Celeste 1.0 brings live voice sessions, transcription and typed decisions to one Python API

Celeste 1.0 is the first stable release of the Python SDK that puts OpenAI, Anthropic, Google and more than a dozen other providers behind one interface. The headline is live sessions, a realtime voice conversation through the same API you already use for text. Transcription, upscaling and typed decisions come with it, and so do some breaking changes.

Slide text

Live voice sessions,
in the same Python API

Celeste 1.0, the first stable release, also adds transcription, upscaling, segmentation and typed decisions.

Live is a new modality. You open a session with celeste.live.connect, send it audio, and read chunks that carry the model's audio, a transcript and any tool calls. Gemini Live also takes text and images. It runs on Gemini 3.8 Live, also on Vertex, and on OpenAI's GPT Live 1.

Slide text

celeste-ai 1.0.0

Hold a realtime voice session with celeste.live.connect

async with celeste.live.connect( model="gemini-3.8-live", ) as livestream: await livestream.send( "Say hello in one short sentence." ) async for chunk in livestream: if chunk.finish_reason: await livestream.end() output = livestream.output # audio, usage, transcript

Google

gemini-3.8-live
gemini-3.8-live-extended-thinking

Also on Vertex through GoogleADC

OpenAI

gpt-live-1

Beyond live sessions, 1.0 adds new operations. Transcribe speech with Whisper on Groq, OpenAI's models, ElevenLabs Scribe or Mistral's Voxtral. Generate music with Google's Lyria, upscale images through Topaz Labs, segment them with SAM 3, and edit video with Gemini Omni. And decide answers typed questions about text or an image, with probabilities, using TypeSafe's Jev models.

Slide text

celeste-ai 1.0.0

Transcribe, upscale, segment and decide, one call each

Transcribe

celeste.audio.transcribe

Groq Whisper, OpenAI, ElevenLabs Scribe, Mistral Voxtral

Generate music

celeste.audio.generate

Google Lyria

Upscale

celeste.images.upscale

Topaz Labs Gigapixel

Segment

celeste.images.segment

SAM 3 on fal.ai

Edit video

celeste.videos.edit

Gemini Omni

Decide

celeste.text.decide

Typed yes/no, choice, score and box answers from TypeSafe Jev

Answers now carry their evidence. Pass the WebSearch tool, and the response's grounding field lists the queries, sources and citations, on OpenAI, Anthropic, Google, xAI and Groq. Reasoning models return their thinking in a reasoning field, and one message can now mix text, images, audio, video and documents.

Slide text

celeste-ai 1.0.0

Answers now show their sources and their reasoning

from celeste import WebSearch response = await celeste.text.generate( "What changed in the latest uv release?", model="...", tools=[WebSearch()], ) response.grounding.sources response.grounding.citations response.reasoning

grounding

Queries, sources and citations, streamed or not

reasoning

The model's thinking, with a signature to send back next turn

list[MessagePart]

Text, image, audio, video and document parts in one message

tool_choice

AUTO, REQUIRED, NONE or one named function

Two changes matter in production. Ordinary requests now retry by themselves, twice, with exponential backoff, on timeouts, rate limits and server errors. Streams are not retried. And with the new otel extra and a tracer provider, every call emits OpenTelemetry spans with the model, the provider and token usage.

Slide text

celeste-ai 1.0.0

Requests retry on their own and show up in your traces

Retried

408 429 500 502 503 504

Plus timeouts and network errors
2 retries, exponential backoff
Streams are not retried

Traced

gen_ai.request.model
gen_ai.provider.name
gen_ai.usage.*
gen_ai.client.token.usage

OpenTelemetry, with the otel extra and a tracer provider

Five providers join. Meta's Muse models, OpenRouter, fal, Topaz Labs and TypeSafe. The catalog adds Claude Opus, Sonnet and Haiku 5.5, GPT 6.1 Sol, Gemini 4 Argon and Grok 4.7, and models the providers retired are gone from it.

Slide text

celeste-ai 1.0.0

Five providers join, and the model catalog is current

Meta

Muse models

OpenRouter

Any model slug

fal.ai

SAM 3 segmentation

Topaz Labs

Image upscaling

TypeSafe

Jev decisions

New in the catalog

claude-opus-5-5 claude-sonnet-5-5 claude-haiku-5-5 gpt-6.1-sol gemini-4-argon grok-4.7 flux-3-image eleven_v4

Now the breaking changes. The legacy capability argument is gone, so create clients with a modality and an operation. The web search flag becomes a WebSearch tool. On OpenAI and xAI, named efforts move from thinking budget to thinking level, and the old form only logs a warning, so search for it. APIKey is now AuthHeader, and Topaz reads a renamed environment variable.

Slide text

celeste-ai 1.0.0

Rename these before you upgrade

capability=Capability.TEXT_GENERATION

→

modality=Modality.TEXT,
operation=Operation.GENERATE

create_client, list_models, register_models

web_search=True

→

tools=[WebSearch()]

Same for x_search and code_execution

thinking_budget="high"

→

thinking_level="high"

OpenAI and xAI. The old form is ignored with a warning.

APIKey
TOPAZLABS_API_KEY

→

AuthHeader
TOPAZ_API_KEY

A few more changes to check. On Anthropic, max tokens now defaults to the model's ceiling instead of 1,024, so cap it yourself if you relied on short replies. Every earlier ElevenLabs voice fails validation, and the default is now Janet. OpenAI's video provider is gone, along with Sora 2. And specific exceptions now import from celeste dot exceptions.

Slide text

celeste-ai 1.0.0

More to check after the upgrade

1,024
→ the ceiling

Anthropic default max_tokens

Up to 128,000 tokens per model. Pass max_tokens= to cap replies.

Rachel
→ Janet

ElevenLabs default voice

All 52 earlier ElevenLabs voices, Adam and George too, now fail validation.

Sora 2
removed

OpenAI videos provider

OpenAI shut down its Videos API. Google, BytePlus and xAI still generate video.

Specific exceptions and internal types left the top-level celeste import. Import them from celeste.exceptions, celeste.core and the other modules.

To upgrade, pin 1.0.0 with uv add. If you maintain a custom provider, set aside time first, because the header and stream hooks on APIMixin became async. The full list of breaking changes is in section 10 of the changelog. That's Celeste 1.0, live voice sessions in the same Python API.

Slide text

celeste-ai 1.0.0

Upgrade with one line

uv add "celeste-ai==1.0.0"

Maintain a custom provider? Plan time first: APIMixin header and stream hooks are async now.

Every breaking change: CHANGELOG.md, section 10
github.com/withceleste/celeste-python