🤖 Autonomous Agent / Tool Use frontier
As of November 16, 2026, the highest Autonomous Agent / Tool Use score on this frontier is Gemini 4 Carbon at 95.2, with an estimated monthly cost of $180.00.
An agent that runs multi-step jobs on its own: shell commands, code changes, tool calls. Weighted toward agentic benchmarks that grade finished tasks (Terminal-Bench, SWE-bench) plus novel-problem reasoning, with a slice for self-correction because a long run fails if the model cannot recover from its own mistakes. Cost matters because agent loops resend large contexts on every step.
The y axis is this scenario's weighted score; the x axis starts at estimated monthly cost for 500 requests a day. Switch axes to compare against speed, latency or context. See the full ranking.
On the frontier
| Model | Score | Est. $/month |
|---|---|---|
| Gemini 4 CarbonGoogle | 95.2 | $180 |
| GPT-6.1 TerraOpenAI | 90.0 | $90 |
| Gemini 3.8 FlashGoogle | 79.7 | $68 |
| DeepSeek V4.1 Flash 1017DeepSeek | 77.0 | $24 |
| Claude Haiku 5.5Anthropic | 72.3 | $18 |
| GPT-6 LunaOpenAI | 71.7 | $9 |
| DeepSeek V4 Flash 0731DeepSeek | 65.0 | $9 |
| GPT-5 nanoOpenAI | 28.4 | $6 |