| Anthropic |
Claude Opus 5.5 |
2026-09-22 |
1M |
The first flagship of the Claude 5.5 family. On most work it matches Fable 5.1, and typical runs cost about 40% less than Opus 5. 1M context, with cybersecurity, biology, and anti-distillation safeguards at the Fable 5.1 level. |
| OpenAI |
GPT-6 Sol |
2026-09-22 |
1050K |
The GPT-6 balanced tier (Sol). It carries Astra's alignment and capability into a faster, cheaper model for complex coding and agentic workflows, with 1.05M context. |
| OpenAI |
GPT-6 Luna |
2026-09-22 |
1050K |
The most efficient GPT-6 tier (Luna), for high-throughput, repeatable lightweight tasks. 1.05M context. Free and Go users can try it in the desktop app. |
| xAI |
Grok-4.7 |
2026-09-21 |
500K |
xAI's latest flagship as of September. A larger base model with a longer reinforcement-learning run. Stronger than Grok 4.6 on coding, long-horizon agents, and knowledge work. 500K context, at the same price as 4.6. |
| DeepSeek |
DeepSeek-V4.1-Flash |
2026-09-10 |
1M |
The smallest model in the V4.1 architecture family, with native vision, a causal encoder-decoder, and 1 million context. It beats V4-Pro on benchmarks and cost. From September 14, V4-Pro API traffic started routing to this model. |
| OpenAI |
GPT-6 Astra |
2026-09-03 |
1050K |
The GPT-6 flagship (Astra), built for the hardest end-to-end work: complex reasoning, coding, computer use, research, and long-horizon agents. 1.05M context. Advanced cyber capabilities first shipped through the Daybreak Blue program. |
| Gemini |
Gemini 3.8 Flash |
2026-09-02 |
1048576 |
The latest public Flash as of early September. Long-horizon coding, autonomous agents, and multi-step reasoning continue past 3.7, at the same speed and price. About 1049K context. |
| Anthropic |
Claude Fable 5.1 |
2026-09-01 |
1M |
The latest public Mythos-class tier, for demanding reasoning and ultra-long-horizon agents. Use it when high-effort Opus 5 is still not enough. |
| Qwen |
Qwen3.8-Flash-Next |
2026-08-26 |
262144 |
An open multimodal MoE (about 125B, about 6B active). Native 262K context, extendable to 1 million with YaRN. An early preview of the Qwen4 architecture. The hosted version, Qwen3.8-Flash, offers 1 million context by default. |
| Zhipu GLM |
GLM-5.3 |
2026-08-14 |
1M |
Post-training redone on the 5.2 base. Migration requires thinking to be on. Rolled out in stages for Coding Plan and the API. |
| DeepSeek |
DeepSeek-V4-Pro |
2026-08-13 |
1M |
The official V4 flagship. About 1.6T / 49B active, MIT-licensed, aimed at the strongest open-source general capability. |
| Gemini |
Gemini 3.7 Flash |
2026-08-13 |
1048576 |
A Flash-tier iteration. Coding, knowledge work, and agents are clearly stronger than 3.6 Flash. About 1049K context. It also became the model behind Gemini Spark. |
| xAI |
Grok-4.6 |
2026-08-12 |
500K |
The latest public Grok generation as of August 2026. About 500K context, for general chat, search, and reasoning. |
| Qwen |
Qwen3.8-Max |
2026-08-03 |
1M |
A native-vision MoE flagship of about 2.4T. Hybrid thinking is on by default, with 1 million context. It was Qwen's strongest model at the time. |
| DeepSeek |
DeepSeek-V4-Flash |
2026-07-31 |
1M |
The official V4 efficient tier. About 284B / 13B active, aimed at low latency and large-scale calls. |
| Anthropic |
Claude Opus 5 |
2026-07-24 |
1M |
The Claude 5 flagship. Near Fable intelligence at about half the price, for complex enterprise coding and long-horizon agents. |
| Gemini |
Gemini 3.6 Flash |
2026-07-21 |
1048576 |
The latest public Flash as of July 2026. About 1049K context, with coding and speed clearly stronger than 2.5 Flash. |
| Kimi |
Kimi K3 |
2026-07-16 |
1M |
A 2.8T open-source flagship with native vision and 1 million context, for long-horizon coding, knowledge work, and reasoning. Full weights opened on July 27. |
| OpenAI |
GPT-5.6 Sol |
2026-07-09 |
1050K |
The GPT-5.6 frontier tier (Sol). After public release it was the strongest GPT at the time, with programmatic tool calling and multi-agent orchestration. |
| OpenAI |
GPT-5.6 Terra |
2026-07-09 |
1050K |
The GPT-5.6 balanced tier (Terra). A tradeoff between intelligence and cost, suited to most production workloads. |
