Artificial Analysis’s independent run of Terminal-Bench 2.1 — a benchmark that tests an AI model’s ability to navigate and complete real tasks in a sandboxed terminal — now has GPT-5.6 Sol (running at maximum reasoning effort) narrowly ahead of Claude Opus 5 (also at maximum effort): 89.5% versus 89.1%. Other independent benchmark runs put the gap slightly wider, and Claude Opus 5’s own scoring has been complicated by its use of Opus 4.8 as a refusal fallback on a handful of tasks. But every recent version of this leaderboard tells the same story: the top two frontier coding models are now within a point or two of each other, not the multi-point gaps that used to clearly separate a “best” model from the field.
This is a meaningfully different situation to six months ago, when picking a frontier model for a coding-heavy workload was, in large part, a benchmark-chasing exercise — pick whichever lab was ahead this quarter. That gap has now compressed to the point of statistical noise for most real-world purposes.
So what
When the top models are functionally tied on raw capability, the decision that actually determines outcomes moves elsewhere: which agent integrates best with your existing toolchain, which vendor’s default security posture and permission model fit your risk tolerance (see this week’s Black Hat disclosures for why that’s not a minor point), which pricing structure fits your usage pattern, and which team actually knows how to get consistent, well-scoped output out of the tool rather than impressive demo output. For a business commissioning AI-assisted development, that’s better news than it sounds — it means the choice of AI coding agent is no longer the highest-leverage decision in the project, the choice of team is. If you want a partner who picks tooling on that basis rather than chasing whichever model topped last month’s leaderboard, see our approach to AI-assisted development or get in touch to talk through your project.