Anthropic vs OpenAI: whose flagship models build better? 32 builds, no scores

Which models represent Anthropic and OpenAI here?

Each company's two strongest generally available models for building software, run in each company's own coding agent.

Prices are from Anthropic's pricing page and OpenAI's GPT-6.1 Sol announcement. Anthropic also offers Claude Mythos 5.1, the same model as Fable 5.1 with different safeguards, only through trusted access programmes, according to its announcement.

How do Anthropic's and OpenAI's models work differently?

Like two engineering cultures. The Claude agents work iteratively: they write, open the result in a headless browser, look, and fix, over dozens of steps. On our briefs that meant 41 to 174 turns and 9 to 128 shell commands per build. The GPT agents work in one long, considered pass: a single turn of 10 to 33 commands, with almost all of their input served from the prompt cache.

The results follow from the method. Claude's builds are larger, take longer and include more of what nobody asked for. GPT's builds are compact, arrive in minutes and stick closely to the brief.

Whose models build better websites and apps?

Anthropic's, if "better" means more distinctive; OpenAI's, if it means faster to a clean result. On the landing page brief, the two Claude models produced two different brands and the two GPT models produced near-twins with the same typefaces. On the phone finance app, all four completed the send-money flow correctly, with Claude Opus 5.5 adding haptic feedback and a full-screen confirmation, and the GPT apps sharing a layout and fonts. On the open brief, where each model invented a product, the Claude models wrote stories, a night train and a sky lamp, and the GPT models both invented small gadgets for using your phone less.

Whose models build better games?

Anthropic's look better; OpenAI's finish sooner. The Claude shooters are atmospheric, with neon lighting and more interface; the GPT shooters are block-built and plainer. Claude Opus 5.5 was still refining both its shooter and its MOBA when our 150-minute limit arrived. Under our scripted input, the player died before the script ended in three of the four GPT game runs, and in Claude Opus 5.5's shooter and Claude Fable 5.1's MOBA too.

What does the MOBA brief reveal about each company's models?

That each vendor's two models think alike. Both Claude MOBAs are called Riftline, and both give their hero a dash called Blink Step. Both GPT MOBAs end on an ultimate called Starfall, and so does Claude Fable 5.1's, which shows how much these models share across company lines. In execution the split is familiar: the two Claude lanes are dense, with shops, minimaps and named enemy heroes, and took 131 minutes and the full 150; the two GPT lanes are lighter and took 20 and 25 minutes.

What does each company charge for the same work?

Much more at Anthropic, per build. At list API prices, OpenAI's sixteen builds would have cost about $37 in total, $31 for Astra and $6 for Sol. Each Claude build cost between $7 and $25, so a single brief built by both Claude models can cost as much as all sixteen GPT builds together. Subscriptions blur this, since both companies sell plans at $20 a month, but agent work eats into Anthropic's usage limits much faster.

Do the companies' launch claims hold up in the builds?

Mostly, yes, which is more than launch claims usually manage. Anthropic says Claude Opus 5.5 "performs at the level of Claude Fable 5.1 on most work" and that a tester who had several Claude models build a game from a single prompt found Opus 5.5's the strongest on graphics and polish, in its Opus 5.5 announcement. Our builds agree on both counts: Opus matched or exceeded Fable on every brief both finished, and its games are the most polished on the site.

OpenAI says GPT-6.1 Sol delivers "near-Astra intelligence for a fifth of the price". On our briefs that undersells it for building: Sol's output was often near-identical to Astra's, at roughly a fifth of the cost per build or less. Where the claims overreach is in the comparisons each company chooses to show, not in the products.

How fast are the two companies shipping?

Very fast, and in step. In September 2026 alone, Anthropic released Claude Fable 5.1 on September 1 and Claude Opus 5.5 on September 22, with Claude Sonnet 5.5 following within a week, while OpenAI released GPT-6 Astra on September 3 and 4, GPT-6 Sol and Luna on September 22, GPT-6.1 Sol on September 29 and GPT-6 Instant on October 7, according to Wikipedia's GPT-6 entry and the public Terminal-Bench leaderboard's release dates. Any comparison between the two has a shelf life of weeks, which is why every number on this page carries its date and every build its model version.

How do their safety approaches show up when building?

Quietly, which is the point. OpenAI reports that on a new test of whether a model oversteps its authorised scope, GPT-6 Astra did so in 0% of cases against 48% for GPT-5.6 Sol without safeguards. Anthropic reports that Opus 5.5 scored best of any model on its automated behavioural audit and is much less likely to take hard-to-reverse actions. In our runs both agents worked with full permissions, but inside an empty folder outside any real project, which is the way we would recommend running either company's agent unattended, whatever its safety record. For agents you leave alone with real code, these numbers matter as much as any benchmark.

What do the benchmarks say about the two companies?

They agree with the builds on direction. The public Terminal-Bench 4.0 leaderboard puts Claude Opus 5.5 first at 64.8%, the two GPT-6 models tied at 58.2% and Claude Fable 5.1 at 57.9%. It also shows the cost: Opus 5.5 used about 8 billion tokens for its run against 1.5 billion for each GPT model. Each company's own launch posts claim leadership on benchmarks it selected; OpenAI's post for GPT-6 Astra compares it with Claude models released before it, which is normal and worth knowing.

Which company should you bet on for building?

Both, for different jobs. Use Anthropic's models through Claude Code when the first attempt has to be impressive and you can give it an hour. Use OpenAI's through Codex when you want many quick, cheap attempts. The two companies are converging on price at the top and diverging on method, and the method is what you will feel day to day. Before committing to either, run the blind test on the brief closest to your work.

Frequently asked questions

Is Anthropic or OpenAI ahead in AI coding?

On the public Terminal-Bench 4.0 leaderboard, Anthropic's Claude Opus 5.5 leads at 64.8% against 58.2% for OpenAI's GPT-6 models. On cost and speed, OpenAI leads clearly: its models finished our builds several times faster and far more cheaply.

What is the difference between Anthropic and OpenAI models?

In our builds, Anthropic's Claude models worked longer, tested their own output more and produced more elaborate, more varied results. OpenAI's GPT-6 models worked in one efficient pass and produced complete, consistent results that often looked alike.

Which company has the better coding agent, Claude Code or Codex?

Claude Code produced the richer builds and Codex the faster, cheaper ones. Our Codex vs Claude Code page compares them brief by brief.

Are Anthropic and OpenAI models priced similarly?

At the top end, yes: Claude Fable 5.1 and GPT-6 Astra both cost $10 per million input tokens and $50 per million output tokens. Below that, OpenAI's GPT-6.1 Sol at $2 and $10 undercuts Claude Opus 5.5 at $4 and $20.