| xAI |
Grok-4 Fast |
2025-09-19 |
2M |
A high-speed version that uses far fewer thinking tokens. Context can expand to 2 million, with much lower cost per unit of intelligence than early frontier models. |
| OpenAI |
GPT-5-Codex |
2025-09-15 |
400K |
A GPT-5 variant optimized for agentic coding. It became the default Codex model and is strong at repo-scale changes and long-horizon coding. |
| Kimi |
Kimi K2-0905 |
2025-09-05 |
256K |
A K2 instruction-tuning update. Context rose to 256K, with more stable coding and tool calling. |
| DeepSeek |
DeepSeek-V3.1 |
2025-08-21 |
128K |
One model that supports both thinking and non-thinking modes. Tool use and agent tasks are clearly stronger than V3/R1. |
| OpenAI |
GPT-5 |
2025-08-07 |
400K |
A generation that unified the fast model and deep-thinking model. A realtime router sets thinking depth. It became ChatGPT's default, and the API offers 400K context. |
| OpenAI |
GPT-5 mini |
2025-08-07 |
400K |
A mid-size GPT-5 that keeps unified routing and reasoning. Price and latency are a better fit for large-scale apps. |
| OpenAI |
GPT-5 nano |
2025-08-07 |
400K |
The smallest GPT-5 size, aimed at high throughput and low cost, while still offering basic reasoning. |
| Anthropic |
Claude Opus 4.1 |
2025-08-05 |
200K |
An incremental Opus 4 upgrade. More stable on complex reasoning, analysis, and creative work, and the transitional flagship before 4.5. |
| Zhipu GLM |
GLM-4.5 |
2025-07-28 |
128K |
A native-agent hybrid-reasoning flagship, 355B / 32B active, MIT-licensed. Coding and tool use are the core selling points. |
| Qwen |
Qwen3-Coder |
2025-07-22 |
256K |
A repo-scale agent coding specialist. Long context and tool calling are optimized for software engineering. |
| Kimi |
Kimi K2 |
2025-07-11 |
128K |
An open-source MoE flagship of about 1T parameters, with about 32B active. Focused on agents, tools, and coding, and seen as a major open-source milestone. |
| xAI |
Grok-4 |
2025-07-09 |
256K |
A multimodal flagship with native tool use and realtime search. The API offers 256K context, and Grok-4 Heavy launched with it. |
| xAI |
Grok-4 Heavy |
2025-07-09 |
256K |
The highest-compute Grok-4 tier for SuperGrok Heavy. It spends more compute for higher reasoning quality. |
| Gemini |
Gemini 2.5 Flash |
2025-06-17 |
1M |
The official 2.5 high-speed tier. Thinking budget is adjustable, for a flexible tradeoff between cost and intelligence. |
| DeepSeek |
DeepSeek-R1-0528 |
2025-05-28 |
128K |
A mid-cycle R1 update with stronger reasoning and fewer hallucinations. It also added JSON output and function calling. |
| Anthropic |
Claude Opus 4 |
2025-05-22 |
200K |
The Claude 4 flagship, focused on long-horizon agents and coding, with extended thinking. It was positioned as the strongest coding model at the time. |
| Anthropic |
Claude Sonnet 4 |
2025-05-22 |
200K |
The Claude 4 balanced tier. A broad upgrade over 3.7 in coding, instruction following, and thinking stability, also available to free users. |
| Qwen |
Qwen3 |
2025-04-29 |
131072 |
A hybrid thinking / non-thinking generation, including MoE sizes such as 235B-A22B, fully Apache-2.0 open source. |
| OpenAI |
o3 |
2025-04-16 |
200K |
The full o-series reasoning flagship. It can agentically use all tools in ChatGPT and reached the then-highest level on math, science, and complex tasks. |
| OpenAI |
o4-mini |
2025-04-16 |
200K |
A cost-efficient reasoning model released with o3. Full tool-calling support, for cases that need reasoning without o3's cost. |
