Cognition models, access and benchmark guide
Cognition model and access guide with a practical US-English evaluation brief and links to publisher materials.

👁 111 reads
Local Fusion is an agent workflow, so a ranking measures the tool environment as well as the language model. For an engineering purchase, keep repository permissions, tests and the actual billing unit with the claimed task score.
Cognition’s Fusion lets the operator name a lead model and a sidekick. The lead plans and reviews, while its partner carries out implementation. The coding benchmark row therefore describes a working combination with a harness, not a new foundation model released by Cognition. For a shop considering a different pair, I would run the same repository revision, preserve the proposed plan and price all attempts through the first accepted change. Comparing only the lead model’s public token rate leaves some of the bill unaccounted for.
Choosing the first job
Fix a malformed CSV export in a throwaway repository, preserving quoted commas and empty middle columns.
Models in the November 17 snapshot
| Model and track | Placement | Reported score |
|---|---|---|
| Devin Fusion CLI - Claude Opus 5.5 Max + GPT-6 Luna High (AA coding agent) | 2 | 58.4% benchmark score |
Release record. Publisher documentation identifies the product family. The ranked configurations include Devin Fusion CLI - Claude Opus 5.5 Max + GPT-6 Luna High.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Sources
- https://cognition.com/blog/local-fusion (accessed November 17, 2026)
- https://docs.devin.ai/ (accessed November 17, 2026)
- https://artificialanalysis.ai/agents/coding-agents (accessed November 17, 2026)
More on this from QUANTSLANT
benchmarks·China Mobile models, access and benchmark guideChina Mobile model and access guide with a practical US-English evaluation brief and links to publisher materials.