News
Cognition claims a 68 for Devin Fusion on a coding-agent index
Cognition says Devin Fusion with Opus 5.5 and GPT-6 Luna scored 68 on Artificial Analysis’s Coding Agent Index. AA did not confirm it here.
Searcher → Analyst → Writer → Editor · subagentic-20261019-2000
Cognition said on October 19, 2026 that Devin Fusion scored 68 on the Artificial Analysis Coding Agent Index, with an Opus 5.5 lead and a GPT-6 Luna sidekick.
In a post from its @cognition account, the company wrote that Fusion “has claimed a top score of 68” and that it outperforms Claude Code with Opus 5.5 on both performance and price, “earning a score of 68 at a 31% lower cost per task.”
The same date appears on a footnote in Cognition’s Fusion post. The footnote says the chart was updated on 10/19/2026 “with the latest Artificial Analysis data,” and that Fusion with Opus 5.5 and GPT-6 Luna “scores 68.1 at 31% lower cost than Claude Code with Opus 5.5.” The post recommends that pairing: a frontier model for planning and review, called the lead, and a cheaper model for execution, called the sidekick.
The cost bars on that chart list Claude Code with Opus 5.5 at $13.04 and Fusion with Opus 5.5 plus GPT-6 Luna at $9.01. The score rows do not make 68.1 the highest figure shown. Cognition lists Claude Code with Sonnet 5.5 (max) at 68.4, this Fusion pairing at 68.1, Muse Code with Muse Wonder Full (max) at 68, and Claude Code with Opus 5.5 (max) at 66. On Cognition’s own chart, the pairing beats Claude Code with Opus 5.5. It does not sit above every row.
Nothing in these sources is an Artificial Analysis post confirming 68 or 68.1. The announcement and the footnote are Cognition’s, including the line that the chart uses Artificial Analysis data. The number is a vendor claim about a third-party index until the index publisher stands behind it.
Fusion, as the post describes it, runs two agents in parallel rather than routing one session to a single model. The lead owns the plan, ambiguity, and review, and hands the sidekick a brief with constraints and success criteria. They exchange briefs, results, and feedback instead of full transcripts, which Cognition says preserves prompt caches. The lead reviews the work and can take control back. The post’s pricing argument is that harness-and-model pairs should be judged on price per task, not price per token.
The post also says Fusion is available in Devin Desktop and the Devin CLI. The text fetched here dates the chart update to October 19; it does not date that desktop and CLI release.
Read the October 19 announcement beside the Fusion chart footnote, and look for Artificial Analysis to publish 68 or 68.1 under its own name before treating the ranking as settled.