Devin Fusion pairs Opus 5.5 with GPT-6 Luna and beats Claude Code on Opus 5.5 for 31% less per task
Cognition launched Devin Fusion late on October 19: Opus 5.5 as the lead, GPT-6 Luna as a cheaper sidekick that takes the grunt work. Artificial Analysis ran it independently on October 21 and the numbers held up. Devin Fusion scored 68.1 on the AA Coding Agent Index, basically tied at the top with Claude Code on Sonnet 5.5 at 68.4 and Muse Code on Muse Wonder Full at 68.0. Against Claude Code on Opus 5.5 at 66.0, it costs $9.01 per task instead of $13.04 and averages 55 minutes instead of 65. Index: https://artificialanalysis.ai
Look at what is being ranked. The top three entries are a Claude harness, a Meta harness and a Cognition harness, and the Cognition one mixes models from two rival labs. The score belongs to the pairing and the routing logic, not to either model. Breakdown from AA: Fusion leads on DeepSWE v1.1, 72.6% versus 68.4%, and trails slightly on Terminal-Bench 4.0, 62.1% versus 63.1%. So the cheap sidekick costs a point on terminal work and buys back a third of the bill.
The same week, the plumbing for this kind of mixing got built at hyperscaler level. OpenAI announced its frontier models and Codex agentic workflows are coming to Google Cloud's Vertex Model Garden, billable against existing GCP commits, and said they will soon be available in Google Antigravity and the Gemini Agent, Google's own agent harnesses. Microsoft remains the primary cloud partner, OpenAI added. Link: https://openai.com/index/openai-models-google-cloud/
A year ago the debate was which lab's model you bet on. Now Google's harness will run OpenAI's model, Cognition's harness runs Anthropic and OpenAI together, and the index leader is whoever routes best. Model loyalty is becoming a cost line. The harness is where the switching cost lives.
← Back to all articles
Look at what is being ranked. The top three entries are a Claude harness, a Meta harness and a Cognition harness, and the Cognition one mixes models from two rival labs. The score belongs to the pairing and the routing logic, not to either model. Breakdown from AA: Fusion leads on DeepSWE v1.1, 72.6% versus 68.4%, and trails slightly on Terminal-Bench 4.0, 62.1% versus 63.1%. So the cheap sidekick costs a point on terminal work and buys back a third of the bill.
The same week, the plumbing for this kind of mixing got built at hyperscaler level. OpenAI announced its frontier models and Codex agentic workflows are coming to Google Cloud's Vertex Model Garden, billable against existing GCP commits, and said they will soon be available in Google Antigravity and the Gemini Agent, Google's own agent harnesses. Microsoft remains the primary cloud partner, OpenAI added. Link: https://openai.com/index/openai-models-google-cloud/
A year ago the debate was which lab's model you bet on. Now Google's harness will run OpenAI's model, Cognition's harness runs Anthropic and OpenAI together, and the index leader is whoever routes best. Model loyalty is becoming a cost line. The harness is where the switching cost lives.
Comments