What’s new

Remove retired Moonshot and HuggingFace text registrations (#542)

Groq: Remove the shut-down llama-4-scout registration (#541)

Groq: Parse grounding from streamed executed_tools (#540)

Prepare the 1.0.0 release gate (#539)

OpenAI: Remove the shut-down videos provider (#538)

Meta: Accept META_API_KEY alongside MODEL_API_KEY (#537)

Deps dev: Bump pre-commit from 4.6.2 to 4.7.0 in the uv group (#536)

Align SDK templates and response handling (#534)

Replace the README logo with the Celeste sculpture mark (#532)

Anthropic: Drop the temperature step on claude-opus-4-6 (#530)

Black Forest Labs: Add flux-3-image-klein to the images catalog (#531)

ElevenLabs: Add scribe_v2_medical to the audio catalog (#486)

Anthropic: Return empty max_tokens responses instead of raising (#523)

Google: Add audio/mulaw and audio/alaw to Gemini 3.8 TTS output formats (#490)

ElevenLabs: Remove Thai from the v2.5 TTS languages (#522)

BytePlus: Label image outputs with their declared format instead of PNG (#520)

Anthropic: Lower the claude-opus-4-6 max_tokens minimum to 0 (#487)

ElevenLabs: Mark eleven_v3 as streaming (#491)

ElevenLabs: Add Bulgarian and Croatian to the v2.5 TTS languages (#489)

Black Forest Labs: Add flux-3-image editing via per-model input image fields (#519)

Google: Skip kind in the Interactions WebSearch unsupported-field warning (#518)

Google: Stop sending google_search.exclude_domains on Interactions (#517)

xAI: Add multi-image editing via reference images (#516)

Anthropic: Send the documented Claude on Google Cloud request body and basic web search (#515)

Black Forest Labs: Add the grounding image parameter for flux-3-image (#513)

Black Forest Labs: Stop polling on the No Card Generated status (#512)

Black Forest Labs: Send native aspect_ratio for flux-3-image (#511)

Test(audio): fix stream speak chunk content assertion (#510)

ElevenLabs: Replace the retired Rachel default voice with its documented replacement (#509)

BytePlus: Add Seedream image editing and reference images (#508)

xAI: Map quality to the images resolution field (#507)

Black Forest Labs: Add flux-3-image to the images catalog (#506)

Google: Add resolution support to gemini-omni-1.1-flash (#505)

ElevenLabs: Replace the expiring default voices with their documented replacements (#504)

Google: Drop Preview from the Lyria 3.5 display name (#503)

Meta: Add Meta provider with Muse Wonder and Muse Spark 1.3 (#502)

Text: Move reasoning effort strings from thinking_budget to thinking_level (#501)

Openresponses: Join every output message's text (#500)

Google: Add Gemini 4 Argon text model (#499)

Deps: Commit uv.lock and install it with --locked in CI (#498)

Types: Type output_schema as OutputSchema and drop the protocols mypy override (#497)

Google: End Interactions streams at interaction.completed (#496)

Anthropic: Add Claude Haiku 5.5 text model (#495)

Google: Add gemini-nano-banana-2.1 to the images catalog (#493)

Deps: Pin mypy below 2.4 (#494)

ElevenLabs: Mark eleven_multilingual_v2 as streaming (#485)

ElevenLabs: Add eleven_v4 to the audio catalog (#484)

OpenAI: Add gpt-4o-transcribe-diarize to the audio catalog (#483)

Add stream_post_bytes for binary stream framings (#480)

OpenAI: Add GPT-6.1 Sol text model (#479)

Live: Add the live modality with Gemini 3.8 Live and GPT-Live-1 (#477)

Anthropic: Add Claude Sonnet 5.5 text model (#468)

Anthropic: Bound claude-opus-4-6 thinking_budget to 1024 and its 128K max output (#469)

Google: Add 512 image size to gemini-3.1-flash-lite-image (#472)

Google: Remove audio/mp3 from Gemini 3.1 and 3.8 TTS output formats (#470)

Google: Add 1:4, 1:8, 4:1 and 8:1 aspect ratios to gemini-3-pro-image (#471)

xAI: Add xhigh reasoning effort to grok-4.5 and grok-4.3 (#475)

Anthropic: Lower the max_tokens minimum to 0 on twelve Claude text models (#476)

Google: Add minimal thinking level to Gemini 3.5 Flash-Lite and 3.6 Flash (#473)

OpenAI: Enable streaming for gpt-image-1.5 (#474)

OpenAI: Add gpt-image-2.5-sunburst to the images catalog (#455)

Google: Add gemini-3.8-flash-tts to the audio catalog (#452)

OpenAI: Add gpt-image-2.5-flare to the images catalog (#456)

xAI: Add grok-4.7 to the text catalog (#458)

OpenAI: Replay the full output turn when it holds more than messages and calls (#465)

Style(mistral): add the one-line docstring to _init_request (#467)

Google: Resend Interactions built-in tool steps and send dict results as the Vertex response (#464)

Anthropic: Carry toolset_name through tool calls and results (#459)

Mistral: Replay the ThinkChunk on reasoning turns (#463)

Moonshot: Send the tool name on tool result messages (#462)

DeepSeek: Send reasoning_content back on every assistant turn (#461)

Google: Accept Vertex function calls without an id (#460)

OpenAI: Add structured outputs to gpt-4o-mini (#457)

Google: Add medium thinking level to Gemini 3.5 Flash-Lite and 3.6 Flash (#453)

Google: Add gemini-3.8-flash-lite-tts to the audio catalog (#451)

Google: Add 1:4, 1:8, 4:1 and 8:1 aspect ratios to gemini-3.1-flash-lite-image (#450)

BytePlus: Add dola-seedream-5-0-flash-260915 to the images catalog (#449)

Anthropic: Raise thinking_budget ceiling to the 64K max output on Haiku 4.5 and Opus 4.5 (#447)

BytePlus: Enable streaming for seedream-4-5-251128 (#448)

Anthropic: Add Claude Opus 5.5 text model (#446)

OpenAI: Register GPT-6 Sol and Luna text models (#438)

Black Forest Labs: Correct image family contracts and reference limits (#436)

Google: Remove unused Imagen backend (#435)

Anthropic: Align Messages request and response contracts (#428)

Google: Remove discontinued Imagen 4 registrations (#418)

OpenAI: Remove three shut-down direct API registrations (#432)

Black Forest Labs: Honor settled image costs and terminal polling states (#426)

OpenAI: Correct reasoning efforts and schema support (#394)

xAI: Preserve every image output and its metadata (#374)

Anthropic: Retain compaction state for continuation (#430)

Openresponses: Retain hosted tool billing markers (#431)

OpenAI: Add GPT Transcribe with plural language hints (#413)

BytePlus: Accept unaligned Seedream Pro dimensions (#408)

Google: Complete Gemini TTS language metadata (#414)

Google: Rebase Interactions citations across text parts (#420)

OpenAI: Preserve applied speech format and support WAV (#412)

Google: Reconcile Lyria family and final song responses (#405)

OpenAI: Register GPT-6 Astra text model (#433)

Black Forest Labs: Correct FLUX.2 multi-reference image limits (#399)

Google: Drop reference images from veo-3.1-lite-generate-preview (#402)

Google: Add lyria-3.5 to the audio catalog (#400)

Close SDK transports and preserve provider response metadata (#401)a

Google: Add the Lyria 3.5 music models to the audio catalog (#397)

Google: Add medium thinking level to Gemini 2.5 Pro and 2.5 Flash-Lite (#396)

Anthropic: Floor thinking_budget at 1024 on extended-only models (#395)

Anthropic: Allow forced tool use on claude-fable-5 (#393)

Google: Register Gemini 3.8 Flash text model (#391)

OpenAI: Add the gpt-4o-mini-transcribe-2025-12-15 snapshot to the audio catalog (#388)

BytePlus: Add seedream-4-5-251128 to the images catalog (#390)

Google: Add referenceImages support to Veo 3.1 Fast (#389)

Anthropic: Raise Opus 4.6 and Sonnet 4.6 max output to the documented 128K (#387)

Google: Add the documented TTS languages to the audio catalog (#382)

BytePlus: Sync Seedream 5.0 pro size tiers with API reference (#386)

OpenAI: Narrow tts-1 and tts-1-hd voices to the documented set (#385)

xAI: Add grok-imagine-image-2.0 and 21:9, 5:2 aspect ratios (#381)

Google: Register gemini-omni-1.1-flash in the videos catalog (#383)

Anthropic: Add Claude Fable 5.1 and Mythos 5.1 text models (#384)

Google: Preserve all Vertex text response parts (#373)

OpenAI: Preserve multiple image outputs (#356)

Anthropic: Add Mythos 5 and Opus 5 models (#360)

OpenAI: Support transparent GPT Image 2 backgrounds (#357)

OpenAI: Register chat-latest text model (#355)

Remove OpenSpec remnants (#354)

Groq: Include usage in streamed chat completions (#351)

Groq: Register Qwen 3.8 27B text model (#350)

xAI: Register Grok 4.6 text model (#349)

Google: Register Gemini 3.7 Flash text model (#347)

Google: Route mixed tools through Interactions (#342)

Segmentation: Add fal SAM 3 image-RLE modality (#341)

Images: Wire Operation.UPSCALE with Topaz Labs Gigapixel (#340)

Audio: Add OpenAI, ElevenLabs, and Mistral transcription (#339)

Audio: Add transcribe with Groq Whisper STT (#337)

Google: Fall back for mixed interaction tools (#336)

Google: Validate mixed interaction tools (#334)

Google: Constrain 3.5 Flash Lite / 3.6 Flash thinking to low|high (#332)

Models: Model-watch catalog corrections (#331)

Google: Register Gemini 3.5 Flash Lite and 3.6 Flash (#330)

Google: Lyria music rejects response_format.mime_type — drop output_format (#329)

Google: Map audio/mp4 to the Interactions m4a literal (#328)

Google: Send Interactions tool results as typed content blocks (#327)

Videos: Gemini Omni video generation + editing via Interactions (#326)

Google: Seed parameter + UrlContext tool (#325)

Audio: GENERATE operation + Google Lyria music models (#324)

Google: Restore generateContent parity on the Interactions path (#323)

Google: Migrate to the Interactions API (#321)

Templates: _transform_output takes Unpack[{Modality}Parameters] in the dispatcher template (#322)

API-addition playbook, naming rules, templates (#318)

Test(core): dispatcher delegation guard (#317)

Unify composition-layer conventions (#316)

Commit agent guide and fix stale docs (#315)

Provider bugs and catalog corrections from model-watch findings (#314)

Models: Sync provider catalogs and capabilities (#313)

Models: Refresh catalogs and provider capabilities (#312)

Mistral: Align reasoning constraints and catalog with the live API (#311)

Clean up test suite and remove legacy surfaces (#310)

Google: Replay native tool context across calls (#309)

Add voice descriptions (#308)

Exclude audio bytes from metadata (#307)

Fix Gemini grounding source domains (#306)

Models: Register GPT-5.6 models (#305)

xAI: Add grok-4.5 text model (#304)

Add BytePlus Seedream 5 Pro model (#302)

Streaming: Assemble raw_response so streamed usage is complete (#300)

Anthropic: Code-execution replay, max_tokens ceiling; tool kind discriminator (#299)

Fix anthropic native replay (#298)

Anthropic: Stream thinking summaries, fix haiku thinking budget, dynamic-filtering web search (#297)

Anthropic: Echo thinking-bearing assistant turns verbatim in tool loops (#296)

Anthropic: Move thinking effort to output_config (#295)

Google: Add gemini-3.1-flash-lite-image and fix Gemini image catalog errors (#294)

Anthropic: Add Claude Sonnet 5, Opus 4.8, and Fable 5 to catalog (#293)

Google: Add minimal thinking level to gemini-3.5-flash (#292)

Update reasoning effort metadata (#291)

Streaming: Surface native web-search activity per chunk (#290)

Retry transient HTTP failures + require all properties in strict schemas (#288)

Text: Add web search grounding (#287)

Telemetry: Collapse 6 OpenTelemetry no-op stubs into one proxy (#286)

Google: Set includeServerSideToolInvocations for mixed tools (#285)

Text: Add OpenRouter provider via OpenResponses (#284)

Support multimodal text messages (#283)

Models: Add high-priority provider models (#282)

Google: Promote Gemini 3.1 Flash Lite to GA model id (#281)

Models: Remove 17 retired provider models from catalogs (#280)

Google: Add gemini-3.5-flash text model (#279)

Validate returned tool call arguments (#278)

Telemetry: Drop use_span from _TracedStream.__anext__; eager-prime SSE under captured ctx (#277)

Telemetry: Emit gen_ai.response.model from Output.metadata["model"] (#275)

Telemetry: Widen usage attrs, add metrics histograms, opt-in content events (#273)

Telemetry: Emit gen_ai.usage attrs on streaming spans (#272)

Telemetry: GenAI OpenTelemetry spans on _predict and _stream (#270)

Tools: ToolOutput / ToolError types + fix Pydantic content lowering (#267)

DeepSeek: Register V4-Flash and V4-Pro models (#265)

Models: Register 7 new provider models (#264)

Google: Send tool schemas via parametersJsonSchema (#262) (#263)

Parameters: Describe Parameters TypedDicts and Constraint base (#261)

Revert: AuthenticationContext ambient primitive (#259) (#260)

Add AuthenticationContext ambient primitive (#259)

Drop call-arg exclusion from internal modality/provider mypy overrides (#258)

Add total=False to modality Parameters TypedDicts (#256)

Add reasoning output support for text generation (#208) (#254)

Add tool_choice parameter for forced tool calling (#253)

Consolidate trivial parameter mapper wrappers (#252)

Correct URL handling for document and media input (#251)

Add document input support for text modality (#207) (#249)

Centralize operation methods on base modality clients (#248)

Add multimodal embeddings with gemini-embedding-2-preview (#246)

Clean up re-exports in celeste/__init__.py (#233) (#245)

Add unified CHANGELOG.md (#244)

Unified tools= parameter for tool calling (#199)

Add new model definitions for Google, OpenAI, and xAI (#213)

Warn when unsupported parameters are silently ignored (#212)

Correct model ID in README Quick Start (#211)

Make _parse_content a pure parser, add Chunk type param (#203)

Enrich streaming HTTP errors with provider-specific messages (#196)

Add extra_headers support to all client methods (#193)

Detect and raise SSE stream error events instead of silently discarding them (#144) (#192)

Empty/whitespace credentials should raise MissingCredentialsError (#190)

Make ParameterMapper generic on Content type (#189)

Move _transform_output from provider _parse_content to base _predict pipeline (#187)

Catch UnicodeDecodeError in _handle_error_response (#170) (#171)

Deduplicate streaming methods across all modalities (#168)

Remove _build_headers() side effect in Anthropic client (#159) (#163)

Add _json_headers() helper to APIMixin (#162)

Add FieldMapper base class and migrate 41 simple mappers (#161)

Extract _content_fields ClassVar to eliminate duplicate _build_metadata() overrides (#160)

Extract Chat Completions protocol from 4 providers (#154)

Migrate OpenResponses to protocol layer (#153)

Use strict=False in json.loads for structured output parsing (#152)

Centralize usage/finish_reason type conversion in base classes (#151)

Add auth passthrough to all namespace methods (#146)

Add base64 serialization config for Output and Chunk models (#145)

Fix Opus 4.6 model config for claude-code-action v1 (#141)

Add web search support for Anthropic models (#140)

Let Artifact validator handle base64 decoding across all providers (#139)

Add Vertex AI support for all providers (#135)

Add Claude Code GitHub Workflow (#136)

Deps: Make google-auth optional under celeste-ai[gcp] (#133)

Deps: Remove requests and ipykernel from runtime deps (#132)

Rewrite CONTRIBUTING.md with accurate workflow (#131)