feat(anthropic): add Claude Fable 5.1 by mateo-berri · Pull Request #39151 · BerriAI/litellm

mateo-berri · 2026-09-01T18:29:30Z

TLDR

Problem this solves:

Claude Fable 5.1 (claude-fable-5-1) is unknown to LiteLLM
Requests with reasoning_effort get rejected with a 400
Cached and uncached requests are billed at $0
Forced tool calls get an opaque provider 400

How it solves it:

Adds the Anthropic cost map entry with Fable 5.1 pricing and capabilities
Cache reads priced at 0.025x input, a quarter of Fable 5
New supports_forced_tool_use: false flag gates tool_choice any/tool
Without drop_params that is a clear client-side 400, with it a downgrade to auto
Lists the model in the setup wizard and reasoning-effort grid
Registers the new flag in the cost map JSON schema

User Flow

Before: a developer pointing their app at Fable 5.1 gets 400s and $0 spend

They add anthropic/claude-fable-5-1 to the proxy model list and send POST http://localhost:4000/v1/chat/completions with "reasoning_effort": "low"
They get HTTP 400 saying anthropic does not support parameters: ['reasoning_effort']
They drop reasoning_effort and resend with a cached system prompt: HTTP 200, but x-litellm-response-cost is 0 and https://litellm-domain/ui/?page=logs shows the request at $0 spend
They send the same route with tools and "tool_choice": "required": HTTP 400 quoting the raw provider error tool_choice: type "tool" and "any" are not supported for this model

After: the same requests are priced and the forced tool call fails clearly or downgrades

They add anthropic/claude-fable-5-1 to the proxy model list and send POST http://localhost:4000/v1/chat/completions with "reasoning_effort": "low"
They get HTTP 200 with the answer and x-litellm-response-cost set (input $10/MTok, output $50/MTok)
They resend with a cached system prompt: the first call bills the cache write at $12.50/MTok and the second bills the cache read at $0.25/MTok, and the logs page shows real spend
They send the same route with tools and "tool_choice": "required": HTTP 400 saying claude-fable-5-1 does not support forced tool use and pointing them at tool_choice="auto" or drop_params
On a deployment with drop_params: true, the same request gets HTTP 200 with a get_weather tool call and finish_reason: tool_calls

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
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

Shared setup, real Anthropic API calls through a local proxy:

# fable_5_1_config.yaml
model_list:
  - model_name: claude-fable-5-1
    litellm_params:
      model: anthropic/claude-fable-5-1
      api_key: os.environ/ANTHROPIC_API_KEY
  - model_name: claude-fable-5-1-drop-params
    litellm_params:
      model: anthropic/claude-fable-5-1
      api_key: os.environ/ANTHROPIC_API_KEY
      drop_params: true

general_settings:
  master_key: sk-1234

python litellm/proxy/proxy_cli.py --config fable_5_1_config.yaml --port 4000

cache_body.json used by the prompt caching case (system prompt is 89 "Rule N: ..." sentences, about 3.4k tokens, prefixed with a fresh nonce per run so the first call is a cache write):

{
  "model": "claude-fable-5-1",
  "max_tokens": 50,
  "messages": [
    {"role": "system", "content": [{"type": "text", "text": "Session 3b46abb9. Rule 1: ... Rule 89: ...", "cache_control": {"type": "ephemeral"}}]},
    {"role": "user", "content": "Reply with the single word OK."}
  ]
}

Before (`f2a4172`)

Chat completion with reasoning_effort

Command:

curl -s http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"claude-fable-5-1","messages":[{"role":"user","content":"In one sentence, what model are you?"}],"reasoning_effort":"low","max_tokens":200}' -D -

Output (HTTP 400):

{"error": {"message": "litellm.UnsupportedParamsError: anthropic does not support parameters: ['reasoning_effort'], for model=claude-fable-5-1. To drop these, set `litellm.drop_params=True` or for proxy:\n\n`litellm_settings:\n drop_params: true`\n. \n If you want to use these params dynamically send allowed_openai_params=['reasoning_effort'] in your request.. Received Model Group=claude-fable-5-1\nAvailable Model Group Fallbacks=None", "type": "None", "param": null, "code": "400"}}
x-litellm-response-cost: 0

Forced tool_choice without drop_params

Command:

curl -s -w "\nHTTP %{http_code}\n" http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"claude-fable-5-1","messages":[{"role":"user","content":"What is the weather in Paris?"}],"tools":[{"type":"function","function":{"name":"get_weather","description":"Get weather for a city","parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}}],"tool_choice":"required","max_tokens":300}'

Output, the raw provider error after a round trip to Anthropic:

{"error":{"message":"litellm.BadRequestError: AnthropicException - {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"tool_choice: type \\\"tool\\\" and \\\"any\\\" are not supported for this model.\"},\"request_id\":\"req_011CedBqNQu6djk3P3sSGdTe\"}. Received Model Group=claude-fable-5-1\nAvailable Model Group Fallbacks=None","type":null,"param":null,"code":"400"}}
HTTP 400

Forced tool_choice with drop_params

Command:

curl -s -w "\nHTTP %{http_code}\n" http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"claude-fable-5-1-drop-params","messages":[{"role":"user","content":"Use the get_weather tool to answer: what is the weather in Paris?"}],"tools":[{"type":"function","function":{"name":"get_weather","description":"Get weather for a city","parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}}],"tool_choice":"required","reasoning_effort":"low","max_tokens":300}'

Output, drop_params does not help because nothing knows the param is unsupported:

{"error":{"message":"litellm.BadRequestError: AnthropicException - {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"tool_choice: type \\\"tool\\\" and \\\"any\\\" are not supported for this model.\"},\"request_id\":\"req_011CedBqb1RrGf5hJumoFXwZ\"}. Received Model Group=claude-fable-5-1-drop-params\nAvailable Model Group Fallbacks=None","type":null,"param":null,"code":"400"}}
HTTP 400

Prompt caching cost

Command, run twice:

for n in 1 2; do curl -s http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" -d @cache_body.json -D headers_$n.txt | jq -c '{content: .choices[0].message.content, usage}'; grep -i "^x-litellm-response-cost" headers_$n.txt; done

Output for request 1 (cache write, billed at $0):

{"content": "OK", "prompt_tokens": 3401, "cache_creation_input_tokens": 3385, "cache_read_input_tokens": 0}
x-litellm-response-cost-input: 0.0
x-litellm-response-cost-output: 0.0

Output for request 2 (cache read, billed at $0):

{"content": "OK", "prompt_tokens": 3401, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 3385}
x-litellm-response-cost-input: 0.0
x-litellm-response-cost-output: 0.0

After (`40cd92d`)

Chat completion with reasoning_effort

Command:

curl -s http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"claude-fable-5-1","messages":[{"role":"user","content":"In one sentence, what model are you?"}],"reasoning_effort":"low","max_tokens":200}' -D -

Output (HTTP 200), 19 x $10/MTok + 21 x $50/MTok:

{"model": "claude-fable-5-1", "content": "I'm Claude, an AI model made by Anthropic.", "prompt_tokens": 19, "completion_tokens": 21}
x-litellm-response-cost: 0.0012400000000000002
x-litellm-response-cost-input: 0.00019
x-litellm-response-cost-output: 0.0010500000000000002

Forced tool_choice without drop_params

Command:

curl -s -w "\nHTTP %{http_code}\n" http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"claude-fable-5-1","messages":[{"role":"user","content":"What is the weather in Paris?"}],"tools":[{"type":"function","function":{"name":"get_weather","description":"Get weather for a city","parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}}],"tool_choice":"required","max_tokens":300}'

Output, rejected client-side before any provider call:

{"error":{"message":"litellm.UnsupportedParamsError: claude-fable-5-1 does not support forced tool use (tool_choice='required' or a named tool). Use tool_choice='auto' and tell the model in the prompt when to call the tool, or set `litellm.drop_params = True` to downgrade to 'auto' automatically.. Received Model Group=claude-fable-5-1\nAvailable Model Group Fallbacks=None","type":"None","param":null,"code":"400"}}
HTTP 400

Forced tool_choice with drop_params

Command:

curl -s -w "\nHTTP %{http_code}\n" http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"claude-fable-5-1-drop-params","messages":[{"role":"user","content":"Use the get_weather tool to answer: what is the weather in Paris?"}],"tools":[{"type":"function","function":{"name":"get_weather","description":"Get weather for a city","parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}}],"tool_choice":"required","reasoning_effort":"low","max_tokens":300}'

Output, tool_choice downgraded to auto and the model still calls the tool:

{"finish_reason": "tool_calls", "tool_calls": [{"index": 0, "caller": {"type": "direct"}, "function": {"arguments": "{\"city\": \"Paris\"}", "name": "get_weather"}, "id": "toolu_019Ysxs7ab6e1axvBYUEkqoa", "type": "function"}], "prompt_tokens": 386, "completion_tokens": 50}
HTTP 200

Prompt caching cost

Command, run twice:

for n in 1 2; do curl -s http://localhost:4000/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" -d @cache_body.json -D headers_$n.txt | jq -c '{content: .choices[0].message.content, usage}'; grep -i "^x-litellm-response-cost" headers_$n.txt; done

Output for request 1, cache write: 3395 x $12.50/MTok = 0.0424375, plus 16 uncached input and 4 output tokens:

{"content": "OK", "prompt_tokens": 3411, "cache_creation_input_tokens": 3395, "cache_read_input_tokens": 0}
x-litellm-response-cost: 0.0427975
x-litellm-response-cost-input: 0.00016000000000000042
x-litellm-response-cost-output: 0.0002

Output for request 2, cache read: 3395 x $0.25/MTok = 0.00084875, plus the same 16 input and 4 output tokens:

{"content": "OK", "prompt_tokens": 3411, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 3395}
x-litellm-response-cost: 0.0012087500000000002
x-litellm-response-cost-input: 0.0001600000000000001
x-litellm-response-cost-output: 0.0002

Type

🆕 New Feature

Caveats (if any)

Medium

The forced tool_choice guard covers /v1/chat/completions only
- Native /v1/messages requests with tool_choice any/tool still get the provider's own 400

Low

Anthropic direct only: Bedrock, Vertex AI and Azure AI entries come in a follow-up
supports_forced_tool_use is a new cost map flag, only Fable 5.1 sets it so far
Tests that needed an unmapped Fable id now use claude-fable-5-2 as the placeholder

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

Registers claude-fable-5-1 in the cost maps so the model prices, reports its 1M/128K limits, and advertises adaptive always-on thinking instead of falling through the generalization rules at zero cost. Pricing matches Fable 5 at $10/$50 per MTok except cache reads, which Anthropic prices at 0.025x base input ($0.25/MTok) on this model instead of 0.1x. Fable 5.1 rejects forced tool use (tool_choice type any/tool) with a 400 because thinking is always on and a forced call would skip it. A new supports_forced_tool_use flag on the entry drives the Anthropic chat transformation to downgrade such a tool_choice to auto when drop_params is set, and otherwise raise a clean client-side UnsupportedParamsError that explains the workaround. Anthropic API only for now. Adds the model to the setup wizard and the reasoning effort e2e grid, and moves the tests that used claude-fable-5-1 as an unmapped placeholder id onto claude-fable-5-2.

codspeed-hq · 2026-09-01T18:31:41Z

Merging this PR will not alter performance

✅ 31 untouched benchmarks

_{Comparing litellm_add_claude_fable_5_1 (40cd92d) with litellm_internal_staging (82dd36c)}

greptile-apps · 2026-09-01T18:32:42Z

Greptile Summary

Adds Claude Fable 5.1 support for Anthropic, including pricing, capabilities, reasoning-effort coverage, and model discovery.

Registers direct Anthropic pricing, caching rates, context limits, and capability metadata.
Rejects unsupported forced tool choices clearly or downgrades them to auto when parameter dropping is enabled.
Adds setup-wizard registration, schema support, and focused configuration and transformation tests.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/common_utils.py	Adds a model-capability-driven guard that rejects or downgrades unsupported forced tool choices.
litellm/llms/anthropic/chat/transformation.py	Applies the forced-tool-choice guard after normalizing OpenAI tool-choice inputs.
model_prices_and_context_window.json	Registers Claude Fable 5.1 pricing, limits, and Anthropic capability metadata.
litellm/model_prices_and_context_window_backup.json	Keeps the bundled backup cost map synchronized with the primary model registry.
model_prices_and_context_window.schema.json	Adds schema validation for the forced-tool-use capability flag.
litellm/setup_wizard.py	Adds Claude Fable 5.1 to the established hand-maintained Anthropic model list.
tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py	Covers forced-tool rejection, downgrade behavior, parallel-tool preservation, and unaffected models.
tests/test_litellm/test_claude_fable_5_1_config.py	Validates Fable 5.1 pricing, caching, capabilities, bundled-map synchronization, and provider resolution.

_{Reviews (2): Last reviewed commit: "chore(model-prices): register supports_f..." | Re-trigger Greptile}

codecov · 2026-09-01T18:40:06Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

…p schema

mateo-berri · 2026-09-01T19:04:58Z

@greptileai

mateo-berri · 2026-09-01T21:05:32Z

Superceded by #39148

greptile-apps Bot reviewed Sep 1, 2026

View reviewed changes

Comment thread litellm/setup_wizard.py

chore(model-prices): register supports_forced_tool_use in the cost ma…

40cd92d

…p schema

This was referenced Sep 1, 2026

feat(claude): add Claude Fable 5.1 model pingdotgg/t3code#9078

Merged

feat(claude): add Claude Fable 5.1 pingdotgg/t3code#9077

Closed

mateo-berri closed this Sep 1, 2026

mateo-berri deleted the litellm_add_claude_fable_5_1 branch September 1, 2026 21:05

Navigation Menu

Uh oh!

Uh oh!

feat(anthropic): add Claude Fable 5.1 - #39151

feat(anthropic): add Claude Fable 5.1#39151
mateo-berri wants to merge 2 commits into
litellm_internal_stagingfrom
litellm_add_claude_fable_5_1

mateo-berri commented Sep 1, 2026 •

edited

Loading

Uh oh!

codspeed-hq Bot commented Sep 1, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot commented Sep 1, 2026 •

edited

Loading

Important Files Changed

Uh oh!

Uh oh!

codecov Bot commented Sep 1, 2026

Uh oh!

mateo-berri commented Sep 1, 2026

Uh oh!

mateo-berri commented Sep 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Uh oh!

Conversation

mateo-berri commented Sep 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

TLDR

User Flow

Relevant issues

Linear ticket

Pre-Submission checklist

Delays in PR merge?

Screenshots / Proof of Fix

Before (f2a4172)

Chat completion with reasoning_effort

Forced tool_choice without drop_params

Forced tool_choice with drop_params

Prompt caching cost

After (40cd92d)

Chat completion with reasoning_effort

Forced tool_choice without drop_params

Forced tool_choice with drop_params

Prompt caching cost

Type

Caveats (if any)

Medium

Low

Final Attestation

Uh oh!

codspeed-hq Bot commented Sep 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merging this PR will not alter performance

Uh oh!

greptile-apps Bot commented Sep 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 5/5

Important Files Changed

Uh oh!

Uh oh!

codecov Bot commented Sep 1, 2026

Codecov Report

Uh oh!

mateo-berri commented Sep 1, 2026

Uh oh!

mateo-berri commented Sep 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

mateo-berri commented Sep 1, 2026 •

edited

Loading

Before (`f2a4172`)

After (`40cd92d`)

codspeed-hq Bot commented Sep 1, 2026 •

edited

Loading

greptile-apps Bot commented Sep 1, 2026 •

edited

Loading