Claude vs GPT Coding Benchmarks: What the 2026 Scores Actually Say

The benchmark answer depends on the job

A useful Claude-vs-GPT coding comparison starts by naming the job. Software repair benchmarks such as SWE-Bench Pro ask for a patch that passes tests. Terminal-agent benchmarks such as Terminal-Bench measure a wider loop: planning, shell work, test iteration, environment handling, and tool coordination. Provider launch posts add another layer, because they typically report the model under a chosen internal harness and effort setting.

The Claude lineup changed twice since the July baseline. Anthropic now lists Claude Fable 5.1 as the most capable generally available model, Claude Opus 5 as the recommended starting point for most workloads, Claude Sonnet 5 as the speed and intelligence balance, and Claude Haiku 4.5 as the fastest and cheapest. Opus 4.8 and Fable 5 are legacy. The full routing picture is in the lineup report.

OpenAI meanwhile has GPT-5.5 generally positioned for agentic coding and GPT-5.6 Sol in limited preview. A comparison that ignores availability is not actionable, and one that ignores benchmark version is not even arithmetic.

Version numbers are not interchangeable

This is now the dominant failure mode in Claude-vs-GPT coding coverage. OpenAI reports GPT-5.5 at 82.7% on Terminal-Bench 2.0. Anthropic reports Fable 5.1 at 55.8% on Terminal-Bench 4.0. Those two numbers look comparable, share a benchmark name, and are not comparable at all: they come from different task sets, different difficulty calibrations, different harnesses, and different reporting parties.

The same trap exists elsewhere in the current evidence. OpenAI reports GPT-5.5 on OSWorld-Verified; Anthropic reports Fable 5.1 on OSWorld 2.0. OpenAI reports Humanity's Last Exam with tools; so does Anthropic, under its own harness. In each case the shared name invites a subtraction that the underlying data does not support.

A benchmark version bump usually exists precisely because the previous version saturated or was found unrepresentative. Scores generally drop when the tasks get harder, which means naive cross-version comparison systematically flatters whichever provider reported against the older, easier set. The only safe reading is within a version, within a harness, and ideally within a single reporting party.

What SWE-Bench Pro tells you

SWE-Bench Pro is the more useful public coding-repair reference than older Verified-only scoreboards because it was explicitly designed around contamination, task diversity, realistic underspecification, and reproducible environments. Scale defines the primary metric as resolve rate: a submitted patch has to pass fail-to-pass tests and avoid regressions on pass-to-pass tests.

That makes the public leaderboard valuable, but still not self-contained. The Scale public table includes model names, uncertainty, and the scaffold note. In the July 6, 2026 snapshot used for this site's baseline, gpt-5.4 (xHigh) appeared at 59.10 +/- 3.56 and claude-opus-4-6 (thinking) at 51.90 +/- 3.61. Both of those are now historical model versions, and neither the current Claude flagship nor Opus 5 had a comparable public row at the time of research. The absence is recorded rather than filled in.

For buyers, the strongest conclusion is not "GPT wins" or "Claude wins." It is that harder, more realistic repair benchmarks expose larger gaps between model tiers and make harness disclosure non-negotiable.

The September 2026 buying frame

For maximum Claude capability, Fable 5.1 is the named top tier at $10/$50 per million tokens. For most coding work, Anthropic itself points to Opus 5 at $5/$25 as the starting point, with escalation to Fable 5.1 when evaluations at higher effort fall short. For high-volume work where cost dominates, Sonnet 5 at a standing $2/$10 deserves a separate slot rather than being assumed inferior.

For GPT, GPT-5.5 remains the general public comparison point, while GPT-5.6 Sol should be discussed as preview-only unless the organization has access. Treat GPT-5.6 claims as relevant to frontier direction, not as a replacement option for every buyer.

A proper evaluation plan runs your own task set across Opus 5 or Sonnet 5, GPT-5.5 or GPT-5.6 if available, and the actual agent harness your team will deploy, with a fixed effort setting and a budget cap. Public benchmarks should pick the shortlist; private evals should make the purchase decision.

FAQ

Cite this page