| xAI |
Grok-3 |
2025-02-17 |
1M |
A reasoning-era flagship trained at Colossus scale. It offers Think mode, with context up to 1 million tokens. |
| xAI |
Grok-3 mini |
2025-02-17 |
1M |
A lighter reasoning version of Grok-3 that keeps Think mode, for lower cost and faster responses. |
| Gemini |
Gemini 2.0 Pro |
2025-02-05 |
2M |
The 2.0 experimental flagship, focused on coding and complex prompts, with context up to 2 million. |
| OpenAI |
o3-mini |
2025-01-31 |
200K |
The first reasoning model opened to free ChatGPT users. Near o1 on math and coding, with faster speed and lower price. |
| Qwen |
Qwen2.5-Max |
2025-01-28 |
131072 |
Alibaba Cloud's hosted flagship MoE. It targeted then-current first-tier closed models, for general chat and complex tasks. |
| Kimi |
Kimi K1.5 |
2025-01-20 |
128K |
The start of the numbered model line. Stronger multimodal reasoning and RL, and the start of Kimi's thinking path. |
| DeepSeek |
DeepSeek-R1 |
2025-01-20 |
128K |
A reasoning model trained with pure RL. Open-sourcing it drew global attention. R1-Zero and several distilled sizes shipped with it. |
| DeepSeek |
DeepSeek-V3 |
2024-12-26 |
128K |
A 671B MoE (about 37B active) with FP8 training and MTP. It became the base for later R1 and the app. |
| Gemini |
Gemini 2.0 Flash |
2024-12-11 |
1M |
The default high-speed model of the 2.0 era. Stronger native tools, multimodal output, and realtime capability. Veo 2 launched with it. |
| OpenAI |
o1 |
2024-12-05 |
200K |
The full o1 reasoning model, launched with ChatGPT Pro. The API opened on December 17, and context rose to 200K. |
| Anthropic |
Claude 3.5 Haiku |
2024-10-22 |
200K |
The 3.5-generation fast model. Speed near Haiku and intelligence near early Sonnet, for low-latency production traffic. |
| Qwen |
Qwen2.5 |
2024-09-19 |
131072 |
The most widely used open-source generation. Full sizes from 0.5B to 72B, with broad gains in coding, math, and instruction following. |
| OpenAI |
o1-preview |
2024-09-12 |
128K |
Preview of the o-series reasoning models. It runs an internal chain of thought before answering and was clearly stronger than then-current GPT-4o on math, coding, and science. |
| OpenAI |
o1-mini |
2024-09-12 |
128K |
A lighter reasoning version of o1, focused on STEM and coding. Lower cost and faster, for cases that need some reasoning but not the strongest results. |
| DeepSeek |
DeepSeek-V2.5 |
2024-09-05 |
128K |
Merged V2-0628 and Coder-V2-0724 so general chat and coding share one model. |
| xAI |
Grok-2 |
2024-08-13 |
128K |
A full-generation Grok upgrade with stronger coding and writing, plus realtime X search and image generation. |
| OpenAI |
GPT-4o mini |
2024-07-18 |
128K |
A cost-efficient small multimodal version of GPT-4o. It replaced GPT-3.5 Turbo in ChatGPT and fits high-volume lightweight tasks. |
| Anthropic |
Claude 3.5 Sonnet |
2024-06-20 |
200K |
A breakout Sonnet generation. Coding, vision, and tool use far surpassed Claude 3 Opus, with excellent value. |
| DeepSeek |
DeepSeek-Coder-V2 |
2024-06-17 |
128K |
A code-specialized model based on V2. Covers more programming languages and is strong at long-context repo Q&A. |
| Qwen |
Qwen2 |
2024-06-06 |
131072 |
The third Qwen generation, stronger in multilingual and long-context work. 7B and above generally support 128K. |
