fix(models): registry audit: new Together/Fireworks/Gemini/Mistral/xAI models, xai retirement repricing, bedrock grok-4.6 caching, deprecation dates

[devin-ai-integration]

TLDR

Problem this solves:

• Registry missed new Together, Fireworks, Gemini, Mistral, xAI models
• Registry had stale xAI pricing after the May 15, 2026 retirements
• Bedrock Grok 4.6 wrongly claimed explicit prompt caching support
• Registry carried a Gemini API model the API does not serve
• Some deprecation dates missing for OpenAI, Gemini, and xAI models

How it solves it:

• Adds Together/Fireworks GLM-5.3, Gemini Gemma 4, xAI video, Mistral aliases
• Reprices every retired xAI slug at its redirect target's rates
• Keeps Bedrock Grok 4.6 implicit cache-read billing, drops the explicit caching flag
• Removes gemini/gemini-live-2.5-flash-native-audio (the Gemini API 404s it)
• Adds missing deprecation_date values from official provider pages

User Flow

Before: a developer calling a retired xAI slug gets billed at stale rates

1. They send POST https://litellm-domain/v1/chat/completions with "model": "xai/grok-3-mini"
2. xAI serves the request via grok-4.3 (the retirement redirect) but spend logs price it at the old grok-3-mini rates ($0.30/M input)
3. https://litellm-domain/ui/?page=logs shows spend that does not match the xAI invoice

After: the same call is priced at the grok-4.3 redirect rates

1. They send the same POST https://litellm-domain/v1/chat/completions with "model": "xai/grok-3-mini"
2. Spend logs price it at grok-4.3 rates ($1.25/M input, $2.50/M output, $0.20/M cached, 200K tier)
3. https://litellm-domain/ui/?page=logs matches the xAI invoice

Relevant issues

Supports #26900 (deprecation metadata)

Linear ticket

Resolves LIT-6472

Changes in the review pass (2026-08-29, third pass)

xAI, per https://docs.x.ai/developers/migration/may-15-retirement and https://docs.x.ai/developers/pricing:

• Repriced the ten retired slugs the first pass missed (grok-3-beta, grok-3-fast-beta, grok-3-fast-latest, grok-4-1-fast, and the six grok-3-mini variants) at grok-4.3 rates, dropping their stale 128K tiers
• Added their deprecation dates: 2026-02-28 for the grok-3-mini family, matching the date its sibling entries already carry from xAI's announcement, and 2026-05-15 for the rest
• The redirected-slug test now covers all 21 repriced slugs, and its coverage check fails when a dated retired chat slug is missing from the list
• Removed output_cost_per_second_1080p from xai/grok-imagine-video-1.5: the pricing page lists one flat $0.08/second rate with no per-resolution pricing

Bedrock Grok 4.6, per the AWS model card:

• Restored cache_read_input_token_cost ($0.55/M us, $0.50/M global) on both entries: AWS documents implicit prompt caching with cache-read billing for this model
• supports_prompt_caching stays false because that flag gates explicit cachePoint breakpoints, which Bedrock rejects for Grok; without the rate, implicit cache hits would bill at the full input rate
• The cache-control hook test now asserts both halves

Gemini:

• Removed gemini/gemini-live-2.5-flash-native-audio: the Gemini API returns 404 for it on v1beta, v1alpha, and v1, and the Live websocket rejects the model. The Vertex entry stays, and the realtime test now asserts the model remains Vertex-only

Groq, per https://console.groq.com/docs/model/qwen3.8-27b:

• groq/qwen/qwen3.8-27b gains supports_function_calling, supports_response_schema, and supports_tool_choice (the doc lists Tool Use, JSON Object Mode, and JSON Schema Mode)

Together:

• together_ai/zai-org/GLM-5.3 added to the serverless coverage test list
• GLM-5.2 and GLM-5.3-Flash still carry the sync's inflated max_output_tokens; LIT-6489 tracks fixing the sync's context-length fallback and sweeping those entries

Changes in the second pass (2026-08-29)

Together AI, per https://docs.together.ai/docs/serverless-models and https://www.together.ai/models/glm-5-3 (absorbed from #38733 with corrections):

• Added together_ai/zai-org/GLM-5.3 ($1.40/M in, $4.40/M out, $0.26/M cached, 1048575 ctx). Corrected feat(models): sync together_ai model registry #38733's max_output_tokens from 1048575 to 128000 (the documented output ceiling) and added the capability flags feat(models): sync together_ai model registry #38733 flagged as missing (tools, reasoning, response schema, prompt caching)
• together_ai/Qwen/Qwen3.8-2.4T-A95B repriced: in 2e-06 -> 2.5e-06, out 6e-06 -> 6.25e-06, cache read 2.5e-07 -> 5e-07

Gemini, per https://ai.google.dev/gemini-api/docs/pricing and GET generativelanguage.googleapis.com/v1beta/models:

• Added gemini/nano-banana-pro-preview (API alias for Gemini 3 Pro Image; priced at the published gemini-3-pro-image rates, 131072 input limit from the models API)
• Added gemini/gemma-4-26b-a4b-it and gemini/gemma-4-31b-it (free of charge per the pricing page; 262144/32768 limits from the models API)

Mistral, per GET api.mistral.ai/v1/models (the API's own aliases field): 14 official alias entries cloned from their canonical registry entries: mistral-medium-3.5, mistral-vibe-cli-latest, mistral-vibe-cli-with-tools, mistral-vibe-cli-fast, mistral-code-latest, mistral-code-fim-latest, mistral-code-agent-latest, mistral-ocr-3, mistral-ocr-3-0, mistral-ocr-4, voxtral-mini-latest, voxtral-mini-realtime-2602, voxtral-mini-realtime-latest, labs-leanstral-1-5-1

Fireworks AI, per https://docs.fireworks.ai/serverless/pricing and GET api.fireworks.ai/inference/v1/models:

• Added fireworks_ai/accounts/fireworks/models/glm-5p3 ($1.40/M in, $4.40/M out, $0.26/M cached; 1048576 ctx from the models API, 128000 output per the GLM-5.3 model docs)
• Added fireworks_ai/accounts/fireworks/models/qwen3-embedding-8b ($0.10/M input per the embeddings pricing table, 40960 ctx)

xAI, per https://docs.x.ai/developers/models/grok-imagine-video and https://docs.x.ai/developers/models/grok-imagine-video-1.5:

• Added xai/grok-imagine-video ($0.05/second output) and xai/grok-imagine-video-1.5 ($0.08/second output)

Chose not to change:

• Fireworks glm-5p3-flash and the kimi-k2p6-turbo router: in the models API but no row on the pricing page
• xAI video per-resolution rates: the pricing page lists flat per-second rates only, so no resolution-specific cost fields are recorded
• Gemini antigravity-preview-*, deep-research-*: billed at underlying model rates, no fixed per-token price to record; aqa: no published pricing
• Mistral Voxtral TTS aliases already present; OCR/transcribe per-page and per-minute rates re-checked against https://docs.mistral.ai/inference/pricing, already correct

Checked, nothing missing or wrong: OpenAI (models API vs registry; computer-use-preview and current snapshots present), Anthropic (models API; entries current). DeepSeek model-list API rejected the configured key, so that provider was not audited this run; Groq was checked against its public docs in the review pass. OpenRouter/Bedrock/Azure/Cohere/Perplexity/Nebius/Voyage were not reached this run; they stay on the audit backlog.

Changes in the first pass (2026-08-29)

xAI May 15, 2026 retirements, per https://docs.x.ai/developers/migration/may-15-retirement and https://docs.x.ai/developers/models/grok-4.3 and https://docs.x.ai/developers/models/grok-build-0.1: retired chat slugs (grok-3, grok-4, grok-4-0709, grok-4-1-fast-, grok-4-fast-) now priced at grok-4.3 rates (1.25e-06 in, 2.5e-06 out, 2e-07 cache read, 200K tier at 2.5e-06/5e-06/4e-07); stale 128K tiers removed. grok-code-fast* aliases priced at grok-build-0.1 rates (1e-06 in, 2e-06 out, 2e-07 cache read)

Bedrock Grok 4.6 (us.xai.grok-4.6, global.xai.grok-4.6): explicit prompt caching flag removed; Bedrock rejects cachePoint breakpoints for this model. The review pass above restores the implicit cache-read billing this initially dropped

OpenAI deprecation dates, per https://platform.openai.com/docs/deprecations: deprecation_date: 2026-12-01 added to the 15 sized/quality gpt-image-1.5-2025-12-16 variants (the base snapshot entry already had it)

Gemini, per https://ai.google.dev/gemini-api/docs/deprecations: gemini/gemini-omni-flash-preview gets deprecation_date: 2026-09-30 (replacement gemini-omni-1.1-flash). The vertex_ai copy was left alone because the Vertex deprecations page does not list it

Verified as current, no change needed: Anthropic dates against https://docs.anthropic.com/en/docs/about-claude/model-deprecations, other Gemini dates, OpenAI exact-name dates

Absorbed from other PRs

• feat(models): sync together_ai model registry #38733 (Together sync): fully absorbed with the GLM-5.3 output-limit correction and capability flags described above
• fix(model_prices): bedrock Grok 4.6 does not support prompt caching #38699 (Bedrock Grok 4.6 caching): absorbed, then corrected in the review pass: the explicit caching flag stays off while the implicit cache-read rate returns
• fix(model_prices): price the redirected xAI slugs at grok-4.3 rates #38730 (xAI redirected slug repricing): absorbed with a correction, its test priced grok-code-fast* as grok-4.3 redirects but xAI docs alias them to grok-build-0.1

Dropped as unverifiable

• fix(model_prices): annotate retired xAI grok-2-era models and mark grok-4.20-multi-agent Responses-only #38526's deprecation_date: 2026-02-28 on Grok-2-era xAI models: the announcement URL it cites returns 404 and no live official xAI page states that date. Its mode: responses changes are already on the base branch

Not absorbed (code/behavior or structural PRs that only incidentally touch the JSON): #36233, #29898, #30383, #36416, #36422

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• 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. Leave the suites (make test-unit-*, make test-unit) to CI: it finishes in ~15 minutes where a laptop takes an hour or more
• 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 (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Registry data change; validated at the PR tip:

1. python3 ci_cd/check_files_match.py: Passed, both registry copies match
2. 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 tests/test_litellm/llms/xai/test_xai_cost_calculator.py tests/test_litellm/llms/xai/test_xai_redirected_slug_pricing.py tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py tests/test_litellm/test_together_ai_model_metadata.py tests/test_litellm/integrations/test_anthropic_cache_control_hook.py: all passed

Type

🐛 Bug Fix

Caveats (if any)

Medium

• Retired xAI slug spend rises to redirect-target rates, matching xAI invoices
• DeepSeek, OpenRouter, Bedrock, Azure, Cohere audits stay on the backlog

Low

• gemini/gemini-live-2.5-flash-native-audio entry removed; the API already 404s it
• GLM-5.2 and GLM-5.3-Flash output ceilings still inflated (LIT-6489)

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

Link to Devin session: https://app.devin.ai/sessions/1eb47e52d5a0484985cc0e51592871d5
Open in Devin Desktop: https://app.devin.ai/desktop/session/1eb47e52d5a0484985cc0e51592871d5?variant=devin

[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

[devin-ai-integration]

Exited without changes: #38693 has no labels, so it lacks the required enterprise label (author/repo checks did pass). No risk label or Linear routing applied.

[greptile-apps]

Greptile Summary

The PR updates model registry entries and corresponding regression coverage across several providers.

• Adds and revises model pricing, capability, context-window, and deprecation metadata.
• Updates Gemini realtime, xAI redirect pricing, Together metadata, and Bedrock caching tests.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Updates the primary model registry with new models, corrected pricing, capabilities, caching metadata, and deprecation dates.
litellm/model_prices_and_context_window_backup.json	Mirrors the primary registry changes in the bundled backup catalog.
tests/test_litellm/llms/gemini/realtime/test_gemini_realtime_transformation.py	Updates realtime model fixtures and verifies that the native-audio model remains available only through Vertex AI.
tests/test_litellm/llms/xai/test_xai_redirected_slug_pricing.py	Adds comprehensive coverage for retired xAI model slugs and their redirect-target pricing.
tests/test_litellm/integrations/test_anthropic_cache_control_hook.py	Verifies Bedrock Grok cache-read billing without enabling explicit prompt-cache breakpoints.
tests/test_litellm/test_together_ai_model_metadata.py	Extends Together AI serverless model metadata coverage.

_{Reviews (3): Last reviewed commit: "fix(registry): correct xai retired slug ..." | Re-trigger Greptile}

[greptile-apps]

Catalog Test Misses Metadata

This test checks only entry presence and gemini_native_audio, so regressions in the newly added endpoints, modalities, pricing, limits, or capability flags will still pass CI. The assertion also exceeds the repository's 120-character line limit.

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[mateo-berri]

Superseded in 2bd7b58: the gemini/ entry was removed as unserved, this test now asserts the vertex-only invariant, and every line fits 120 chars

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ mateo-berri
❌ devin-ai-integration[bot]
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing devin/1787944648-registry-audit-rolling (2bd7b58) with litellm_internal_staging (194a3cc)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (e0ed0a4) during the generation of this report, so 194a3cc was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mateo-berri]

@greptileai

[mateo-berri]

@greptileai

[mateo-berri]

Superseded by #38804, the same commits on a litellm_ branch so CircleCI actually builds them. Review continues there.