Claude Opus 5.5 review: eight one-shot builds, unedited

What is Claude Opus 5.5?

Anthropic's recommended default model, released on September 22, 2026 as the first model of the Claude 5.5 family. Anthropic says it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5", in its announcement. It costs $4 per million input tokens and $20 per million output tokens, with cached reads at $0.20, according to Anthropic's pricing page, and offers a 1-million-token context window with up to 128,000 output tokens.

Anthropic's launch notes include two claims directly relevant here: a tester who had several Claude models build a game from a single prompt found Opus 5.5's strongest on graphics and polish, and Opus 5.5 "generates output more than 30% faster than Opus 5". On the public Terminal-Bench 4.0 leaderboard, Opus 5.5 in Claude Code ranks first at 64.8%.

What did Claude Opus 5.5 build?

Eight projects, one prompt each, in Claude Code at maximum effort, with no human edits.

1. Landing page. "Easy to book. Hard to interrupt." in Fraunces and Geist, with a working booking calendar in the hero and an unrequested sign-up flow. 43 minutes.
2. Analytics dashboard. Correct figures, an annotated revenue peak and average line, and a chart-to-table switch. 81 minutes, after a first attempt was stopped when the account hit its usage limit.
3. Physics simulation. Glossy numbered balls in a glowing hexagon, with a small physics readout in the corner. 40 minutes.
4. 3D shooter. Iron Tide, a neon city at night with a minimap and a pulse carbine. Opus was still refining it when our 150-minute limit arrived, so its files were published as they stood; its recording ends in a game over at 25 seconds.
5. Phone finance app. A balance card with a month sparkline, haptic feedback on supported phones and a full-screen "$42.50 sent" confirmation. 69 minutes, after three earlier attempts were interrupted by an API outage, a usage limit and a stopped harness.
6. One-lane MOBA. Riftline, a stone plaza with a glowing rune circle opening onto a lamp-lit forest lane, with a shop, a minimap, an enemy hero called Vexis, the Void Herald, and abilities running Radiant Bolt, Rift Eruption, Blink Step and a beam ultimate called Oblivion Ray. Like the shooter, it was still being refined when our 150-minute limit arrived; the files on disk were published, and the hero survives our whole 60-second script.
7. Kanban board. A compact board that tracks "3 of 14 done" in its header, tilts the card you drag, flags overdue cards in red and saves card edits automatically, strikingly close to Fable 5.1's board. 101 minutes, after a first attempt was cut off by an API outage.
8. Open brief. Nod Night Railway, "The 22:40 to Nowhere": a sleeper train that loops through the night and returns you to the same platform, with berths from £58 to £230 and a section of synthesised train sound. 74 minutes and the largest single page on the site, 178 KB.

How did Claude Opus 5.5 work?

In many small, checked steps. Across its eight builds Opus used 81 to 174 turns per build and 23 to 128 shell commands, the most of any model on the site. A typical loop: write a section, start a local server, load the page or game in a headless browser, take a screenshot, read it, fix what it saw. On the MOBA it wrote seventeen script files and was still running that loop at the 150-minute mark.

That loop is expensive in input tokens, because each step re-reads the growing project, but most of those reads are cache hits at $0.20 per million. For the six builds where Claude Code reported a total, Opus cost about $87 at list prices, against about $111 for Claude Fable 5.1's eight builds. Opus is cheaper per token and, on most briefs, did more with it.

What is Claude Opus 5.5 good at?

Ambition that holds together. Opus's builds do more than the brief asks without falling apart: every finished build logged no runtime errors in our recordings, and the extras it adds are the kind a senior developer would add, demos, feedback, transitions, readable data. Where a brief leaves room for invention, Opus invents the most: its open-brief site is a complete little world with a timetable, prices, reviews and sound.

It also tests relentlessly. On the landing page it read files 59 times and ran 28 shell commands, about a dozen of them loading the page in a headless browser; on the shooter it ran 118 commands. That loop is why its builds arrive polished.

Where does Claude Opus 5.5 fall short?

Time, tokens and knowing when to stop. Opus took the longest on every brief it shared with the other models and was the only model to hit our 150-minute limit, twice: on the shooter and on the MOBA. It reads its growing project on every step, about 10 million input tokens for one landing page and about 70 million each for the shooter and the MOBA, which is why the builds where Claude Code recorded a total cost $7 to $21 each at list prices despite a low token price. On a subscription, that appetite meets usage limits quickly; six of its attempts across our briefs were interrupted by infrastructure, an API outage, usage limits and a stopped harness, and none by the model.

How does Opus 5.5 compare with Claude Fable 5.1?

Denser, slower and, per token, cheaper. Fable 5.1 costs $10 and $50 per million tokens against Opus's $4 and $20, but because Opus works longer the two cost about the same per build. Opus built more on most briefs; Fable built calmer, more restrained pages faster. Anthropic's own positioning matches what we saw: Opus as the default, Fable for long, demanding work.

How does Opus 5.5 compare with the GPT-6 models?

More of everything, at several times the time and cost. GPT-6 Astra and GPT-6.1 Sol finished every brief in 7 to 25 minutes for $0.54 to $7.27 per build at list prices, with clean, consistent results that often matched each other closely. Opus's builds are more distinctive and more complete, and it leads the public Terminal-Bench leaderboard by about six points. Opus 5.5 against GPT-6 Astra puts all eight briefs side by side.

Who should use Claude Opus 5.5?

Anyone handing off a build that has to be right the first time: client-facing sites, polished prototypes, interactive tools you will demo. Pair it with a cheaper model for quick iterations, run one heavy session at a time on a subscription, and give it a clear scope, because Opus fills whatever room you leave it.

Frequently asked questions

Is Claude Opus 5.5 good?

It is the most thorough builder we have tested. It completed every brief, produced the most elaborate results, and leads the public Terminal-Bench 4.0 leaderboard at 64.8%. It is also the slowest model on our briefs, running into our 150-minute limit twice.

Is Opus 5.5 better than GPT-6 Astra?

For building, it produces richer, more distinctive work and scores higher on Terminal-Bench 4.0, 64.8% against 58.2%. Astra is several times faster and cheaper per build. Which is better depends on whether you value the first draft or the time and cost.

Is Opus 5.5 cheaper than Opus 5?

Yes. Anthropic prices Opus 5.5 at $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5, with cached reads 60% cheaper at $0.20, and says typical workloads cost about 40% less.

Is Opus 5.5 better than Fable 5.1?

On our builds, usually yes: Opus built more on most briefs at a lower token price, and Anthropic itself says Opus 5.5 performs at Fable 5.1's level on most work. Fable is faster and more restrained, which some briefs reward.