Sync prices from the HK observation point

[john2ai]

The HK observation point's daily run found changes in the vendors' published prices.

Each commit below is one entry, so every diff is the change to that vendor
alone. The last is the global.json version bump — the published table is
version-gated, so without it the app never sees any of this.

Read it the way the checklist asks: in / out / cache_read against
the vendor's own page. A correct reading reports no change, so anything
here is a source disagreeing with what the entry carries.

This run exited 1 — at least one source could not be read from
this box and was skipped (today that is openai and togetherai, the US
runner's half). The log below names every one.

Commits

• a5111abb Bump the version for 5 price(s), 9 new, 1 gone
• 32c8ff34 Sync xai from its source — 1 new
• ed731b42 Sync volcesark from its source — 5 new
• d3ec75df Sync openrouter from its source — 5 price(s), 3 new, 1 gone

Full run log

✗ anthropic  (https://www.anthropic.com/pricing)
  no current-model price blocks matched — the page has been restructured
  Skipped. Check the source, then update its adapter in scripts/sources/.

✗ baidu-qianfan  (https://cloud.baidu.com/doc/qianfan/s/wmh4sv6ya)
  baidu-qianfan: the source no longer lists "GLM-5.3", which the entry flags as flagship — pick a new flagship before this entry can be written
  Skipped. Check the source, then update its adapter in scripts/sources/.

cerebras  (https://inference-docs.cerebras.ai/llms.txt)
  ! 2 model pages read from the vendor's own docs index

deepinfra  (https://api.deepinfra.com/v1/openai/models)
  ! not token-billed (audio/seconds/images): 85 row(s), e.g. Bria/fibo, PrunaAI/p-video, BAAI/bge-m3-multi

friendli  (https://friendli.ai/api/public/model-apis)
  ! 1 row(s) skipped — non-text output or not token-billed

google-gemini  (https://aistudio.google.com/docs/pricing.md)
  ! 28 models with a Standard table, 7 of them in the entry

kimi  (https://platform.kimi.com/docs/pricing/chat.md)
  ! the vendor marks these retired: kimi-k2.5, moonshot-v1-8k, moonshot-v1-32k, moonshot-v1-128k, moonshot-v1-auto, moonshot-v1-8k-vision-preview, moonshot-v1-32k-vision-preview, moonshot-v1-128k-vision-preview, kimi-k2-0905-preview, kimi-k2-0711-preview, kimi-k2-turbo-preview, kimi-k2-thinking, kimi-k2-thinking-turbo

novita  (https://api.novita.ai/v3/openai/models)
  ! 7 row(s) skipped — non-chat or unpriced

perplexity  (https://docs.perplexity.ai/docs/router/models.md)
  ! 4 rows from the vendor's own rate card
  ! cache writes billed at each row's input rate (the page lists no dedicated write rate)
  ! except perplexity/glm-5.3-flash, which supports no cache writes

siliconflow-cn  (https://siliconflow.cn/pricing + https://siliconflow.com/pricing)
  ! input-only rows skipped (embeddings/rerankers): 0
  ! zero-priced (free-tier) rows skipped: 0

stepfun  (https://platform.stepfun.com/docs/zh/guides/pricing/details.md + https://platform.stepfun.ai/docs/en/guides/pricing/details.md)
  ! 3/3 rows priced from the vendor's China page

stepfun-intl  (https://platform.stepfun.com/docs/zh/guides/pricing/details.md + https://platform.stepfun.ai/docs/en/guides/pricing/details.md)
  ! 3/3 rows priced from the vendor's international page

✗ togetherai  (https://docs.together.ai/docs/serverless-models.md)
  https://docs.together.ai/docs/serverless-models.md: HTTP 403
  Skipped. Check the source, then update its adapter in scripts/sources/.

✗ openai  (https://developers.openai.com/api/docs/pricing.md)
  https://developers.openai.com/api/docs/pricing.md: HTTP 403
  Skipped. Check the source, then update its adapter in scripts/sources/.

openrouter  (https://openrouter.ai/api/v1/models + /{id}/endpoints)
  ! 336 of 402 models priced from their upstreams; 402 text-out models enumerated (:free mirrors excluded)
  ! priced differently per upstream: openai/gpt-6-luna-pro (6 endpoints, 4 prices), openai/gpt-6-sol-pro (6 endpoints, 4 prices), openai/gpt-6-luna (7 endpoints, 4 prices), openai/gpt-6-sol (7 endpoints, 4 prices), anthropic/claude-opus-5.5 (11 endpoints, 3 prices), x-ai/grok-4.7 (4 endpoints, 2 prices), deepseek/deepseek-v4.1-flash (26 endpoints, 13 prices), openai/gpt-6-astra (6 endpoints, 4 prices), openai/gpt-6-astra-pro (5 endpoints, 4 prices), google/gemini-3.8-flash (4 endpoints, 3 prices), ibm-granite/granite-4.2-8b (2 endpoints, 2 prices), z-ai/glm-5.3-flash (31 endpoints, 13 prices)
  ! left unpriced this run (rows keep their numbers): x-ai/grok-4.7 (no agreed price among 4 upstreams — left as is), ibm-granite/granite-4.2-8b (no agreed price among 2 upstreams — left as is), qwen/qwen3.8-27b (no agreed price among 16 upstreams — left as is), x-ai/grok-4.6 (no agreed price among 5 upstreams — left as is), deepseek/deepseek-v4-flash-0731 (no agreed price among 30 upstreams — left as is), thinkingmachines/inkling-small (no agreed price among 2 upstreams — left as is), inclusionai/ling-3.0-flash (no agreed price among 2 upstreams — left as is), google/gemini-3.5-flash-lite (no agreed price among 6 upstreams — left as is), x-ai/grok-4.5 (no agreed price among 4 upstreams — left as is), openrouter/fusion (no endpoints), x-ai/grok-build-0.1 (no agreed price among 4 upstreams — left as is), google/gemini-3.1-flash-lite (no agreed price among 6 upstreams — left as is)

openrouter  (https://openrouter.ai/api/v1/models + /{id}/endpoints)
  ~ deepseek-v4-pro-0813.context: 1048576 -> 1024000
  ~ z-ai/glm-5.3:batch.in: "0.72" -> "0.45"
  ~ z-ai/glm-5.3:batch.out: "2.4" -> "2"
  ~ z-ai/glm-5.3:batch.cache_read: "0.12" -> "0.1"
  ~ nvidia/nemotron-3.5-lightning.max_output: 235929 -> 131072
  + aion-labs/aion-3.5-mini is new
  + aion-labs/aion-3.5 is new
  + upstage/solar-mini4 is new
  - mistralai/devstral-2512 is gone from the source

qianwenai  (https://help.aliyun.com/zh/model-studio/billing-for-model-studio)
  ! cache rates are not read: these models are excepted from every published percentage (20% implicit, 10% explicit-hit) and the vendor points at the console instead

volcesark  (https://www.volcengine.com/api/doc/getDocDetail?DocumentID=1544106, https://www.volcengine.com/api/doc/getDocDetail?DocumentID=1925114, https://www.volcengine.com/api/doc/getDocDetail?DocumentID=2366394)
  + doubao-seed-2.1-pro is new
  + doubao-seed-2.1-lite is new
  + doubao-seed-2.0-mini is new
  + kimi-k2.8-preview is new
  + deepseek-v4.1-flash is new

xai  (https://docs.x.ai/docs/models.md)
  + grok-4.7 is new

xiaomi-mimo  (https://mimo.mi.com/static/docs/price/pay-as-you-go.md)
  ! batch API, a different billed mode with no field here: mimo-v2.6-pro
  ! batch API, a different billed mode with no field here: mimo-v2.6-flash
  ! billed by duration, no field for it: mimo-v2.5-asr

xiaomi-mimo-token-plan  (https://mimo.mi.com/static/docs/price/pay-as-you-go.md)
  ! batch API, a different billed mode with no field here: mimo-v2.6-pro
  ! batch API, a different billed mode with no field here: mimo-v2.6-flash
  ! billed by duration, no field for it: mimo-v2.5-asr

3 entr(ies) to change: 5 price(s), 9 new, 1 gone
4 entr(ies) skipped — see the ✗ above
no source, still hand-maintained: ai21, azure-openai, baichuan, baidu-qianfan-token-plan, baseten, bedrock, cloudflare-workers-ai, cohere, databricks, digitalocean, fireworks, groq, huggingface, kimi-for-coding, mistral, nebius, oci, ovhcloud, qianwenai-token-plan, sap-ai-core, siliconflow, tencent-token-plan, tencent-tokenhub, upstage, vertex, watsonx
✓ wrote entries/openrouter/models.json
✓ wrote entries/volcesark/models.json
✓ wrote entries/xai/models.json
✓ bumped global.json version 61 -> 62
✓ committed entries/openrouter/models.json
✓ committed entries/volcesark/models.json
✓ committed entries/xai/models.json
✓ committed global.json

commits are local; nothing was pushed

---

Opened by scripts/sync-hk.sh on the HK box. Nothing here is merged
automatically.