
 - **Universal Streaming:** `keyterms_prompt`, `min_turn_silence`, `max_turn_silence`
-- **universal-3-5-pro / u3-rt-pro:** `mode`, `prompt`, `keyterms_prompt`, `min_turn_silence`, `max_turn_silence`, `continuous_partials`, `vad_threshold`, `interruption_delay`, `agent_context`
+- **universal-3-5-pro:** `mode`, `prompt`, `keyterms_prompt`, `min_turn_silence`, `max_turn_silence`, `continuous_partials`, `vad_threshold`, `interruption_delay`, `agent_context`
 
 Send a JSON message:
 
@@ -199,7 +199,7 @@ Authorization: API_KEY
 
 Real-time PII redaction in streaming sessions. Detected PII is replaced in **final turns only** before being sent to the client.
 
-- Supported models: `universal-3-5-pro`, `u3-rt-pro`, `universal-streaming-english`, `universal-streaming-multilingual`
+- Supported models: `universal-3-5-pro`, `universal-streaming-english`, `universal-streaming-multilingual`
 - When `redact_pii=true`, `include_partial_turns` defaults to `false` automatically — partials would otherwise leak unredacted text
 - Audio redaction is **not** available for streaming. For redacted audio files, use [pre-recorded PII redaction](https://www.assemblyai.com/docs/guardrails/pii-redaction) with `redact_pii_audio`
 - Same policies as pre-recorded redaction (`person_name`, `phone_number`, `email_address`, `credit_card_number`, `us_social_security_number`, `date_of_birth`, etc.)
@@ -234,7 +234,7 @@ Enable speaker diarization by setting query parameters on the WebSocket URL:
 
 ### Revised speaker labels (SpeakerRevision)
 
-When the session ends, the server runs a final refinement pass over the whole conversation and emits a **single `SpeakerRevision` message** (when `speaker_labels` is enabled). It arrives **right before `Termination`**, after the client sends `Terminate`. (Streaming diarization itself is supported across current streaming models; the `SpeakerRevision` message is defined in the Universal-3 Pro-family streaming message set.)
+When the session ends, the server runs a final refinement pass over the whole conversation and emits a **single `SpeakerRevision` message** (when `speaker_labels` is enabled). It arrives **right before `Termination`**, after the client sends `Terminate`. (Streaming diarization itself is supported across current streaming models; the `SpeakerRevision` message is defined for universal-3-5-pro.)
 
 - A session emits **zero or one** `SpeakerRevision` message.
 - It contains a `revisions` array with **only the turns whose speaker labels changed** — unchanged turns are omitted.
@@ -261,11 +261,11 @@ When the session ends, the server runs a final refinement pass over the whole co
 
 ---
 
-## Context Carryover (universal-3-5-pro / u3-rt-pro)
+## Context Carryover (universal-3-5-pro)
 
-Universal-3 Pro-family streaming models automatically carry prior **finalized** turns (`end_of_turn: true`) forward as context to improve accuracy on the next turn. This is **on by default** — no configuration required — and is per-session (closing the WebSocket clears it).
+universal-3-5-pro automatically carries prior **finalized** turns (`end_of_turn: true`) forward as context to improve accuracy on the next turn. This is **on by default** — no configuration required — and is per-session (closing the WebSocket clears it).
 
-**Defaults:** context carryover enabled, ~3 prior entries carried (controlled by `previous_context_n_turns`, default `3`), ~1500-character max context. Older entries drop first. Set `previous_context_n_turns: 0` at connection time to disable automatic context carryover entirely.
+**Defaults:** context carryover enabled, up to **5** prior entries carried (controlled by `previous_context_n_turns`, server default `5`, range 0–100). Older entries drop first. Set `previous_context_n_turns: 0` at connection time to disable automatic context carryover entirely.
 
 You can additionally pass your voice agent's spoken reply (TTS text) via **`agent_context`** so the model knows the question the user is about to answer — especially valuable for short replies (`"yes"`, `"7pm"`, a single name) and spelled-out entities (emails, account IDs). For example, after the agent asks `"What's your email address?"`, `agent_context` helps the model produce `"user@assemblyai.com"` instead of `"user at assemblyai dot com"`.
 
@@ -278,11 +278,11 @@ Two ways to set it:
 { "type": "UpdateConfiguration", "agent_context": "Sure — what date would you like to book?" }
 ```
 
-**Limits:** `universal-3-5-pro` or `u3-rt-pro` only. Per-value cap ~1500 chars. Not billed separately (streaming is billed on session duration).
+**Limits:** `universal-3-5-pro` only. Per-value cap 1750 chars (`agent_context` and `prompt`). Not billed separately (streaming is billed on session duration).
 
 ---
 
-## Voice Focus (Noise Suppression, universal-3-5-pro / u3-rt-pro)
+## Voice Focus (Noise Suppression, universal-3-5-pro)
 
 Voice Focus isolates the primary voice and suppresses background chatter, keyboard clicks, fan hum, and room echo **before** the audio reaches the transcription model. Set the `voice_focus` connection parameter when you open the WebSocket. Pick the variant by how close the speaker is to the mic:
 
@@ -291,7 +291,7 @@ Voice Focus isolates the primary voice and suppresses background chatter, keyboa
 | Near field | `near-field` | Headsets, handsets, and other close-talking microphones |
 | Far field | `far-field` | Conference rooms, drive-thru speakers, laptop mics, other distant capture |
 
-Optionally tune `voice_focus_threshold` (float `0.0`–`1.0`) to control how aggressively background audio is suppressed — higher = more aggressive. Omit `voice_focus` to disable. Universal-3 Pro-family streaming models only.
+Optionally tune `voice_focus_threshold` (float `0.0`–`1.0`, default `0.7`) to control how aggressively background audio is suppressed — higher = more aggressive. Omit `voice_focus` to disable. universal-3-5-pro only.
 
 ```python
 CONNECTION_PARAMS = {
diff --git a/skills/assemblyai/references/voice-agents.md b/skills/assemblyai/references/voice-agents.md
index 9d17ccc..c160069 100644
--- a/skills/assemblyai/references/voice-agents.md
+++ b/skills/assemblyai/references/voice-agents.md
@@ -3,8 +3,8 @@
 AssemblyAI supports four paths for building voice agents:
 
 1. **Speech-to-Speech API** — single WebSocket for full voice agent (speech-in → LLM → speech-out)
-2. **LiveKit Agents** — fastest path to deployment using U3 Pro STT
-3. **Pipecat (by Daily)** — open-source, maximum customizability using U3 Pro STT
+2. **LiveKit Agents** — fastest path to deployment using Universal-3.5 Pro STT
+3. **Pipecat (by Daily)** — open-source, maximum customizability using Universal-3.5 Pro STT
 4. **Direct WebSocket** — fully custom STT builds (see `streaming.md`)
 
 ## Voice Agent API
@@ -125,6 +125,10 @@ The first `session.update` configures the agent one of **two mutually exclusive*
 
 `agent_id` is **mutually exclusive** with the inline fields — sending both in the same `session.update` raises a validation error. `agent_id` can only be set in the **first** `session.update`, before `session.ready`. After binding, you can still send later `session.update` messages to change mutable fields.
 
+### Bring your own LLM (`llm`)
+
+By default a voice agent uses AssemblyAI's **managed** conversational model. To run it on a different model, set the `llm` field on a **stored agent** (via the Agents REST API — this is a REST/create-time field, *not* a `session.update` field) to your own **OpenAI-compatible** endpoint: `{"llm": [{"base_url": "https://api.openai.com/v1", "model": "gpt-4o-mini", "api_key": "sk-..."}]}`. The list accepts **one entry today**; the model must support **streamed** chat completions; `api_key` is write-only. Send `"llm": []` to revert to the managed model. To use a frontier model without your own provider account, point `base_url` at the LLM Gateway (`https://llm-gateway.assemblyai.com/v1`) and pass your AssemblyAI key. Latency then depends on your endpoint. See `references/api-reference.md` for the full schema.
+
 ### Tool Definition Fields
 
 | Field | Type | Default | Notes |
@@ -243,7 +247,7 @@ For browser apps, enable echo cancellation via `getUserMedia({ audio: { echoCanc
 
 Set a voice via `session.output.voice` in `session.update` **before `session.ready`**. `output.voice` and `output.format` are immutable once the session is established — the voice **cannot be changed mid-conversation**. (`output.volume` is the exception — it remains mutable.) Default is `ivy`.
 
-**Every voice speaks all output languages** — 🇺🇸 English, French, German, Italian, Portuguese, Spanish, Hindi, Mandarin, Russian, Korean, and Japanese. The two groups below differ only by the voice's *primary accent*, not which languages it can speak. (Input recognition covers en/fr/de/it/pt/es only, so an agent can speak an output language it can't transcribe — useful for translation-style flows.)
+**Every voice speaks all output languages** — 🇺🇸 English, French, German, Italian, Portuguese, Spanish, Hindi, Mandarin, Russian, Korean, and Japanese. The two groups below differ only by the voice's *primary accent*, not which languages it can speak. Officially supported **output** languages (those with at least one native/primary-accent voice) are English, French, Italian, Spanish, Hindi, and Russian; German, Portuguese, Turkish, Dutch, and Swedish are on the roadmap. **Input** recognition uses Universal-3.5 Pro Streaming and covers **18 languages** with native code-switching (en, es, de, fr, pt, it, tr, nl, sv, no, da, fi, hi, vi, ar, he, ja, zh) — an agent can speak an output language it can't transcribe (useful for translation-style flows).
 
 **American-English accent** (carried into other languages):
 `ivy`, `james`, `tyler`, `winter`, `bella`, `david`, `kyle`, `helen`, `martha`, `river`, `emma`, `victor`, `eleanor`
@@ -286,19 +290,21 @@ In browsers, pre-handshake failures (like `UNAUTHORIZED`) surface as `close` eve
 
 **`universal-3-5-pro`** is the recommended default model for new STT-based voice agent work. The raw streaming API also defaults to it when `speech_model` is omitted; set it explicitly when pinning behavior or using an SDK that requires the field.
 
-| Feature | universal-3-5-pro | u3-rt-pro | universal-streaming-english | universal-streaming-multilingual |
-|---------|-------------------|-----------|------------------------------|----------------------------------|
-| Turn detection | Punctuation-based | Punctuation-based | Confidence-based | Confidence-based |
-| Custom prompting | Yes | Yes | No | No |
-| Keyterms boosting | Yes | Yes | Yes | Yes |
-| Speaker diarization | Yes | Yes | Yes | Yes |
-| Dynamic mid-session updates | Yes | Yes | Yes | Yes |
-| Multilingual code switching | Yes | Yes | No | Yes |
-| Languages | 18 (en, es, de, fr, pt, it, tr, nl, sv, no, da, fi, hi, vi, ar, he, ja, zh) | 6 (en, es, fr, de, it, pt) | English only | Multiple |
+| Feature | universal-3-5-pro | universal-streaming-english | universal-streaming-multilingual |
+|---------|-------------------|------------------------------|----------------------------------|
+| Turn detection | Punctuation-based | Confidence-based | Confidence-based |
+| Custom prompting | Yes | No | No |
+| Keyterms boosting | Yes | Yes | Yes |
+| Speaker diarization | Yes | Yes | Yes |
+| Dynamic mid-session updates | Yes | Yes | Yes |
+| Multilingual code switching | Yes | No | Yes |
+| Languages | 18 (en, es, de, fr, pt, it, tr, nl, sv, no, da, fi, hi, vi, ar, he, ja, zh) | English only | Multiple |
+
+(`u3-rt-pro` / Universal-3 Pro Streaming was removed from the streaming model picker and spec enum in July 2026 — superseded by `universal-3-5-pro`.)
 
-`end_of_turn_confidence_threshold` does NOT work with `universal-3-5-pro` or `u3-rt-pro` — it only applies to older universal-streaming models.
+`end_of_turn_confidence_threshold` does NOT work with `universal-3-5-pro` — it only applies to the Universal Streaming (English/Multilingual) models.
 
-## Turn Detection (universal-3-5-pro / u3-rt-pro)
+## Turn Detection (universal-3-5-pro)
 
 1. User pauses for `min_turn_silence` (e.g., 100ms)
 2. Model checks for terminal punctuation (`.` `?` `!`)
@@ -308,7 +314,7 @@ In browsers, pre-handshake failures (like `UNAUTHORIZED`) surface as `close` eve
 
 ## Silence Settings by Use Case
 
-These are for **Universal Streaming** models. U3 Pro-family defaults differ.
+These are for **Universal Streaming** models. universal-3-5-pro defaults differ (`max_turn_silence` 1536ms, or 768ms with `speaker_labels`).
 
 | Profile | min_turn_silence | max_turn_silence | Use Case |
 |---------|-----------------|-----------------|----------|