fix(model_prices): rolling registry audit - verified models and rates for Novita, DeepInfra, W&B, Bedrock Sol, Gemini, Fireworks, Azure gpt-5.6, Mistral, Together

[devin-ai-integration]

TLDR

Problem this solves:

• Bedrock openai.gpt-5.6-sol billed 25% over published rates
• openai.gpt-5.6-cyber on Bedrock had no entry
• Claude Haiku 3 and Opus 3 had wrong 1h cache-write rates
• 117 live Novita, DeepInfra and W&B models were missing entirely
• Gemini 2.5 Flash / Flash-Lite aliases had 10x-off cached-input rates
• Azure gpt-4.1-nano carried a retirement date six months early
• Two retired Together models had no deprecation_date
• Fireworks DeepSeek V4 Pro (0813) was missing
• Azure gpt-5.6 entries billed cache writes at the plain input rate
• Azure US/EU gpt-5.6 priority rates sat at 1.25x Global instead of 1.1x
• Ten live Mistral ids (ministral 3b/14b, voxtral, embed, aliases) had no entry
• Together published cache-read rates the registry lacked
• Databricks GLM 5.2 and Kimi K3 had no entry
• Z.AI GLM-5.3 had no entry
• xAI still listed eight retired Grok era models
• Four Together models removed 2026-08-27 had no deprecation_date
• Azure sora-2 gained an official retirement date

How it solves it:

• Every value taken from the provider's own page or public API, read raw
• Rolling registry PR, one open PR instead of many pricing PRs
• Field-level merges only, existing tiered pricing keys untouched

User Flow

Before: a developer running new serverless models through the gateway sees $0 spend or spend that disagrees with the provider's invoice

1. They send POST https://litellm-domain/v1/chat/completions with "model": "deepinfra/moonshotai/Kimi-K2.5" and a 10k token prompt
2. The call succeeds, and https://litellm-domain/ui/?page=logs shows it at $0 spend because the model has no cost-map entry
3. They switch to "model": "gemini/gemini-flash-latest" with a cached prefix and the log shows cached input billed at $0.0000003 per token, 10x the published rate
4. They send "model": "bedrock/us.openai.gpt-5.6-sol" and the log shows $5.50 per million input tokens while AWS bills $4.40

After: those calls are priced, and the numbers match the provider pages

1. The same deepinfra/moonshotai/Kimi-K2.5 request is logged at $0.45 in, $2.25 out per million tokens, with cache reads at $0.07
2. gemini/gemini-flash-latest cached input is logged at $0.03 per million tokens, matching Google's pricing table
3. bedrock/us.openai.gpt-5.6-sol is logged at $4.40 in and $22.00 out, and prompts over 272k at $8.80 / $33.00
4. Requests to models the provider marks deprecated or does not publish a price for are unchanged, so nothing silently gains a guessed rate

Relevant issues

Registry audit run against provider docs, no issue.

Linear ticket

Changes by provider

Bedrock (mantle) and Anthropic

bedrock_mantle/openai.gpt-5.6-sol, us.openai.gpt-5.6-sol, global.openai.gpt-5.6-sol: https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-sol.html gives In-Region and Geo CRIS $4.40 in, $5.50 30m cache write, $0.44 cache read, $22.00 out at the 272K tier, and $8.80 / $11.00 / $0.88 / $33.00 at the 1M tier. Global CRIS is $4.00 / $5.00 / $0.40 / $20.00 and $8.00 / $10.00 / $0.80 / $30.00. The registry held launch rates, 25% high on every field.

bedrock_mantle/openai.gpt-5.6-cyber (new): https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-cyber.html gives $13.75 in, $17.1875 30m cache write, $1.375 cache read, $82.50 out, 272K context, Responses only at /openai/v1/responses, image plus text input, prompt caching and server-side tool calling. Geo and Global are unsupported, so only the bedrock_mantle/ key exists.

claude-3-haiku-20240307, claude-3-opus-20240229: https://docs.anthropic.com/en/docs/about-claude/pricing documents 1h cache writes at 2x base input, so $0.50 and $30.00 per MTok. Both carried $6.00, the Sonnet rate.

Gemini and Vertex cached input

https://ai.google.dev/gemini-api/docs/pricing and https://cloud.google.com/vertex-ai/generative-ai/pricing list context-cache reads at $0.03 per million tokens for Gemini 2.5 Flash and $0.01 for 2.5 Flash-Lite. Corrected on gemini/gemini-2.5-flash-preview-09-2025, gemini/gemini-flash-latest, gemini/gemini-2.5-flash-lite-preview-06-17, gemini/gemini-flash-lite-latest, and the Vertex gemini-2.5-flash-preview-09-2025 and gemini-2.5-flash-lite-preview-06-17 entries, which had drifted 10x from the dated models they alias.

Novita (50 added, 40 fields corrected)

Source: Novita's own catalog API https://api.novita.ai/v3/openai/models plus https://novita.ai/pricing and https://novita.ai/docs/guides/llm-function-calling. Added the active, priced, slash-namespaced chat models the API returns that had no entry, each with context and output limits, published input/output and cache-read prices, and capability flags taken from the API's own features (function-calling, structured-outputs, reasoning) and input_modalities. Corrections include novita/qwen/qwen3-coder-480b-a35b-instruct ($0.38 / $1.55), novita/xiaomimimo/mimo-v2-flash ($0.11 / $0.33, $0.024 cache read), cache-read plus supports_prompt_caching on novita/zai-org/glm-4.5-air and novita/moonshotai/kimi-k2-thinking, and output-limit fixes on the Kimi K2, DeepSeek V3 and Qwen entries. Inactive, unpriced and catalog-only rows were skipped.

DeepInfra (67 added, 25 fields corrected)

Source: DeepInfra's public model API https://api.deepinfra.com/models/list (pricing page: https://deepinfra.com/pricing), which publishes cents_per_input_token, cents_per_output_token, the cached-input rate as a fraction of input, context length, tags and a deprecation flag. Added the non-deprecated text-generation models with published token pricing and no entry, among them deepinfra/deepseek-ai/DeepSeek-V4-Pro, DeepSeek-V4-Flash, moonshotai/Kimi-K3, Kimi-K2.5, Kimi-K2.6, Kimi-K2.7-Code, zai-org/GLM-5, GLM-5.1, GLM-5.2, the Qwen/Qwen3.5-*, Qwen3.6-*, Qwen3.8-* and Qwen3-VL-* families, ByteDance/Seed-*, google/gemma-4-*, google/gemini-3.x, anthropic/claude-* served by DeepInfra, nvidia/NVIDIA-Nemotron-3-*, MiniMaxAI/MiniMax-M3, XiaomiMiMo/MiMo-V2.5*, thinkingmachines/Inkling*, tencent/Hy3 and stepfun-ai/Step-3.7-Flash. Flags come from the API tags: tools sets function calling and tool choice, structured-output/json sets response schema, reasoning sets reasoning, multimodal sets vision. Corrections are price-only where the API disagrees with the registry, e.g. deepinfra/openai/gpt-oss-120b ($0.037 / $0.17 against $0.05 / $0.45), Qwen/Qwen3-30B-A3B, Qwen/Qwen3-Next-80B-A3B-Instruct, deepseek-ai/DeepSeek-V3.1, meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 and Gryphe/MythoMax-L2-13b, plus cache-read rates the API publishes and the registry lacked. Entries the API marks deprecated were left untouched.

Weights & Biases Inference (19 added, 6 corrected)

Source: https://wandb.ai/site/pricing/tokens/ and https://docs.wandb.ai/inference/models, read as raw table rows. Added wandb/deepseek-ai/DeepSeek-V4-Pro ($1.15 / $2.55, $0.20 cache hit), DeepSeek-V4-Flash and DeepSeek-V4-Flash-0731, zai-org/GLM-5.2 ($0.76 / $2.42), Qwen/Qwen3.8-27B ($0.40 / $3.00), Qwen/Qwen3.6-27B, MiniMaxAI/MiniMax-M3, moonshotai/Kimi-K2.7-Code, google/gemma-4-31B-it and the rest of the priced rows with no entry. Corrected wandb/openai/gpt-oss-120b ($0.03 / $0.17), gpt-oss-20b ($0.03 / $0.13), Qwen/Qwen3-Coder-480B-A35B-Instruct ($1.00 / $1.50), meta-llama/Llama-3.1-8B-Instruct and Llama-3.3-70B-Instruct, and deepseek-ai/DeepSeek-V3.1.

Fireworks

https://docs.fireworks.ai/serverless/pricing lists DeepSeek V4 Flash (0731) at $0.22 in, $0.007 cached, $0.66 out. fireworks_ai/accounts/fireworks/models/deepseek-v4-flash-0731 carried $0.14 / $0.028 / $0.28, which is Novita's rate for the same model. Every other Fireworks serverless row on that page already matches the registry.

Azure and Together deprecation dates (2026-08-26 run)

azure/gpt-4.1-nano and azure/gpt-4.1-nano-2025-04-14: https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule lists gpt-4.1-nano version 2025-04-14 as Legacy with retirement no earlier than 2027-04-14, matching the gpt-4.1 and gpt-4.1-mini entries already in the registry. Both carried 2026-10-14, which appears nowhere on the current schedule

together_ai/meta-llama/Llama-Guard-4-12B and together_ai/google/gemma-3n-E4B-it: Together's deprecations page (https://docs.together.ai/docs/deprecations) lists both with a 2026-08-25 removal date, read from the raw table rows. Added deprecation_date: 2026-08-25 to each and moved them from the serverless list to the deprecated map in tests/test_litellm/test_together_ai_model_metadata.py, which encoded the pre-2026-08-25 provider state

The DeepInfra-served Claude entries added earlier in this PR now carry the model-intrinsic Anthropic flags the family invariant tests enforce (supports_adaptive_thinking, supports_sampling_params: false, prompt_cache_min_tokens, and thinking_always_on on Fable 5), mirrored from the first-party entries for the same models, including the later-added claude-sonnet-4-6 and claude-haiku-4-5 rows

Fireworks DeepSeek V4 Pro (0813), new entry (2026-08-26 run)

https://docs.fireworks.ai/serverless/pricing now lists DeepSeek V4 Pro (0813) at $1.32 in, $0.044 cached, $3.96 out, and the model page (https://app.fireworks.ai/models/fireworks/deepseek-v4-pro-0813) confirms serverless availability, function calling and 1M context. Added fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813, absorbed from #38218 after re-verifying every field against those pages

Azure gpt-5.6 cache writes and Data Zone priority rates (2026-08-26 run)

The Azure retail price API (https://prices.azure.com/api/retail/prices, cross-checked with https://azure.microsoft.com/en-us/pricing/details/cognitive-services/openai-service/) publishes gpt-5.6 Sol/Terra/Luna cache writes at 1.25x input on every meter: Sol $6.25, Terra $2.50, Luna $0.25 per MTok standard, doubled for the above-272k tier, and matching priority meters. Added cache_creation_input_token_cost, ..._above_272k_tokens and ..._priority to the 12 azure/gpt-5.6* entries. The same API prices US/EU Data Zone standard and priority meters at 1.1x Global, so the azure/us/... and azure/eu/... priority input, cached-input and output rates were corrected from 1.25x to 1.1x. #38370 merged these same fields into the base branch mid-run, so this branch now matches staging for the azure gpt-5.6 entries, including the long-context priority cache-write rates. Regression tests in tests/test_litellm/test_cost_calculator.py pin the 1.25x cache-write and 1.1x Data Zone invariants

Mistral missing models (2026-08-26 run)

Ten ids Mistral's /v1/models serves had no registry entry. Each was added from its official model card: mistral/ministral-14b-2512 and -latest ($0.20/$0.20, 256k, https://docs.mistral.ai/models/ministral-3-14b-25-12), mistral/ministral-3b-2512 and -latest ($0.10/$0.10, 128k, https://docs.mistral.ai/models/ministral-3-3b-25-12), mistral/mistral-embed-2312 (https://docs.mistral.ai/models/mistral-embed-23-12), mistral/mistral-medium-3 (listed as an API name on https://docs.mistral.ai/models/mistral-medium-3-5-26-04, same rates as mistral-medium-3-5), mistral/voxtral-small-2507 and -latest ($0.004/min audio, $0.10/$0.40 per MTok, 32k, https://docs.mistral.ai/models/voxtral-small-25-07), plus the -latest aliases for voxtral-mini-transcribe-realtime and voxtral-mini-tts. The 3b card advertises 256k, but Mistral's /v1/models caps ministral-3b-2512 at 131072, so the 3b entries and the pre-existing mistral/ministral-3-3b-2512 id all carry 131072, matching what the hosted API accepts

Together cache-read pricing (2026-08-26 run)

https://www.together.ai/pricing publishes cached-input rates, read from the raw rows. Added cache_read_input_token_cost plus supports_prompt_caching to MiniMax-M3 ($0.06), Qwen3.8-2.4T-A95B ($0.50), Qwen3.5-397B-A17B ($0.35), DeepSeek-V4-Flash-0731 ($0.03), DeepSeek-V4-Pro ($0.20), DeepSeek-V4-Pro-0813 ($0.13), Muse-Glimmer-30B ($0.04), Kimi-K2.7-Code ($0.19), Kimi-K3 ($0.30), nemotron-3-ultra-550b-a55b ($0.20), Inkling ($0.17), Inkling-Small ($0.10) and GLM-5.2 ($0.26). Qwen3.7-Max keeps its $1.25/$3.75 rates, which the pricing page still shows, and gains the page's $0.13 cached rate

Databricks and Z.AI additions (2026-08-27 run)

databricks/databricks-glm-5-2 and databricks/databricks-kimi-k3: added from https://www.databricks.com/product/pricing/foundation-model-serving, which lists GLM 5.2 at 20.000 DBU in / 62.857 DBU out / 3.714 DBU cache read per MTok and Kimi K3 at 42.857 / 214.286 / 4.286, converted at $0.070 per DBU like the sibling Databricks entries. zai/glm-5.3: added from https://docs.z.ai/guides/overview/pricing, $1.40 in, $0.26 cached, $4.40 out per MTok. Absorbed from #38415 after re-verifying every rate against those pages, including its cost-calculator test

xAI retired models and multi-agent endpoint (2026-08-27 run)

https://docs.x.ai/docs/models no longer lists grok-2, grok-2-1212, grok-2-latest, grok-2-vision, grok-2-vision-1212, grok-2-vision-latest, grok-beta or grok-vision-beta, so those entries are removed. https://docs.x.ai/developers/model-capabilities/text/multi-agent documents grok-4.20-ma and grok-4.20-ma-fast as Responses-only, so both entries switch to mode: responses with supported_endpoints: ["/v1/responses"]. Absorbed from #38510 with its registry test

Together and Azure deprecation dates (2026-08-27 run)

https://docs.together.ai/docs/deprecations now lists nvidia/Nemotron-3-ultra-550b-a55b, pearl-ai/gemma-4-31b-it, deepseek-ai/DeepSeek-V4-Pro and moonshotai/Kimi-K2.7-Code with a 2026-08-27 serverless removal date, read from the raw table rows. Each gets deprecation_date: 2026-08-27 and moves to the deprecated map in tests/test_litellm/test_together_ai_model_metadata.py. The three successor pointers at DeepSeek-V4-Pro now point at DeepSeek-V4-Pro-0813, the newer snapshot of the same model still on the serverless list. azure/sora-2 gets deprecation_date: 2026-10-15 from the sora-2 2025-12-08 row on https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirement-schedule. Groq, Cohere, Anthropic, Gemini, Vertex and OpenAI deprecation schedules were re-read raw this run; every date already in the registry matches the provider pages, so none changed

Absorbed from other open registry PRs

• fix(model_prices): correct cache read pricing for Gemini flash latest and preview aliases #38215 (Gemini cached input): same numbers, confirmed against Google's raw pricing tables, absorbed.
• fix(pricing): correct wandb model prices off by 100,000x #33947 (W&B pricing): re-derived from the current W&B token-pricing table, absorbed. Two of its rows are now stale on the provider page, so this PR carries the current numbers rather than that PR's.
• fix(pricing): add azure gpt-5.6 cache write rates and correct data zone priority #38370 (Azure gpt-5.6 cache write and Data Zone): its rates were confirmed against the Azure retail price API and applied here; it then merged into the base branch mid-run, so this branch was re-merged onto it and now carries the identical values.
• fix(cost): apply Together AI cache read pricing and per-model registry rates #38280 (Together cache reads): the registry fields were re-verified against Together's pricing page and absorbed, except its Qwen3.7-Max change to $2.50/$7.50 with a $0.50 cached rate, which contradicts the page's $1.25/$3.75/$0.13; the page rates are used instead. The PR's cost-calculator code change (route Together models with registry pricing away from the parameter-size buckets) is a behavior change this PR does not carry, so fix(cost): apply Together AI cache read pricing and per-model registry rates #38280 stays open for that part.
• feat(model_prices): add databricks kimi k3, databricks glm 5.2 and zai glm 5.3 #38415 (Databricks GLM 5.2, Kimi K3, Z.AI GLM-5.3): all rates re-verified against the Databricks and Z.AI pricing pages, absorbed with its test.
• fix(models): drop retired xAI models and mark multi-agent responses-only #38510 (xAI retired models, multi-agent Responses mode): confirmed against the current xAI model listing and multi-agent docs, absorbed with its registry test.
• fix(model_prices): add supports_native_structured_output to claude-haiku-4-5 direct API entries #31221 (Anthropic Haiku 4.5 supports_native_structured_output): both flags it adds are already on this branch, confirmed against https://platform.claude.com/docs/en/build-with-claude/structured-outputs, so it is superseded with no new fields to take.

Deliberately not applied

• feat(cost): add missing Fireworks AI Serverless model cost entries (#37274) #38218 (Fireworks additions): deepseek-v4-pro-0813 is absorbed here, its pricing now appears on the Fireworks serverless pricing table and matches the PR. qwen3p8-2p4t-a95b is dropped: neither the pricing table nor the model page publishes a serverless price for it, so its numbers are unverifiable.
• fix: set supports_vision to false for novita/openai/gpt-oss-120b #37585 (Novita GPT-OSS-120B vision): Novita's catalog API (https://api.novita.ai/v3/openai/models) reports input_modalities: [text, image] for openai/gpt-oss-120b, so the provider's own source contradicts setting supports_vision: false. Not applied.
• feat(model_prices): add deepinfra/moonshotai/Kimi-K2.5 to model registry #35776, feat(model_prices): add deepinfra/moonshotai/Kimi-K2.7-Code to model registry #35891 (DeepInfra Kimi K2.5 / K2.7 Code): both entries land in this PR with the prices and 262144 context the DeepInfra API publishes, instead of the zero prices and 32768 context those PRs carry.
• feat(models): add missing Together AI serverless models to the cost map #38230 (Together AI): almost entirely deprecation_date plus successor metadata, which a separate automation owns. Its Llama 3.3 70B Turbo price change still needs an official Together source, so nothing was taken from it and it stays open.
• feat: update EmpirioLabs models in model_prices_and_context_window.json #37972 (EmpirioLabs): ~60 new entries with no source, on a different base branch, each needing a per-model check against https://docs.empiriolabs.ai/models/. Not verified here, left open.
• Code or behavior PRs that only incidentally touch the JSON are untouched: fix: python 3.10 typing compatibility in compact editor and claude 3 cache pricing (#38076, #38056) #38189 (its Claude 3 1h cache registry fix is already carried here), feat(dashscope): add kimi-k3, MiniMax-M3, and deepseek-v4-pro-0813 to the cost map #38165, feat(friendli): auto-sync Friendli model metadata into price registry #35918, feat(providers): add Bitdeer AI as a JSON-configured provider #36734, fix(proxy): restore model pricing in /v2/model/info after a request #37747, test(batches): assert completed Vertex batch cost at batch rates #37443, refactor(model_prices): move fallback_generalizations under sample_spec, above all models #36416 and refactor(model_prices): move fallback_generalizations into its own file and remote URL #36422 (fallback-generalizations block plus the code that reads it), feat(vertex-ai): add Chirp 3 HD STT/TTS via gRPC streaming and WebSoc… #36233 (Vertex Chirp speech code), fix(bedrock-invoke): use native output_format for Claude 4.5+ models that support it #29898 (Bedrock output-format behavior), feat(models): add daily Together AI model registry sync script and workflow #38257 and feat(router): backfill missing cost fields from canonical entry for known models #30383 (no registry files).
• Mistral ids with no published price or card (labs-leanstral-1-5-1, mistral-ocr-3/-4 aliases, mistral-vibe-cli-, mistral-code-, voxtral-mini-realtime-2602, mistral-medium-3.5 dotted alias, voxtral-mini-latest, whose card is deprecated without a price) were not added.
• No speculative deprecation_date edits: only the Azure and Together dates above, each read from the provider's raw schedule page, and no entry gained a price or flag that its provider does not publish, so several models the providers list without public per-token pricing are still absent.

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/test_litellm/<your_test_file>.py -v
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Tests

tests/test_litellm/test_fireworks_serverless_model_costs.py (adapted from #38218, qwen assertions removed) pins the new Fireworks entry and proves the bare accounts/fireworks/models/... id resolves to it. tests/test_litellm/test_anthropic_sonnet_1hr_cache_pricing.py pins the Haiku 3 and Opus 3 1h cache writes in both maps, plus an invariant that every first-party Anthropic entry keeps its 1h cache write at 2x base input. Bedrock Sol and Cyber costing stays covered by the updated cross-region and Mantle Responses tests.

LITELLM_LOCAL_MODEL_COST_MAP=True uv run --no-sync pytest -q \
  tests/test_litellm/test_model_prices_schema.py tests/test_litellm/test_model_cost_aliases.py
33 passed

LITELLM_LOCAL_MODEL_COST_MAP=True .venv/bin/python -m pytest -q tests/test_litellm/test_fireworks_serverless_model_costs.py tests/local_testing/test_get_model_info.py
30 passed

LITELLM_LOCAL_MODEL_COST_MAP=True .venv/bin/python -m pytest -q tests/test_litellm/test_together_ai_model_metadata.py \
  tests/test_litellm/llms/databricks/test_databricks_cost_calculator.py tests/test_litellm/llms/xai/test_xai_model_registry.py
142 passed

python3 ci_cd/check_files_match.py
Passed! Files model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json match.

.venv/bin/python ci_cd/generate_model_prices_schema.py --check
model_prices_and_context_window.schema.json is in sync and model_prices_and_context_window.json validates against it

Link to Devin session: https://app.devin.ai/sessions/e79cabf82dec491aa7b23e502bcf6735

Link to Devin session: https://app.devin.ai/sessions/56ad8758edb14799ae7986746a98af63
Open in Devin Desktop: https://app.devin.ai/desktop/session/56ad8758edb14799ae7986746a98af63?variant=devin

---

Note

Medium Risk
Data-only changes directly drive cost calculation and budgets; wrong rates affect billing accuracy, though scope is limited to the static registry with no runtime logic in this diff.

Overview
Updates model_prices_and_context_window_backup.json (kept in sync with the primary cost map) so LiteLLM spend logging matches current provider pricing.

Rate and metadata fixes include Bedrock openai.gpt-5.6-sol (and US/Global mirrors) at published lower tiers, new openai.gpt-5.6-cyber, Gemini/Vertex 2.5 Flash cached-input rates, Azure gpt-4.1-nano retirement pushed to 2027-04-14, azure/sora-2 deprecation, Fireworks DeepSeek V4 pricing, and large DeepInfra / Novita price realignments plus capability flags (e.g. prompt caching).

New catalog entries cover dozens of Novita, DeepInfra, and W&B models, Databricks GLM 5.2 / Kimi K3, Mistral ministral/voxtral/medium/embed ids, Together cache-read fields and deprecation_date on retiring serverless models (successors pointed at DeepSeek-V4-Pro-0813), and zai/glm-5.3.

Removals / routing: legacy xAI grok-2/beta/vision entries dropped; grok-4.20 multi-agent models switch to mode: responses with /v1/responses. W&B token costs are corrected from implausible per-token magnitudes to per-token decimals with source URLs.

^{Reviewed by Cursor Bugbot for commit a0689f0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[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.}

[devin-ai-integration]

No action taken on #38207 — author is devin-ai-integration[bot] as expected, but the PR has no labels at all, so the required enterprise label is absent and it's out of scope. No GitHub or Linear changes made; no risk label applied.

[greptile-apps]

Greptile Summary

The PR updates model pricing and capability metadata across multiple providers and adds focused regression coverage. The previously reported Claude cache-rate coverage gap is resolved by exact assertions for Haiku 3 and Opus 3 in both registry maps, plus a first-party Anthropic pricing invariant.

• Adds and corrects provider model pricing, limits, capabilities, routing modes, and deprecation metadata.
• Keeps the primary and backup pricing registries synchronized.
• Adds regression tests for Anthropic cache writes, Fireworks pricing, Bedrock mappings, Databricks costs, Together metadata, and xAI registry behavior.

Confidence Score: 5/5

The PR appears safe to merge because the previously reported regression-coverage gap has been fixed and no blocking failure remains.

No blocking failure remains.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Updates the primary model registry with provider pricing, capabilities, context limits, routing modes, and deprecation metadata.
litellm/model_prices_and_context_window_backup.json	Mirrors the primary registry changes in the runtime backup map.
tests/test_litellm/test_anthropic_sonnet_1hr_cache_pricing.py	Fully addresses the prior coverage concern with exact Haiku 3 and Opus 3 assertions against both registry maps and a broader first-party invariant.
tests/test_litellm/test_fireworks_serverless_model_costs.py	Adds focused resolution and cost coverage for the new Fireworks DeepSeek model.
tests/test_litellm/llms/xai/test_xai_model_registry.py	Adds registry coverage for retired xAI entries and Responses-only multi-agent models.
tests/test_litellm/test_together_ai_model_metadata.py	Updates Together serverless and deprecated model expectations to match the registry changes.

_{Reviews (5): Last reviewed commit: "fix(model_prices): cap ministral-3-3b at..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[devin-ai-integration]

Added focused regression tests pinning the Claude 3 one-hour cache writes to 2x base input, so that gap is closed now.

[devin-ai-integration]

@greptileai

[devin-ai-integration]

The failing MCP cost test also fails on four unrelated branches with the same 1.35e-05 == 5.0 assertion, so it predates this PR

[mateo-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 310591f. Configure here.}

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_registry_audit_bedrock_sol_anthropic_1hr (a0689f0) with litellm_internal_staging (8ebcb3e)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (98d231c) during the generation of this report, so 8ebcb3e was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mateo-berri]

bugbot run

[mateo-berri]

@greptileai

[cursor]

Ministral 3B aliases use wrong context

Medium Severity

New mistral/ministral-3b-2512 and mistral/ministral-3b-latest aliases cap context at 131072, while this same change lifts sibling mistral/ministral-3-3b-2512 to 262144 from the same 3B card those aliases cite. Calls above 128k on the short ids can be rejected even though the model accepts 256k.

Additional Locations (1)

• litellm/model_prices_and_context_window_backup.json#L31557-L31563

 

^{Reviewed by Cursor Bugbot for commit 5d26ae0. Configure here.}

[mateo-berri]

Mistral's models API caps ministral-3b-2512 at 131072; only the 3-3b entry was wrong, now lowered to match in a0689f0

[cursor]

Novita Llama 3.3 context window slashed

High Severity

novita/meta-llama/llama-3.3-70b-instruct drops max_input_tokens, max_output_tokens, and max_tokens from a 128k-class window to 12288. That is the only Novita correction that shrinks input context, and Llama 3.3 70B is a 128k model. Gateway context checks will refuse ordinary long prompts that Novita still serves.

 

^{Reviewed by Cursor Bugbot for commit 5d26ae0. Configure here.}

[mateo-berri]

Novita's own models API reports context_size 12288 for meta-llama/llama-3.3-70b-instruct; the entry mirrors the provider's current serving limit

[mateo-berri]

bugbot run

[mateo-berri]

@greptileai

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

2 issues from previous reviews remain unresolved.

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit a0689f0. Configure here.}

[mateo-berri]

@greptileai

[mateo-berri]

LGTM