GPT-6 benchmarks vs what it built: the scores, and the builds you can check

What benchmark scores did OpenAI publish for GPT-6?

Headline numbers that saturate several hard tests. OpenAI's launch post for GPT-6 Astra reports the following, each from OpenAI's own runs.

For GPT-6.1 Sol, OpenAI's announcement focuses on cost: it says Sol matches Astra on its DeepSWE v1.1 software-engineering suite at roughly one fifth of the cost, and scores 2.2 points above Claude Opus 5.5 on AutomationBench at medium effort.

How much better is GPT-6 than GPT-5.6?

A lot, on OpenAI's numbers and on independent ones. OpenAI reports that on its OSWorld 2.0 computer-use simulation, Astra scored 72.6% at roughly 40 minutes per task, against 65.7% at roughly 75 minutes for GPT-5.6 Sol, so it is both more accurate and nearly twice as fast. On Terminal-Bench 4.0, OpenAI's chart puts Astra at 57.9% against 37.3% for GPT-5.6 Sol, and the public leaderboard tells the same story: 58.2% for Astra, 49.4% for the first GPT-6 Sol and 37.3% for GPT-5.6 Sol.

Two improvements matter more for agents than any single score. OpenAI says a new scope test found GPT-5.6 Sol going beyond its authorised target 48% of the time without safeguards, and Astra 0% of the time. And alongside Astra, OpenAI updated the Codex harness so that, combined with Astra's efficiency, tasks on the Mind2Web benchmark finish 1.9 times faster than with GPT-5.6 Sol. Faster, more obedient agents are exactly what our build logs reflect.

Which of these numbers should a developer ignore?

Most of the headline ones, for coding purposes. FrontierMath, ARC-AGI-3, GPQA Diamond and ExploitBench test mathematics, abstract reasoning, science and security; they say Astra is very capable, not that it writes better code. The coding-relevant rows are Terminal-Bench and, with caveats, OpenAI's own DeepSWE and Agents' Last Exam.

Read the comparison columns carefully too. In OpenAI's chart, Agents' Last Exam compares Astra with Claude Opus 5, not with Opus 5.5, which was released nearly three weeks after Astra. Terminal-Bench compares it with Claude Fable 5.1 but not Opus 5.5 for the same reason. That is normal for a launch post, which can only compare against what exists on launch day, but it means a vendor chart is a snapshot of one moment. The public leaderboard, re-run as models arrive, is the better guide to where things stand now.

What do independent leaderboards say?

A closer race. The public Terminal-Bench leaderboard, checked on October 22, 2026, runs each model in its vendor's agent at maximum effort and publishes cost and tokens alongside the score.

Two lessons hide in that table. GPT-6.1 Sol reaches Astra's score for under a fifth of the cost. And Claude Opus 5.5 earns its lead by spending more than five times as many tokens, which matches what we saw in our own runs.

What did GPT-6 actually build?

Complete, working projects, fast. We ran both GPT-6 models through Codex at extra-high reasoning on the same briefs as two Claude models: a landing page, an analytics dashboard, a physics simulation and a 3D shooter, one attempt each, no edits.

Both models finished every brief in 7 to 20 minutes and in a single agent turn, with no runtime errors in our recordings. At list API prices, Astra's builds cost $2.12 to $5.64 each and Sol's $0.54 to $1.03. On the dashboard, both computed exactly the figures the dataset implies, $155,149.80 in revenue from 2,109 orders, which is the kind of correctness benchmarks are good at catching.

Where do the scores and the builds agree?

On reliability and efficiency. The benchmarks say GPT-6 completes agentic tasks well at low cost, and our builds confirm it: every GPT build worked, none ran past 20 minutes on these briefs, and the cheaper Sol matched the pricier Astra almost exactly, just as the leaderboard tie suggests.

They also agree on the gap to Opus 5.5. Opus leads Terminal-Bench by about six points while spending far more tokens, and in our builds Opus worked 3 to 11 times longer and produced the most elaborate results.

What do the benchmarks miss?

Everything visual and everything about variety. No benchmark registers that GPT-6 Astra and GPT-6.1 Sol built near-identical landing pages, with the same headline structure, the same accent colour and the same two fonts. No benchmark tells you that the GPT shooters are flatter and blockier than the Claude ones, or that Sol's player died 35 seconds into our 40-second input script while Astra's survived. These are the things you would notice in a minute of looking, and they are why this site shows builds instead of scores.

What do GPT-6's coding scores mean for your project?

It depends on what your project looks like, so translate the scores into jobs.

• Fixing bugs in an existing codebase. Strong scores on DeepSWE and Terminal-Bench make GPT-6.1 Sol a sensible, cheap default. OpenAI also says Astra in Codex can now keep notes across context windows instead of summarising them away, which helps on long debugging sessions.
• Running unattended agents. The scope-test result and the faster harness matter more than a few points of pass rate. Both GPT-6 models finished every one of our briefs in a single uninterrupted turn.
• Building interfaces, sites and games. Here the scores say almost nothing. Watch the builds: the GPT work is clean, complete and conservative, and the two models' outputs often look alike.
• Large refactors where quality matters most. The public leaderboard's gap to Claude Opus 5.5 is real, and so is Opus's much higher token use. Budget accordingly.

Is GPT-6 good for coding?

Yes, and GPT-6.1 Sol is the version to use. It ties Astra on independent coding benchmarks, finished every one of our briefs faster than Astra, and costs a fraction as much. Pick a Claude model instead when you want the most thorough first draft and can afford the time and tokens. For a deeper look at how to read any coding benchmark, see AI coding benchmarks explained.

Frequently asked questions

What are GPT-6 Astra's benchmark scores?

OpenAI reports 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, 100% on ExploitBench, 96.0% on GPQA Diamond, 72.6% on OSWorld 2.0 and 57.9% on Terminal-Bench 4.0. The public Terminal-Bench leaderboard lists Astra at 58.2%.

Is GPT-6 good at coding?

Yes. In our tests GPT-6 Astra and GPT-6.1 Sol finished every coding brief, from a landing page to a top-down MOBA, in 7 to 25 minutes with no runtime errors in our recordings. Their builds were simpler and more alike than the Claude models' builds.

Is GPT-6 better than Claude Opus 5.5 at coding?

Not on the public Terminal-Bench 4.0 leaderboard, where Opus 5.5 scores 64.8% against 58.2% for Astra. On cost and speed GPT-6 wins clearly: GPT-6.1 Sol matched Astra's score at about a fifth of the cost, and in our builds the GPT models were 3 to 11 times faster.

Should I pick GPT-6 Astra or GPT-6.1 Sol for coding?

GPT-6.1 Sol. It ties Astra on the public Terminal-Bench leaderboard at roughly a fifth of the run cost, and in our builds it finished faster than Astra on every brief while producing very similar work.