chore(release): patch v1.89.0-rc.1 with #30064 (Claude Fable 5) for v1.89.0-rc.2

[mateo-berri]

Relevant issues

Adds #30064 (Claude Fable 5 across Anthropic, Bedrock, Vertex AI, and Azure AI) to the patch/v1.89.0-rc.1 line so it can be tagged as v1.89.0-rc.2. Cherry-picked from the squashed commit that landed on litellm_internal_staging (e15b37a18e).

What is included

• Add Claude Fable 5 across Anthropic, Bedrock, Vertex AI, and Azure AI #30064 Claude Fable 5 model entries (Anthropic, Bedrock converse + geo profiles, Vertex AI, Azure AI), supports_adaptive_thinking wiring, BEDROCK_CONVERSE_MODELS registration, setup wizard entry, the sampling-params gating (supports_sampling_params: false for Fable 5 / Opus 4.7 / Opus 4.8 with drop_params support and clean UnsupportedParamsError), the top_k=0 truthiness fix, and all accompanying tests

No version bump: the 1.89 line keeps pyproject.toml at 1.89.0 and the rc number is a git tag, matching how the 1.88 rc patches were cut (#29637).

One conflict resolved in tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py: the upstream context includes a billing-header test block that exists on litellm_internal_staging but not on the rc1 line; #30064's only change inside that region was two blank lines, so the resolution drops the block and keeps the rest byte-identical (1238 insertions here vs 1240 upstream — the delta is exactly those two blank lines).

Verified on this branch: tests/test_litellm/test_claude_fable_5_config.py (18 passed), sampling/top_k tests in the anthropic transformation file (13 passed) and bedrock converse file (8 passed); Black clean on the resolved file.

Pre-Submission checklist

• The cherry-picked PR carries its own tests
• Scope is limited to backporting an already-merged change

Type

🆕 New Feature (backport to rc line)

Changes

See the commit list. No new code beyond the cherry-pick and the conflict resolution described above.

[greptile-apps]

Greptile Summary

This backport cherry-picks #30064 onto the patch/v1.89.0-rc.1 line to produce v1.89.0-rc.2. It adds Claude Fable 5 model entries across Anthropic, Bedrock (base + global/us/eu geo profiles), Vertex AI, and Azure AI, and introduces a shared _apply_sampling_param gate that drops or rejects temperature ≠ 1, top_p, and top_k for Fable 5 / Opus 4.7 / Opus 4.8 models, with a top_k=0 truthiness fix along the way.

• Model map: Fable 5 entries added for all four providers; supports_sampling_params: false backfilled on all existing Opus 4.7/4.8 variants; supports_adaptive_thinking: true set on every Fable 5 variant.
• Sampling-param gating: New _apply_sampling_param helper is wired into AnthropicConfig.map_openai_params, AnthropicConfig.transform_request (for top_k), and both AmazonConverseConfig paths; drop_params is threaded through correctly with backward-compatible defaults.
• Tests: 18 unit tests for Fable 5 config, 9 for the Anthropic transformation path, and 8 for Bedrock Converse — all mocked, no network calls in tests/test_litellm/.

Confidence Score: 4/5

Safe to merge; the sampling-param gating and model map entries are well-tested and backwards-compatible.

The change is a focused, well-tested backport. The only notable concern is the hardcoded model-name fallback strings added to _supports_sampling_params as a safety net — the team's standing policy is to drive these decisions from the model map alone. All actual Fable 5 variants have the flag in the JSON so the fallback is not exercised in practice, but it extends a fragile pattern that the team has explicitly called out.

litellm/llms/anthropic/common_utils.py — the _supports_sampling_params fallback; everything else looks correct.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/common_utils.py	Adds _supports_sampling_params, _apply_sampling_param, _model_map_lookup_candidates, and _get_model_capability; refactors _supports_model_capability. The fallback in _supports_sampling_params uses hardcoded name strings as a safety net when the model map is absent.
litellm/llms/anthropic/chat/transformation.py	Routes temperature/top_p through _apply_sampling_param; gates top_k at the transform_request boundary via is not None check. Logic is clean and well-tested.
litellm/llms/bedrock/chat/converse_transformation.py	Propagates drop_params through the converse pipeline; fixes the top_k=0 truthiness bug (if val_top_k → if val_top_k is not None); routes sampling params through the shared gate. All changes are backwards-compatible (new params default to False).
model_prices_and_context_window.json	Adds claude-fable-5 entries for Anthropic, Bedrock (base + global/us/eu profiles), Vertex AI, and Azure AI with correct pricing, capability flags, and supports_sampling_params: false. Also backfills supports_sampling_params: false on all existing Opus 4.7/4.8 variants.
tests/test_litellm/test_claude_fable_5_config.py	New test file validating model map entries, pricing, adaptive-thinking flags, and sampling-param gating for all Fable 5 variants. Uses mocked model cost map and no real network calls.

_{Reviews (1): Last reviewed commit: "Add Claude Fable 5 across Anthropic, Bed..." | Re-trigger Greptile}

[greptile-apps]

The name-string fallback inside _supports_sampling_params hardcodes "fable", "opus-4-7", and "opus-4-8" as models that do not support sampling params. The team's rule is to drive these decisions exclusively from model_prices_and_context_window.json so that future model updates work without a code change. When a provider-routed id is missing the flag, the fallback is to assume sampling params are supported (return True), which is the safer default — the user would see a provider 400 instead of a silent incorrect behaviour, and the fix is just a JSON entry rather than a code deployment. The same pattern exists in _is_adaptive_thinking_model but that's a pre-existing issue; adding more hardcoded names here extends the problem.

Suggested change

	if flag is not None:
	return flag
	model_lower = model.lower()
	return not any(
	v in model_lower
	for v in (
	"fable",
	"opus-4-7",
	"opus_4_7",
	"opus-4.7",
	"opus_4.7",
	"opus-4-8",
	"opus_4_8",
	"opus-4.8",
	"opus_4.8",
	)
	)
	if flag is not None:
	return flag
	# No entry in the model map → assume sampling params are supported.
	# The caller will see a provider 400 for genuinely unsupported models,
	# which is easier to diagnose than a silent drop. Add
	# ``"supports_sampling_params": false`` to the model map entry to gate
	# sampling params for new models without a code change.
	return True

Rule Used: What: Do not hardcode model-specific flags in the ... (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!