25 frontier models. All live today.

Anthropic Claude Opus 5.5 (released 2026-09-22) — long-running agentic coding and knowledge work, native 1M context; adaptive thinking always on, default effort medium.

Anthropic Claude Sonnet 5.5 (released 2026-09-28) — the best combination of speed and intelligence, native 1M context; adaptive thinking on by default, default effort high.

Anthropic Claude Haiku 5.5 (released 2026-10-07) — high-volume, latency-sensitive classification, extraction, routing and subagent work, native 1M context; adaptive thinking on by default, default effort medium; prompts over 100k tokens are billed at 5×.

OpenAI's next-generation flagship (released 2026-09-03) — reasoning effort up to max.

OpenAI's most efficient GPT-6 model (released 2026-09-22) for focused, high-volume tasks — summarization, extraction, focused coding.

OpenAI's GPT-6 workhorse (released 2026-09-22) for complex coding and agentic workflows — reasoning effort none through max.

OpenAI's GPT-6.1 workhorse (released 2026-09-29) for coding and everyday work — reasoning effort low through max (none / minimal unsupported).

OpenAI's GPT-6.1 efficient tier (released 2026-10-26) for visual reasoning, precise computer use and clear, natural writing — reasoning effort none through max, default medium (minimal unsupported).

Google's most intelligent model — sustained frontier performance, excels at agentic and coding tasks.

Gemini speech-to-text on /v1/audio/transcriptions — automatic language detection, files up to 14MB.

Lower-cost Gemini text-to-speech on /v1/audio/speech — 24kHz WAV output, Chinese supported.

Gemini text-to-speech on /v1/audio/speech — 24kHz WAV output, Chinese supported, 30 prebuilt voices.

Always points to the newest Gemini Flash generation.

Always points to the newest Gemini Flash-Lite generation.

DeepSeek V4.1 Flash — the vendor positions it as surpassing V4 Pro on quality, cost and speed. Accepts image input; deep thinking on by default and can be disabled.

DeepSeek V4 Pro — significantly stronger agent capability and rich world knowledge; deep thinking on by default, can be disabled manually.

Zhipu's next-gen flagship model with native 1M-token context and adjustable thinking depth — excels at long-horizon coding and complex engineering optimization.

Zhipu's first native multimodal model in the GLM-5 series — 320B total / 18B active parameters, with image/video/file input and 1M-token context.

Kimi Code K2.8 Preview (kimi-for-coding upgraded in place) — close to K3 with more efficient thinking; reasoning_effort low/high/max (default max), image+video, 1M context.

Kimi's flagship multimodal model (text/image/video input) with a 1M-token context window and always-on reasoning; Kimi Code defaults reasoning_effort to high.

xAI's current flagship for coding, agents, and knowledge work — 500k context, tunable reasoning_effort (default high, including xhigh).

Xiaomi's cost-efficient multimodal MiMo model (309B total / 15B active) with a 1M-token context and tool calling; thinking is on by default and can be turned off.

Xiaomi's flagship MiMo reasoning model with omni-modal input (text/image/video/audio), a 1M-token context and tool calling; thinking is on by default and can be turned off.

Alibaba's small Qwen3-VL vision-understanding model (text/image/video input) with a 256K-token context and tool calling; thinking is off by default and can be turned on.

Alibaba's Qwen3.8 Flash multimodal model (text/image/video input) with a 1M-token context and tool calling; thinking is on by default with adjustable effort and can be turned off.