Devin Fusion with Claude Opus 5.5 (max) and a GPT-6 Luna (high) sidekick scores 68 on the Coding Agent Index, 2 points above Claude Code with Opus 5.5 (max) at 31% lower cost per task
@cognition's Devin Fusion runs a frontier lead model alongside a lower-cost sidekick model. The th their own context and tools, with the lead keeping planning, delegation and final review. We benchmarked Devin Fusion CLI with Claude Opus 5.5 (max) + GPT-6 Luna (high).
It scores 68.1 on the Artificial Analysis Coding Agent Index, closely matched at the top with Claude Code with Claude Sonnet 5.5 (max) at 68.4 and Muse Code with Muse Wonder Full (max) at 68.0. Against Claude Code with Claude Opus 5.5 (max) at 66.0, it costs $9.01 vs. $13.04 per task and averages 55 vs. 65 minutes per task.
Across the three benchmarks, the Fusion configuration scores higher on DeepSWE v1.1 (72.6% vs. 68.4%) and SWE-Atlas-QnA (69.6% vs. 66.4%) and slightly lower on Terminal-Bench 4.0 (62.1% vs. 63.1%). Its results reflect both models working together, including their combined API cost. Costs are pay-per-token API costs, not subscription prices.
@cognition's Devin Fusion runs a frontier lead model alongside a lower-cost sidekick model. The th their own context and tools, with the lead keeping planning, delegation and final review. We benchmarked Devin Fusion CLI with Claude Opus 5.5 (max) + GPT-6 Luna (high).
It scores 68.1 on the Artificial Analysis Coding Agent Index, closely matched at the top with Claude Code with Claude Sonnet 5.5 (max) at 68.4 and Muse Code with Muse Wonder Full (max) at 68.0. Against Claude Code with Claude Opus 5.5 (max) at 66.0, it costs $9.01 vs. $13.04 per task and averages 55 vs. 65 minutes per task.
Across the three benchmarks, the Fusion configuration scores higher on DeepSWE v1.1 (72.6% vs. 68.4%) and SWE-Atlas-QnA (69.6% vs. 66.4%) and slightly lower on Terminal-Bench 4.0 (62.1% vs. 63.1%). Its results reflect both models working together, including their combined API cost. Costs are pay-per-token API costs, not subscription prices.
Oct 21, 2026 · 5:10 AM UTC
119
22
562
74,828
Explore the Coding Agent Index, task costs and benchmark results:
artificialanalysis.ai/agents/coding-…
Methodology:
artificialanalysis.ai/methodology/co…
artificialanalysis.ai/agents/coding-…
Methodology:
artificialanalysis.ai/methodology/co…
4
4
18
8,041
@ArtificialAnlys @cognition sonnet 5.5 edging out the fancy combo is funny, it's the one i use daily
1
0
3
415
@ArtificialAnlys @cognition Two points on one index is close. I'd want the spread across reruns first.
0
0
1
517
@ArtificialAnlys How much of the bill goes to Opus reviewing the sidekick's work?
0
0
1
511
@ArtificialAnlys @cognition Impressive fusion achieving top performance with significantly lower cost efficiency
0
0
1
253
@ArtificialAnlys @cognition Pairing models keeps winning lol
0
0
1
226
@ArtificialAnlys @cognition Lead + cheaper sidekick 这个拆法很实在——很多团队其实不是缺最强模型，是缺「贵活贵模型、杂活便宜模型」的路由。CLI 侧如果也能一个入口切模型，体感会更接近这条曲线。
0
0
0
12
@ArtificialAnlys @cognition I think you should be clear about being sponsored by Cognition. There are a bunch of other services that do the same thing and they don't get your 'privileged' spot in the AA Coding Agent Index.
0
0
0
26
@ArtificialAnlys @cognition Fusion pairing cuts cost while matching top coding scores
0
0
0
41
@ArtificialAnlys @cognition Combining frontier lead with efficient sidekick seems smart for cost
0
0
0
162
@ArtificialAnlys @cognition Lead plus sidekick model cutting cost 31% while scoring higher is smart
0
0
0
5
@ArtificialAnlys @cognition 68 vs 66 is within noise tbh. 31% cheaper is the only number that matters here
0
0
2
930
@ArtificialAnlys @cognition 2 分的领先在 benchmark 噪声范围内，真正值得看的是 cost per task 那条线：便宜三成、分数还涨了，说明 coding agent 的瓶颈不在单模型智商，而在怎么分工——生产环境里没人会为 2 分多付三成的钱。
0
0
0
310
@ArtificialAnlyse@cognitionly Devin run alongside this, with the same budget and tools. Comparing Fusion with Claude Code changes both the model team and the harness. That extra baseline would help isolate how much the sidekick adds, especially where Terminal-Bench moves the other way.
0
0
1
223
@ArtificialAnlys @cognition so the best setup is opus planning while a cheaper model does the grunt work. AI just reinvented middle management 😭
0
0
0
66
@ArtificialAnlys @cognition Cost per task is the number I'd want next to this, but counted per merged PR. For me a task isn't done until the PR carries the plan, the reasoning and verification screenshots. Does the Index pass a run on tests alone, or on something a reviewer can check?
0
0
0
52
@ArtificialAnlys @cognition makes sense, let the cheap model do the boring parts
0
0
0
106
@ArtificialAnlys @cognition You can run the lead + sidekick split in Claude Code today: a subagent in .claude/agents/ takes a model: field in its frontmatter, so file reading and search run on a cheaper model while the main thread stays on Opus. Pays off only if it returns a summary, not the files.
0
0
0
49
@ArtificialAnlys @cognition the cheaper sidekick doing most of the work is the part I care about, same pattern works outside coding, the expensive model only gets the calls that need it
0
0
0
70
@ArtificialAnlys @cognition been doing this by hand for months. opus plans and reviews, a cheaper model writes the boring parts. nice to see it actually works on paper too
0
0
0
36
@ArtificialAnlys @cognition How cleanly does the orchestrator decide which subtasks get delegated to Luna versus keeping Opus in the loop?
0
0
0
30
@ArtificialAnlys @cognition Detail in your own numbers: Claude Code with Sonnet 5.5 (max) scores 68.4, a hair above Fusion's 68.1. So the 2-point gain is against Opus in Claude Code. What's the Sonnet setup's cost per task? That comparison shows whether the lead-plus-sidekick design pays.
0
0
0
1
@ArtificialAnlys @cognition Frontier model plans and reviews, cheaper model does the typing, and it wins on cost. Agents have independently rediscovered the tech lead.
0
0
0
62
@ArtificialAnlys2@cognitionhe index is an equal-weight average of 3 benchmarks, and Terminal-Bench has only 66 tasks. One task flipping there moves the index by about 0.5
So the gap is roughly 4 tasks, and the methodology page lists no error bars. The 31% cost difference is the solid number here
So the gap is roughly 4 tasks, and the methodology page lists no error bars. The 31% cost difference is the solid number here
0
0
0
132
@ArtificialAnlys @cognition lead + sidekick beating solo opus while costing less is the real headline.. expensive model plans and reviews, cheap one does the grunt work. feels like every harness ends up here
0
0
0
73
@ArtificialAnlys @cognition Incredible efficiency, Fusion approach delivers top performance at significantly lower cost
0
0
0
13
@ArtificialAnlys @cognition The lead model plus a cheap sidekick setup makes a lot of sense to me. It's kind of how I already split things, Claude Code with Opus for the heavy stuff and a local model behind Hermes for the rest. Curious how much of the 31% comes from the sidekick vs just fewer retries?
0
0
0
54
@ArtificialAnlys @cognition 55 min per task is the number i care about. that's a long time to stare at a terminal. i built a thing so i can talk to claude code, codex or pi out loud while they grind instead. open source if anyone wants something to do during minute 37: github.com/CakeCrusher/fu…
0
0
0
205
@ArtificialAnlys @cognition This matches what I see running Claude Code with agent teams. I've got wrappers that add ChatGPT and Grok as teammates.
The cheaper model doing the grunt work in its own context is where the savings come from. The lead stays clean and just plans and reviews.
The cheaper model doing the grunt work in its own context is where the savings come from. The lead stays clean and just plans and reviews.
0
0
1
189
@ArtificialAnlys @cognition Strong performance combining efficiency with high capability
0
0
0
54
@ArtificialAnlys @cognition was literally about to try sending test reviews to a cheaper model
nice to see it scores higher too
nice to see it scores higher too
0
0
0
42
@ArtificialAnlys @cognition Same idea as my setup: CC builds, Codex reviews. Instead of choosing btw providers just combine them 🫶
0
0
0
40