$ impossibl changelog
gpt-6.1 terra joins the ai api. openai's fast, efficient model for visual reasoning, computer use and natural writing, with a 1,050,000 token context window, text and image input, at $1 per million input tokens and $5 per million output.
kimi k3.1 pro highspeed joins the ai api. moonshot's faster serving of k3.1 pro for coding agents and interactive use, with a 1m context window, text and image input, always-on reasoning with three effort levels, at $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output.
kimi k3.1 pro joins the ai api. moonshot's flagship for long-horizon coding, 3d creation and financial research, with a 1m context window, text and image input, always-on reasoning with three effort levels, at $2 per million input tokens, $0.20 per million cached input tokens and $10 per million output.
minimax m3.1 flash joins the ai api, with always-on reasoning and five effort levels.
deepseek v4.1 flash 1017 joins the ai api. deepseek's october snapshot of its cost-efficient flash model, stronger at coding and agent work, with a 1m context window, text and image input, and low, high and max reasoning effort, at $0.30 per million input tokens and $1.20 per million output.
claude opus 5.5 now supports fast mode on all three generation endpoints. send service_tier: "fast" on chat completions or responses, or speed: "fast" on messages. fast uses anthropic direct and passes through $8 input / $40 output per million tokens, with cache pricing on top. responses report delivered speed; a confirmed standard-speed downgrade bills standard rates. see speed tiers for examples, cache prices and explicit client fallback.
jev 1.14.0 joins the gateway.
muse spark open joins the ai api. meta's open-weight reasoning model for coding and long-running agent work, with a 1m context window, text and image input, and six reasoning effort levels, at $1.25 per million input tokens and $4.25 per million output.
muse wonder joins the ai api. meta's model for complex software engineering and long-running agent tasks, with a 1m context window, text and image input, and six reasoning effort levels, at $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output.
qwen3.6 27b joins the ai api. alibaba's dense open-weight model for agentic coding and repository-level work, with a 229,376 token context window, text and image input, and a thinking on/off switch, at $0.60 per million input tokens and $3.60 per million output.
gemini 4 argon joins the ai api. google's first gemini 4 model, built for careful reasoning on long coding, research and knowledge work, with a 1m context window, text plus image, audio and video input, at $2 per million input tokens and $10 per million output.
Step 3.7 Flash joins the gateway.
Step 5 Preview joins the gateway.
openai's gpt-6 luna decisions joins the gateway on a new endpoint, post /v1/decisions: send an input and typed yes/no, choice or score questions, get back typed answers with probabilities, at $0.10 per million input tokens and nothing for output.
claude haiku 5.5 joins the ai api. anthropic's fastest and cheapest claude, now with a 1m context window and text plus image input, from $0.10 per million input tokens and $0.50 per million output.
Google's Nano Banana 2.1 is live on /v1/images/generations, generating 1K, 2K and 4K square images from text at $1.50 per million input tokens and $30.00 per million image output tokens, with a 65,536 token context window.
gemini robotics er 2 preview and gemini 3.8 flash tts now bill at google's introductory rates through 2026.
Mistral Large 4 joins the gateway.
hunyuan 3 now costs $0.132 in and $0.528 out per million tokens.
hunyuan 4 preview joins the ai api.
mimo v2.6 flash joins the ai api.
mimo v2.6 pro joins the ai api.
glm 5.3 fast joins the ai api. it is a faster serving tier over the same glm-5.3 weights.
gpt-6.1 sol joins the ai api.
claude sonnet 5.5 joins the ai api. it is anthropic's successor to sonnet 5 for everyday coding and writing work.
gemini accepts tool calls carried over from other models without discarding existing thought signatures. mistral on foundry accepts foreign tool call ids and tools without descriptions. mimo v2.5's current route is marked text-only; image requests return a clear 400.
a tool result that carries an image now reaches the model instead of being flattened to text on every wire. /v1/chat/completions accepts image_url parts on tool messages, /v1/messages accepts image blocks inside tool_result, and /v1/responses accepts input_image items in a function_call_output. anthropic, gemini and the openai responses api receive the image natively inside the tool result; chat-completions backends get it as one user message right after the full tool-result group, so parallel tool calls stay intact. inline base64 images are supported; remote image urls in tool results return 400 instead of being fetched by the gateway. explicitly text-only models reject image input with a 400 rather than silently dropping it.
bedrock provider keys now detect every model your aws account can call, route those through your key first, and can be refreshed from the console.
scribe v2 and scribe v2 medical join /v1/audio/transcriptions, billed per second of audio at the scribe rate.
claude opus 5.5 joins the ai api. it is anthropic's flagship opus model for long-running agentic coding, and it serves from our east us azure foundry deployment first, falling back to anthropic direct.
gemini 3.8 flash tts joins /v1/audio/speech with the same 30 steerable voices, billed per token at the flash tts rate.
gpt-6 luna joins the ai api at $0.10 in / $0.50 out per million tokens, with a 1,050,000 token context window and text plus image input.
gpt-6 sol joins the ai api at $2.00 in / $10.00 out per million tokens, with a 1,050,000 token context window and text plus image input.
grok 4.7 joins the ai api. it is spacexai's flagship model for coding and agentic work, with text and image input.
laya and laya-multilingual join the gateway free on /v1/systemone, with an 8k context window.
jev is now pinnable: typesafe-ai/jev-1.13.0 always answers with that version, while typesafe-ai/jev and the new typesafe-ai/jev-latest follow typesafe's rolling alias.
glm 5.3 flashx joins the ai api. it is z.ai's high-speed serving tier over the glm-5.3-flash weights.
typesafe ai's jev joins the gateway on a new endpoint, post /v1/systemone: send a state and typed yes/no, choice or score questions, get back typed answers with probabilities and confidence. mirrors typesafe's native api, so their sdk works by pointing typesafe_base_url at api.impossibl.com.
qwen3.7 and qwen3.8 flash, inkling small, nemotron 3 ultra, kimi k2.7 code, longcat 2, mimo v2.5 pro, gemini image and embedding, qwen embedding and reranking, whisper v3 and grok transcription models join the gateway.
deepseek v4.1 flash joins the gateway.
openai/gpt-image-2.5-flare is live on /v1/images/generations at $5.00 input and $30.00 output per 1M tokens.
openai/gpt-image-2.5-sunburst is live on /v1/images/generations at $5.00 input and $30.00 output per 1M tokens.
hunyuan 3 drops to $0.132 / $0.528 per million tokens, down from $0.20 / $0.80.
hunyuan 4 preview joins the gateway.
gpt-6 astra joins the gateway with a 1,050,000 token context window and text plus image input.
muse spark 1.3 and muse spark 1.3 contributor join the ai api.
workspaces join the console: invite your team, give each member a role — owner, admin, developer, or viewer — and switch between the workspaces you belong to from the top of the console. keys, balance, and usage belong to the workspace.
gemini 3.8 flash joins the ai api with a 1m-token context window. it is promo-priced at $0.75 / $3.75 per million tokens through 2026-12-31.
claude fable 5.1 joins the ai api with a 1m-token context window and adaptive thinking. cache reads bill at $0.25 per million tokens, a quarter of fable 5's rate.
glm 5.3 flash joins the ai api.
qwen3.8 27b joins the ai api.
glm 5.3 joins the ai api.
deepseek v4 pro 0813 joins the ai api.
gemini 3.7 flash joins the gateway.
grok 4.6 joins the gateway.
qwen3.8 2.4t a95b joins the ai api.
muse glimmer 30b joins the ai api.
nemotron 3.5 lightning joins the ai api.
take your request log with you. filter the logs room, press export csv, and the whole filtered window walks itself into a spreadsheet — not just the rows on screen.
a cheap request no longer reads as free. costs under a cent keep their digits down to the credit, so a log of small calls stops printing as a column of $0.00.
qwen3.8 max joins the ai api.
deepseek v4 flash 0731 joins the ai api.
gemini 3.1 flash tts brings 30 steerable voices to /v1/audio/speech, billed per token like every other gemini model.
gemini robotics-er 2 preview is now available.
grok voice tts brings five expressive voices and mp3, wav or pcm output to the new /v1/audio/speech endpoint.
send audio and video to gemini models on chat completions, the responses api, and the gemini-native endpoint, with audio billed at the provider's audio rate.
gpt-transcribe brings high-accuracy speech-to-text for completed files and streamed transcripts to /v1/audio/transcriptions.
the ai api now turns text into vectors on /v1/embeddings, starting with openai/text-embedding-3-small and openai/text-embedding-3-large.
the ai api now generates images, starting with openai/gpt-image-2 on /v1/images/generations.
claude opus 5 joins the ai api with a 1m-token context window and native anthropic fallback.
hunyuan 3 joins the ai api through atlascloud.
qwen 3.7 plus joins the ai api through fireworks.
atlascloud joins the gateway, serving mimo 2.5 and minimax m3.
gemini 3.5 flash-lite is now available.
gemini 3.6 flash is now available.
every request is now logged. browse them in the dashboard, or have agents pull their own from /v1/requests.
mimo 2.5 joins the ai api through atlascloud.
minimax m3 joins the ai api through atlascloud.
usage and balance update live in the dashboard as requests come in.
grok 4.5 joins the gateway.
alibaba's qwen models join the gateway.
qwen's 3.8 max preview joins the gateway.
failed routes no longer slow down every request. in our benchmarks, failover times dropped from 6 seconds to under 1 millisecond.
recharge when your balance runs low.
moonshot joins the gateway.
a better start.
two context tiers: 64k and 256k.
three fast-inference providers join the gateway.
one-liner migration now supported
luna joins the gateway.
sol joins the gateway.
terra joins the gateway.
xai's native web_search and x_search tools.
v4 and v3.2 join the gateway.
bring your own provider keys now supported
fable 5 redeployed.
humans can claim agent-created accounts and merge balances. usage stays at provider price.
sonnet 5 launches.
glm models join the gateway.
api keys, credit top-ups and the model catalog.
anthropic messages support, with image input and tool calling on every surface.
impossibl init
