Add Claude Fable 5 across Anthropic, Bedrock, Vertex AI, and Azure AI

[mateo-berri]

Relevant issues

Linear ticket

Pre-Submission checklist

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

• I have added meaningful tests
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Delays in PR merge?

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

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Screenshots / Proof of Fix

Supersedes #30062 (same commit; branch renamed to follow the litellm_ prefix convention).

Proxy started against this branch's cost map (the hosted map on main does not have the model yet, so LITELLM_LOCAL_MODEL_COST_MAP=True makes the proxy use the bundled copy; once this merges, the hosted map picks it up and the flag is unnecessary):

LITELLM_LOCAL_MODEL_COST_MAP=True python litellm/proxy/proxy_cli.py --config /tmp/fable_proof/config.yaml --port 4000

with deployments anthropic/claude-fable-5, bedrock/us.anthropic.claude-fable-5, vertex_ai/claude-fable-5, and azure_ai/claude-fable-5.

First-party Anthropic, with reasoning_effort translated to adaptive thinking and real spend tracked:

curl -s -D /tmp/h.txt http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "claude-fable-5", "messages": [{"role": "user", "content": "Reply with exactly: fable-5 via anthropic works"}], "reasoning_effort": "low", "max_tokens": 2000}' \
  | jq '{model, content: .choices[0].message.content, prompt_tokens: .usage.prompt_tokens, completion_tokens: .usage.completion_tokens}'; grep -i "^x-litellm-response-cost:" /tmp/h.txt

{
  "model": "claude-fable-5",
  "content": "fable-5 via anthropic works",
  "prompt_tokens": 25,
  "completion_tokens": 13
}
x-litellm-response-cost: 0.0009000000000000001

The cost header matches the new pricing exactly: 25 input tokens at $10/MTok plus 13 output tokens at $50/MTok is $0.0009

Streaming with reasoning_effort: "xhigh" (Fable 5 rejects the legacy budget_tokens shape with a 400, so a 200 here proves the adaptive thinking plus output_config.effort wire format is being sent):

curl -s -N http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "claude-fable-5", "messages": [{"role": "user", "content": "In one short sentence, what is 17*23? Show the result only."}], "reasoning_effort": "xhigh", "stream": true, "stream_options": {"include_usage": true}, "max_tokens": 30000}' | tail -4

data: {"id":"chatcmpl-3211eb1a-...","model":"claude-fable-5","choices":[{"index":0,"delta":{}}],"usage":{"completion_tokens":12,"prompt_tokens":26,"total_tokens":38,...}}

data: [DONE]

reasoning_effort: "none" also returns 200; this matters because Fable 5 is stricter than Opus 4.8 and 400s on an explicit thinking: {"type": "disabled"}, so LiteLLM must omit the param entirely, which it does:

{"content": "no-thinking path works", "completion_tokens": 20}

Sampling params (reported by @emerzon on this PR)

Fable 5 rejects top_p, top_k, and any temperature other than 1 with a 400; Opus 4.7 and 4.8 behave identically (verified against the live API for all three; Opus 4.6 accepts everything). Without drop_params, the raw provider error is now replaced by a clean client-side 400 that explains the opt-in:

curl -s http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "claude-fable-5", "messages": [{"role": "user", "content": "Reply with exactly: ok"}], "temperature": 0.5, "max_tokens": 2000}'

{"error":{"message":"litellm.UnsupportedParamsError: claude-fable-5 does not support temperature=0.5. Only temperature=1 is supported. To drop unsupported params, set `litellm.drop_params = True`. ...","code":"400"}}

With litellm_settings: drop_params: true in the proxy config, the same request plus top_p succeeds because both params are dropped before the request goes out:

curl -s http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "claude-fable-5", "messages": [{"role": "user", "content": "Reply with exactly: dropped and ok"}], "temperature": 0.5, "top_p": 0.9, "max_tokens": 2000}' \
  | jq -c '{content: .choices[0].message.content}'

{"content":"dropped and ok"}

temperature: 1 (the API default) still passes through on the no-drop-params proxy and succeeds:

{"content":"temp1 ok"}

top_k is a provider-specific kwarg that bypasses the OpenAI param mapping entirely, so it is gated at the transform boundary instead (reported by Greptile on this PR). Same behavior on the live proxy, no drop_params:

curl -s http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "claude-fable-5", "messages": [{"role": "user", "content": "Say ok"}], "top_k": 40, "max_tokens": 16}'

{"error":{"message":"litellm.UnsupportedParamsError: claude-fable-5 does not support top_k=40. To drop unsupported params, set `litellm.drop_params = True`. ...","code":"400"}}

and on the drop_params: true proxy the same request with top_k: 40 is stripped and succeeds:

{"content":"dropped and ok"}

Partner clouds

Bedrock routing reaches the real model; the remaining blocker is account provisioning, not LiteLLM. Fable 5 on Bedrock requires opting the AWS account into provider data sharing via the Data Retention API (no console UI at launch per the AWS model card), which our CI account has not done:

curl -s http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model": "bedrock-claude-fable-5", "messages": [{"role": "user", "content": "hi"}], "max_tokens": 100}'

BedrockException - {"message":"The model returned the following errors: data retention mode 'default' is not available for this model"}

(this is from the global. profile; us.anthropic.claude-fable-5 is not yet listed for this account and returns "provided model identifier is invalid", and the base id returns the standard "use an inference profile" error, both consistent with a day-one rollout)

Azure Foundry authenticates and routes correctly; the CI resource just has no claude-fable-5 deployment yet (same current state as claude-opus-4-8 there). Creating the deployment in the Foundry portal under Models + endpoints is the one manual step left:

Azure_aiException - {"error":{"code":"DeploymentNotFound","message":"The API deployment claude-fable-5 does not exist. ..."}}

Vertex could not be exercised live from this sandbox because the ADC credential in the environment has an expired refresh token ("Reauthentication is needed. Please run gcloud auth application-default login"); the model id claude-fable-5 is confirmed against the Vertex docs and the entries mirror the verified Opus 4.8 shape

To re-run Bedrock/Vertex/Azure once provisioned: opt the AWS account into provider_data_sharing, refresh the gcloud ADC credential, create the Foundry deployment, then re-run the curls above with the corresponding model names. The e2e grid cells in tests/llm_translation/reasoning_effort_grid/grid_spec.py carry fail_reason markers for exactly these three gaps and should have those markers removed at the same time

Type

🆕 New Feature

Changes

Adds Claude Fable 5 (released today; $10/$50 per MTok, 1M context, 128K max output, adaptive thinking only) to the model cost map for all four platforms that serve it: claude-fable-5 (Anthropic API, with the inference_geo: us 1.1x multiplier in provider_specific_entry and deliberately no fast key since Fable 5 has no fast mode), anthropic.claude-fable-5 plus global./us./eu. inference profiles (Bedrock converse, geo profiles at the documented 10% regional premium; AWS lists no au/apac/jp profiles for this model), vertex_ai/claude-fable-5 and @default (Vertex publishes the bare id with no date suffix), and azure_ai/claude-fable-5 (Microsoft Foundry serves Fable 5 with the full 1M context window, unlike Opus 4.8 which is capped at 200k there)

All entries carry supports_adaptive_thinking, which is what makes LiteLLM emit thinking: {"type": "adaptive"} plus output_config.effort instead of the legacy budget_tokens shape that Fable 5 rejects. Cache pricing uses the standard multipliers: $12.50/MTok 5m write, $20/MTok 1h write, $1/MTok read

Also registers anthropic.claude-fable-5 in BEDROCK_CONVERSE_MODELS, adds the model to the setup wizard, and extends the reasoning effort e2e grid with Fable 5 rows for all four providers (Bedrock/Vertex/Azure cells xfail with documented provisioning reasons until the CI accounts are set up, mirroring how Opus 4.8 was rolled in)

A follow-up commit fixes sampling param handling reported on this PR: Fable 5, Opus 4.7, and Opus 4.8 removed sampling params, so the API 400s on top_p, top_k, and any temperature other than 1, but the Anthropic and Bedrock converse transformations forwarded temperature/top_p unconditionally, making drop_params a no-op for them. They now mirror the GPT-5/o-series handling: temperature=1 passes through, other values and any top_p are dropped when drop_params is set, and without drop_params a clean client-side UnsupportedParamsError explains the opt-in instead of surfacing the raw provider error. The shared helper lives on AnthropicModelInfo, so the Anthropic, Vertex, Bedrock invoke, and Bedrock converse paths all get it

A second follow-up commit addresses the Greptile review of that fix. The restriction is now declared in the cost map as supports_sampling_params: false on every Fable 5 / Opus 4.7 / Opus 4.8 entry in both map files (28 entries each; the perplexity route is exempt because it is OpenAI-compatible and Perplexity maps sampling params upstream) and read back through a tri-state lookup, with the substring check kept only as a fallback for provider-routed ids whose hosted map entries predate the flag, the same layering supports_adaptive_thinking uses. It also gates top_k, which bypasses map_openai_params as a provider-specific kwarg: once at the shared AnthropicConfig.transform_request boundary (covering the direct, Bedrock invoke, Vertex, and Azure paths) and once in the Bedrock converse _handle_top_k_value path, with drop_params threaded through the converse transform helpers. The same commit fixes the reasoning effort grid cell count assertion (29 x 11 after the four Fable 5 rows), which was the one failing CircleCI job (llm_translation_testing) on the previous commit

A third follow-up commit closes a gap Cursor Bugbot found in that gating: _handle_top_k_value used a truthiness check, so top_k=0 was popped and silently discarded on the converse path while the shared transform_request boundary treats 0 as present. The check is now is not None, so top_k=0 raises or drops on restricted models and is forwarded elsewhere, identical to the other paths, with regression tests for both cases

Tests: tests/test_litellm/test_claude_fable_5_config.py validates pricing, capabilities, regional premiums, backup-map parity, Bedrock converse registration, provider resolution, and adaptive-thinking detection through bedrock/, bedrock/invoke/, vertex_ai/, and azure_ai/ routed ids. One notable assertion: provider_specific_entry == {"us": 1.1} guards against a future copy-paste of the Opus entries adding a fast multiplier that would misprice speed: "fast" requests on a model that has no fast mode. The sampling param fix adds regression tests in both mapped transformation test files covering dropped, raised, temperature=1 passthrough, and 4.6-still-forwarded cases. The Greptile follow-up adds top_k coverage on both transform paths (dropped, raised, still-forwarded), a test proving the gating is driven by the map flag rather than name matching, and a map-level test asserting every affected entry carries supports_sampling_params: false in both files

Known pre-existing issue, not touched here: model_prices_and_context_window.json and the bundled backup already diverge on this branch in the snowflake entries; the new Fable 5 entries are byte-identical in both files and the new test asserts that parity for these entries

https://claude.ai/code/session_01MZarYYT3aS7DxaNjoax6Gm

[emerzon]

litellm.BadRequestError: AnthropicException - b'{"error":{"message":"{\\"type\\":\\"error\\",\\"error\\":{\\"type\\":\\"invalid_request_error\\",\\"message\\":\\"`temperature` is deprecated for this model.\\"},\\"request_id\\":\\"req_vrtx_011CbtVzDVs3puMnkMSYYx94\\"}. Received Model Group=claude-fable-5\\nAvailable Model Group Fallbacks=None","type":"None","param":"None","code":"400"}}'
Model: claude-fable-5 (openai)

I know the PR is still WIP, but please consider this - temperature needs to be dropped when drop_params is set

[mateo-berri]

@emerzon what client are you using? Codex? OpenCode? Claude Code? Something else?

[mateo-berri]

@emerzon confirmed legit, thanks for the report. Fable 5 rejects top_p, top_k, and any temperature other than 1 with a 400, and Opus 4.7/4.8 behave the same way (verified against the live API for all three; 4.6 still accepts everything). Fixed in 1a080a2: with drop_params set, temperature and top_p are now dropped before the request goes out; without it, LiteLLM raises a client-side UnsupportedParamsError that explains the opt-in instead of surfacing the raw provider error. temperature=1 still passes through since the API accepts it. The fix covers the Anthropic, Vertex, and Bedrock paths; proof curls are in the PR description

[codecov]

Codecov Report

❌ Patch coverage is 95.45455% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/llms/anthropic/common_utils.py	93.75%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

Adds Claude Fable 5 ($10/$50 per MTok, 1M context, 128K output, adaptive-thinking-only) to LiteLLM's model registry across Anthropic, Bedrock (including four geo inference profiles), Vertex AI, and Azure AI, and fixes sampling-param handling for Fable 5, Opus 4.7, and Opus 4.8 — all three models reject temperature≠1, top_p, and top_k with a 400.

• All 8 cost-map entries carry supports_adaptive_thinking: true (required for the adaptive thinking wire format) and supports_sampling_params: false; all 28 affected Fable 5/Opus 4.7/4.8 entries in both map files are flagged.
• _apply_sampling_param now gates temperature/top_p in map_openai_params and top_k at transform_request on both the Anthropic/Vertex/Azure and Bedrock converse paths, dropping with drop_params or raising a clean UnsupportedParamsError without it.
• A pre-existing truthiness bug in _handle_top_k_value (if val_top_k → if val_top_k is not None) is fixed so top_k=0 is no longer silently discarded on the converse path.

Confidence Score: 5/5

Safe to merge; changes are additive model registrations plus a sampling-param fix on shared Anthropic/Bedrock/Vertex/Azure transform paths that was already broken before this PR (the provider returned a 400 unconditionally).

All Fable 5 entries are correctly flagged in both cost map files, the sampling-param gating is driven by the map flag with a name-based fallback documented as temporary, and the drop_params plumbing is threaded consistently across all four provider paths. The top_k=0 truthiness fix is a genuine correctness improvement. Tests are thorough and mock-only in the unit test directories.

No files require special attention; the shared transform paths in common_utils.py and converse_transformation.py are the highest-impact changes but are well-tested.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/common_utils.py	Refactors model capability lookups into _get_model_capability (tri-state) and _model_map_lookup_candidates; adds _apply_sampling_param and _supports_sampling_params helpers that gate temperature/top_p/top_k for models with supports_sampling_params: false, with a name-based fallback for pre-flag entries.
litellm/llms/anthropic/chat/transformation.py	Wires _apply_sampling_param into map_openai_params for temperature/top_p, and gates top_k at the transform_request boundary shared by the direct Anthropic, Bedrock invoke, Vertex, and Azure paths.
litellm/llms/bedrock/chat/converse_transformation.py	Threads drop_params through transform_request → _prepare_request_params → _handle_top_k_value; fixes val_top_k truthiness bug (if val_top_k → if val_top_k is not None) that silently dropped top_k=0; applies _apply_sampling_param gating for both top_k and temperature/top_p on the Bedrock converse path.
model_prices_and_context_window.json	Adds 8 Claude Fable 5 entries (Anthropic base, 4 Bedrock profiles, Vertex, Vertex @default, Azure AI) and propagates supports_sampling_params: false to all 28 Fable 5 / Opus 4.7 / Opus 4.8 entries; pricing ($10/$50 MTok) and flags verified correct.
litellm/model_prices_and_context_window_backup.json	Byte-identical mirror of the root cost map additions; parity is asserted by test_fable_5_present_in_bundled_backup.
litellm/constants.py	Registers anthropic.claude-fable-5 in BEDROCK_CONVERSE_MODELS, consistent with other recently-added models that lack versioned IDs.
tests/test_litellm/test_claude_fable_5_config.py	New test file validating pricing, capability flags, geo premium correctness, backup parity, Bedrock registration, provider resolution, adaptive-thinking detection across all provider-routed ids, and presence of supports_sampling_params: false on every affected entry in both cost map files.
tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py	Adds regression tests for temperature/top_p drop and raise on the Anthropic path, temperature=1 passthrough, model-map-flag-driven gating, top_k drop/raise/forward at transform_request, and forwards for models that still accept sampling params.
tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py	Adds parallel regression tests for the Bedrock converse path, including a dedicated top_k=0 test that covers the truthiness-check fix; all new tests are mock-only with no network calls.
tests/llm_translation/reasoning_effort_grid/grid_spec.py	Adds Fable 5 rows for all four providers to the e2e reasoning effort grid; Bedrock/Vertex/Azure cells carry fail_reason markers documenting provisioning blockers, consistent with how Opus 4.8 was rolled in.
tests/llm_translation/reasoning_effort_grid/test_reasoning_effort_grid.py	Updates the cell-count assertion from 25×11 to 29×11 to account for the four new Fable 5 rows.

_{Reviews (4): Last reviewed commit: "fix(bedrock): gate top_k=0 on converse t..." | Re-trigger Greptile}

[mateo-berri]

@greptileai

[mateo-berri]

bugbot run

[mateo-berri]

@greptileai

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Converse skips top_k zero gating
  • Replaced the truthiness check in _handle_top_k_value with if val_top_k is not None, so top_k=0 now enters the same _apply_sampling_param gating as any other value and matches the Anthropic boundary's behavior.

_{You can send follow-ups to the cloud agent here.}

^{Reviewed by Cursor Bugbot for commit b475cf7. Configure here.}

[cursor]

Converse skips top_k zero gating

Low Severity

Bedrock converse only runs sampling gating for top_k when val_top_k is truthy, so top_k=0 is popped and dropped without calling _apply_sampling_param. The shared Anthropic transform_request path treats 0 as present and raises or drops consistently. Converse can silently omit top_k=0 on models that removed sampling params instead of matching that behavior.

Additional Locations (1)

• litellm/llms/anthropic/chat/transformation.py#L1985-L1995

 

^{Reviewed by Cursor Bugbot for commit b475cf7. Configure here.}

[mateo-berri]

@greptileai

[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
❌ cursoragent
_{You have signed the CLA already but the status is still pending? Let us recheck it.}