| OpenAI |
GPT-5.6 Luna |
2026-07-09 |
1050K |
The fastest and cheapest GPT-5.6 tier (Luna), for high-throughput chat and lightweight tasks. |
| xAI |
Grok-4.5 |
2026-07-08 |
500K |
A mid-to-late 4.x flagship with strong coding and text Arena results. Context and reasoning efficiency kept improving. |
| Anthropic |
Claude Sonnet 5 |
2026-06-30 |
1M |
The Claude 5 balanced tier, balancing speed and intelligence. Adaptive thinking is on by default, for most agents and knowledge work. |
| Zhipu GLM |
GLM-5.2 |
2026-06-16 |
1M |
Context expanded to 1 million. Adds IndexShare, MTP, and reasoning-effort control. MIT-licensed. |
| Kimi |
Kimi K2.7 Code |
2026-06-12 |
256K |
A thinking-mode coding model for software engineering. Thinking only, optimized for repo-scale long tasks. |
| Anthropic |
Claude Fable 5 |
2026-06-09 |
1M |
The first Mythos-class high-end model opened to regular users. Aimed at ultra-long-horizon agents and the strongest reasoning, with 1M context. |
| Qwen |
Qwen3.7-Plus |
2026-06-01 |
1M |
The 3.7 balanced hosted tier. 1 million context, covering office production and long-horizon agents. |
| Anthropic |
Claude Opus 4.8 |
2026-05-28 |
1M |
The closing 4.x flagship, continuing to beat 4.7 on coding, agent skills, reasoning, and knowledge work. |
| Qwen |
Qwen3.7-Max |
2026-05-20 |
1M |
The next Max-series flagship. Thinking is on by default, with 1 million context. Later snapshots added visual understanding. |
| Gemini |
Gemini 3.5 Flash |
2026-05-01 |
1048576 |
The 3.5 high-speed tier, for low-latency multimodal apps. Context is about 1049K. |
| DeepSeek |
DeepSeek-V4 Preview |
2026-04-24 |
1M |
A V4 preview with 1 million context by default. Split into Pro and Flash MoE tiers, with open weights. |
| OpenAI |
GPT-5.5 |
2026-04-23 |
1M |
The next default flagship for ChatGPT and Codex. Instant became the full default on May 5, with fewer hallucinations and better personalization. |
| Kimi |
Kimi K2.6 |
2026-04-21 |
256K |
A general multimodal K2 iteration with stronger long-horizon coding and agent workflows. It was the recommended model before K3. |
| Anthropic |
Claude Opus 4.7 |
2026-04-16 |
1M |
A broadly released flagship for complex reasoning and long-horizon coding. Higher visual resolution, with Claude Design launched at the same time. |
| Qwen |
Qwen3.6 |
2026-04-16 |
262144 |
Built on 3.5 with stronger stability and real-world coding. Open-sourced sizes such as 35B-A3B and 27B. |
| Zhipu GLM |
GLM-5.1 |
2026-04-07 |
200K |
Aimed at extra-long tasks, strategy correction, and bug fixing. Officially, agents can run for hours. |
| OpenAI |
GPT-5.4 mini |
2026-03-17 |
1M |
Brings GPT-5.4-class capability down to a mid-small model. Fits high-concurrency agents and tool calling. |
| OpenAI |
GPT-5.4 nano |
2026-03-17 |
1M |
The smallest GPT-5.4 size. Keeps million-token context and fits classification, routing, and light tool calling. |
| OpenAI |
GPT-5.4 |
2026-03-05 |
1M |
The first mainline native computer-use model. API/Codex offer million-token context and tool search, with better token efficiency than GPT-5.2. |
| xAI |
Grok-4.20 |
2026-02-17 |
256K |
A multi-agent preview generation. Default 256K, expandable to 2M. Later Beta 2 strengthened instruction following and reduced hallucinations. |
