fix(model_prices): add Bedrock flex prices, Anthropic web search flags, Gemini shutdown dates, OpenRouter alias drift, Mistral Large 4 context

[devin-ai-integration]

Summary

Rolling registry PR for the scheduled model-pricing audit. Data-only change to model_prices_and_context_window.json and its byte-identical backup; generated schema unchanged. The Claude Haiku 5.5 flag fix, the openrouter/anthropic/claude-haiku-5.5 row and the Mistral Large 4 rows from earlier runs have since landed on main with the same values, so they drop out of this diff

Run of 2026-10-08 evening

Mistral (direct): max_input_tokens and max_tokens on mistral/mistral-large-4 and mistral/mistral-large-4-0 go from 524288 to 1048576. The raw https://docs.mistral.ai/models/mistral-large-4 model card shows Context 1M, and GET https://api.mistral.ai/v1/models returns max_context_length: 1048576 for both ids. max_output_tokens is not set on these rows and was left unset. tests/unit/test_mistral_large_4_model_metadata.py (added in #44870 with the 524288 value) now expects 1048576

OpenRouter (source for every row: GET https://openrouter.ai/api/v1/models, OpenRouter's own price list for its routes)

Key	Field	Before	After
openrouter/mistralai/mistral-large-4-0	max_input_tokens	524288	1048576
openrouter/~anthropic/claude-sonnet-latest	cache_read_input_token_cost	2e-07	1e-07
openrouter/~openai/gpt-sol-latest	cache_read_input_token_cost / _above_272k_tokens	2e-07 / 4e-07	1e-07 / 2e-07
openrouter/anthropic/claude-sonnet-5.5:batch	cache_read_input_token_cost	1e-07	5e-08
openrouter/~anthropic/claude-haiku-latest	prices, limits, *_above_100k_tokens tiers	Haiku 4.5 values (1e-06 / 5e-06, 200k / 64k)	Haiku 5.5 values (1e-07 / 5e-07, cache read 1e-08, write 1.25e-07, 1h write 2e-07, 100k-token tier, 1M / 128k)
openrouter/anthropic/claude-haiku-5.5:batch	new row	none	5e-08 / 2.5e-07, cache read 5e-09, write 6.25e-08, 1h write 1e-07, 100k-token tier at 5x, 1M / 128k

The ~anthropic/claude-haiku-latest alias drifted when OpenRouter repointed it at Haiku 5.5, and the API now returns the same prices and limits as openrouter/anthropic/claude-haiku-5.5, which main already carries. The Sonnet 5.5 and Haiku 5.5 cache-read prices match the raw https://platform.claude.com/docs/en/about-claude/pricing rows ($0.10 and $0.01 per MTok)

Checked this run and not changed: codestral-2508, because /v1/models says 256k context but the raw https://docs.mistral.ai/models/model-cards/codestral-25-08 card says 128k. Anthropic claude-sonnet-4-5-20250929, because /v1/models reports the 200k default while the registry carries the documented 1M beta window. gpt-audio-mini-2025-10-06, with no separate price row. OpenRouter open-weight routes, whose headline price follows the cheapest live provider and moves daily. Fireworks qwen3p8-2p4t-a95b still has no price on https://fireworks.ai/pricing or its model page via r.jina.ai (context 262k only). Bedrock ListFoundationModels profile-only base ids (openai.gpt-6-astra, xai.grok-4.6, zai.glm-5.3 and others) already have us./global. rows

Bedrock flex tier prices (absorbed from #45320 by @roytev, run of 2026-10-08)

input_cost_per_token_flex, output_cost_per_token_flex and, where the model has cache pricing, cache_read_input_token_cost_flex added to 110 Bedrock rows across 43 models (225 fields). Applied at field level from the PR's own merge base, so no existing field changed

Source: AWS Price List API offer file https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/index.json (13315 products, 3157 *-flex usage types). Every one of the 225 values was matched to the *-flex usage type for the same model and region (bedrock/<region>/... rows to that region, us./eu./apac. profiles to the regions whose standard price equals the row's standard price, unprefixed rows to any region), with the per-1K or per-1M unit converted to per token. Example: us.amazon.nova-2-lite-v1:0 maps to USE1/USE2/USW2-Nova2.0Lite-input-tokens-flex 0.000165 per 1K, so input_cost_per_token_flex 1.65e-07, output 1.375e-06, cache read 4.125e-08

Together AI (absorbed in part from #45307)

deprecation_date 2026-09-29 removed from together_ai/deepseek-ai/DeepSeek-V4-Flash-0731 and together_ai/deepseek-ai/DeepSeek-V4-Pro-0813. The raw https://docs.together.ai/docs/deprecations.md table no longer lists either id (it lists only deepseek-ai/DeepSeek-V4-Pro at 2026-08-27), and GET https://api.together.ai/v1/models still serves both on 2026-10-08 with prices matching main, so the lapsed date would wrongly mark live models as retired

Dropped from #45307: the together_ai/Qwen/Qwen3.7-Max price change, because /v1/models returns 2.50 / 7.50 / 0.50 cached per 1M while the raw https://www.together.ai/pricing row says $1.50 / $4.50 / $0.30 cached, so the official sources disagree. The together/Tev1-4B-experimental cache-price removal, because the pricing page publishes no cache column for it. The 19 absent-model warnings, because Together gives no removal date for them

Anthropic (direct litellm_provider: "anthropic")

supports_web_search: true added to 11 rows that lacked it: claude-opus-5-5, claude-fable-5-1, claude-opus-5, claude-sonnet-5, claude-fable-5, claude-opus-4-8, claude-opus-4-7, claude-sonnet-4-6, claude-opus-4-6, claude-opus-4-5-20251101, claude-haiku-4-5-20251001. Source: GET https://api.anthropic.com/v1/models returns capabilities.server_tools.web_search.supported: true for each id, and https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool says to read that Models API capability for per-model availability

Gemini API

Source for every row: https://ai.google.dev/gemini-api/docs/deprecations, raw HTML table rows cross-checked with r.jina.ai

Key	Before	After
gemini/gemini-2.5-flash-image	2026-10-02	2027-03-15
gemini/gemini-omni-flash-preview	2026-09-30	2026-10-22
gemini-2.5-flash-native-audio-preview-12-2025 and gemini/ copy	none	2026-11-17
gemini-3.1-flash-live-preview and gemini/ copy	none	2026-11-17
gemini-2.5-flash-preview-tts and gemini/ copy	none	2026-11-17
gemini/gemini-2.5-pro-preview-tts, gemini/gemini-3.1-flash-tts-preview	none	2026-11-17

Checked in earlier runs and not changed

OpenAI (/v1/models): only fine-tune, dated search, deep-research and computer-use snapshots are absent; no published per-token row was confirmed for the dated aliases

Gemini (/v1beta/models): antigravity-preview-* is a managed agent and aqa uses generateAnswer, so neither is a token-priced chat row. API token limits disagree with https://ai.google.dev/gemini-api/docs/models for image, Lyria, Veo, Deep Research and Live rows, so no limit was changed

Bedrock (ListFoundationModels, ListInferenceProfiles): only provisioned-only Nova/Titan variants and legacy profiles are missing

Azure: https://prices.azure.com/api/retail/prices matches azure/gpt-5.6-luna and azure/gpt-5.6-terra. Voyage: https://docs.voyageai.com/docs/pricing matches. Nebius: pricing page is behind a login, nothing checked. Fireworks: qwen3p8-2p4t-a95b has no published per-token price. Perplexity: https://docs.perplexity.ai/docs/agent-api/models matches. Cohere: model page limits kept over the /v1/models context length. Groq: allam-2-7b has no price on https://groq.com/pricing. Mistral, xAI and DeepSeek: no missing rows

Capability pins: ci_cd/cost_map_pins.json does not exist on main, so no pinned flag could be compared

Open registry PRs re-checked and not absorbed

PR	Why not absorbed
#43879 kimi-k2.5	Same values already on main as moonshot/kimi-k2.5
#43753 deepseek-v3	Same values already on main; https://api-docs.deepseek.com/quick_start/pricing no longer lists it
#43859 tinker	No tinker provider exists in LiteLLM
#32117 Cohere embed-multilingual-light-v3.0	https://cohere.com/pricing no longer lists per-token Embed v3 pricing
#40368 Bedrock Claude 3 Haiku cache pricing	Keys are gone from main
#41376 xAI grok-4 flags	xai/grok-4* keys no longer exist on main
#30775 Bedrock us.xai.grok-4.3	Not returned by Bedrock ListFoundationModels
#36274 Azure gpt-5.6 terra/luna	Contradicted by the Azure retail prices API
#29920 qwen3.7-max	Filed under litellm_provider: "anthropic"
#36084, #31460	Every registry field they change already has the same value on main
#41775 Together date removals	Superseded by the current Together deprecations table
#45386, #45158, #45016, #44822, #44817, #44808, #44651, #44416, #44400, #44379, #44303, #44299, #44168, #44167, #44163, #44091	New-provider or code PRs whose JSON rows depend on code in the same PR, left alone
#45358, #45200, #45026, #45019, #45016, #44969, #44863, #44856, #44839, #44700, #44601, #44432, #43581, #43457, #43390, #43363, #43094, #42829, #42669, #42668, #41857, #38165, #36385, #31018, #29679, #22086	Code/behavior PRs that touch the JSON incidentally, left alone

Superseded and closed: #45320, #45307

Checks: LITELLM_LOCAL_MODEL_COST_MAP=True uv run --no-sync pytest -q tests/unit/test_model_prices_schema.py tests/unit/test_model_cost_aliases.py 86 passed, check_files_match.py passed, cost_map_guard.py --base origin/main --head HEAD --head-ref litellm_model_registry_audit_20261006 passed, duplicate-key JSON parse and jsonschema validation passed, schema generator produced no diff

Link to Devin session: https://app.devin.ai/sessions/74ee32335515402a92a8b25ff7cfdab0
Open in Devin Desktop: https://app.devin.ai/desktop/session/74ee32335515402a92a8b25ff7cfdab0?variant=devin

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Adds pricing and capability data for new model variants.

The registry additions appear safe to merge

What we checked:

• New model names resolve: The lookup constructs provider-prefixed names and checks them before falling back to other names

Summary

Adds Mistral Large 4, its native alias, and its OpenRouter route to both model registries

• Adds token prices, cache-read prices, limits, and supported features
• Enables native Mistral reasoning requests through the existing metadata-driven code
• No actionable issues were established
• devin-ai-integration[bot] acknowledges that launch-discount prices need updating when the sale ends and intentionally uses the API's 524288 context limit despite the model page saying “1M”

_{Reviews (1) · Last reviewed commit: "feat(model_prices): add Mistral Large 4 ..."}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_model_registry_audit_20261006 (bbd7b53) with main (3571b5d)}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}