The model catalog.

The frontier ceiling — deepest reasoning and vision for the hardest problems.

Prior Opus flagship — frontier reasoning, pinned in place.

Frontier generalist — broad, sharp, and steady on long tool chains.

The prior GPT-5 flagship — most of 5.5's strength for less.

Frontier scale — a 1M-token window and strong multimodal reasoning.

Moonshot's flagship — deep reasoning, vision, and a 1M-token window.

DeepSeek's flagship — frontier-grade reasoning at a startling price.

The current Qwen flagship — frontier scale with a 1M-token window.

Thinking Machines' frontier debut — text, images, and audio in one model.

Default. Strong general-purpose agent model with fast responses.

The premium ceiling — top reasoning and vision for the hardest work.

The newest Sonnet — near-frontier quality for everyday premium work.

The fast Claude — premium-family quality at a fraction of the latency.

Deep reasoning over a very large window, at the low end of premium.

Highest-quality GLM. Flagship tuned for quality over raw speed — higher, variable latency.

Fast, reliable coding flagship for everyday heavy work.

Heavy coding flagship — large MoE built for complex code.

Coding-specialist flagship. Quality-first; higher, variable latency.

The prior Qwen flagship — a heavyweight that codes exceptionally well.

Big reasoning model at a cheap-tier price — standout value.

Strong reasoning. Best on open-ended analysis, not strict tool loops.

Extended reasoning; budget output tokens for its hidden chain-of-thought.

Cheap reasoning with a large context window.

Baidu's reasoning specialist — deep deliberation at a low price.

Tencent's reasoning model — strong long-form thinking, very cheap.

StepFun's multimodal reasoner — thinks, sees, and calls tools.

Qwen's biggest open-weight reasoner — flagship thinking, mid-tier price.

The top Nemotron — 550B of deliberate reasoning at a mid-tier price.

Strong all-round workhorse with a very large context.

Hybrid MoE, strong on multi-agent, 256K context.

Efficient workhorse — big context at a low price.

Google's open workhorse — reasoning and a 256K window near the price floor.

Large general-purpose model, fast and capable.

Baidu's flagship generalist — broad knowledge with vision.

Tencent's fast generalist — quick answers at a rock-bottom price.

Google's fast frontier model — 1M context, multimodal, cheap.

MiniMax's frontier agent model — 1M context and vision at a workhorse price.

Kimi's vision generalist — thinks when asked, sees what you show it.

The GLM workhorse — flagship instincts at an everyday price.

Qwen's balanced mid-tier — a 1M-token window at an everyday price.

Affordable eyes — a 235B vision model at workhorse money.

Tencent's newest generalist — reasoning and tools near the price floor.

Meta's dependable open workhorse — a known quantity everywhere.

Cheap coding offload for routine changes.

Balanced coding model with a large context.

Fast coding-general model.

Coding-specialist tuned for software tasks.

MiniMax's coding workhorse — agentic edits at a budget rate.

Kuaishou's coding specialist — direct edits, no reasoning overhead.

256K context for the price of a small model.

Meituan's flagship MoE — a 1M-token window with an unusually deep cached-input discount.

The V4 line's fast sibling — a 1M-token window at cheap-tier prices.

A million tokens of context at one flat, tiny price.

Xiaomi's efficiency play — 1M context and vision for pocket change.

The catalog's cheapest chat tokens — quick, direct answers at volume.

Flagship reasoning with images and a near-million-token window.

Smallest and fastest tier.

Near the price floor — GLM-family quality in the budget tier.

Cheapest tool + reasoning capable model.

StepFun's quick multimodal — sees, thinks, and answers fast for very little.

BAAI's versatile embedding — multilingual, multi-granularity.

The long-standing default English embedding.

384 dimensions — the smallest index and the fastest search.

The most accurate English embedding in the catalog.

Compact Gemma-family embedding with a 2K input window.

Multilingual retrieval with an 8K input window.

Japanese-specialist embedding — the widest vector in the catalog.