What Artificial Analysis measured
Artificial Analysis published results for Devin Fusion CLI on October 21, 2026 at 13:10 Beijing time. Devin Fusion runs two agents in parallel, each with its own context and tools: a frontier lead model and a lower-cost sidekick. The lead keeps planning, delegation and final review. The tested setup paired Claude Opus 5.5 (max) as the lead with GPT-6 Luna (high) as the sidekick. For how Luna compares with OpenAI's larger model, see our comparison of GPT-6 Sol and GPT-6 Luna.
A three-way tie at the top
On the Artificial Analysis Coding Agent Index, the top three configurations are within 0.4 points:
| Agent and model | Index score |
|---|---|
| Claude Code, Claude Sonnet 5.5 (max) | 68.4 |
| Devin Fusion CLI, Claude Opus 5.5 (max) + GPT-6 Luna (high) | 68.2 |
| Muse Code, Muse Wonder Full (max) | 68.0 |
| Claude Code, Claude Opus 5.5 (max) | 66.0 |
Artificial Analysis describes the top three as closely matched. The figures here are from the Artificial Analysis leaderboard page; the original Artificial Analysis post on X gave Devin Fusion 68.1.
Same lead model, lower cost
The most direct comparison is with Claude Code running the same lead model alone:
| Metric | Devin Fusion (Opus 5.5 + GPT-6 Luna) | Claude Code (Opus 5.5) |
|---|---|---|
| Coding Agent Index | 68.2 | 66.0 |
| Cost per task | $9.02 | $13.04 |
| Average time per task | 55 min | 64 min |
| DeepSWE v1.1 | 72.6% | 68.4% |
| SWE-Atlas-QnA | 69.4% | 66.4% |
| Terminal-Bench 4.0 | 62.6% | 63.1% |
That is about 31% lower cost per task. Devin Fusion scored higher on DeepSWE v1.1 and SWE-Atlas-QnA and slightly lower on Terminal-Bench 4.0.
The cost figures are pay-per-token API costs and include both models' usage. They are not subscription prices, so they do not translate directly to what a Claude Code or Devin plan costs. For how Claude Code compares with OpenAI's agent, see our Claude Code vs Codex comparison.
Background
Cognition introduced Devin Fusion on June 29, 2026 (US Pacific time) as a preview, and brought Fusion to Devin Desktop and the Devin CLI on September 11, 2026 (US Pacific time). Separately, Devin can now bill some GPT usage to a user's ChatGPT plan, as covered in our report on Devin and ChatGPT plan token sharing.