GPT-6 Astra review: eight one-shot builds, unedited

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's most capable model, released to the public on September 4, 2026, as the first member of the GPT-6 family. OpenAI calls it "the world's most intelligent and aligned model" in its launch post and highlights computer use, reasoning and professional work. On the API it costs $10 per million input tokens and $50 per million output tokens, with a 1,050,000-token context window and up to 128,000 output tokens, according to its model page.

We ran Astra in OpenAI's Codex agent at extra-high reasoning on eight briefs, one attempt each, under the same rules as every other model on this site: the same prompt word for word, a 150-minute cap and no human edits. The methodology page has the exact command.

What did GPT-6 Astra build?

Eight complete projects. Here is every one, with the facts from its run log.

Brief by brief:

1. Landing page. "Less busy. More in flow." in DM Sans and Instrument Serif on cream, with a terracotta accent, a working yearly switch and a sign-up dialog. Ten minutes.
2. Analytics dashboard. "Good things are growing." over a calm green layout with correct figures, sortable and filterable orders and a dark mode. Ten and a half minutes.
3. Physics simulation. Titled Kinetic, "Order, in motion": small pastel balls in a thin-lined hexagon with a live stats panel. All twenty balls are inside the hexagon at the end. Eight minutes.
4. 3D shooter. Dead Signal: a block-built industrial yard with a door marked Relay. The player survives our whole 40-second input script. Twenty minutes, two files.
5. Phone finance app. "Hello, Jordan." with a four-step send flow marked 01 of 04; sending $42.50 to Maya Patel updates the balance to $4,783.69. Nine minutes.
6. One-lane MOBA. Emberfall, a low-poly forest lane with a river, towers and a hero called Aeris whose ultimate is Starfall. The hero dies once during our 60-second script. Twenty-five minutes, five files.
7. Open brief. It invented peep, a pocket ceramic bird whistle, on a sunny yellow page headed "Small bird. Big mood." Thirteen minutes.
8. Kanban board. A sidebar layout with a "Small steps. Good things ahead." card, working drag and drop, undo and filters. Twelve and a half minutes.

How did GPT-6 Astra work?

In one long, quiet pass per brief. Codex ran each Astra build as a single agent turn of 10 to 30 shell commands: reading the brief, writing the files, and running a few checks before finishing. Across all eight builds Astra read about 9.4 million input tokens, 94% of them served from the prompt cache, and wrote about 340,000 output tokens including its reasoning. For comparison, Claude Opus 5.5 read about 10 million input tokens for a single landing page.

That economy is the most interesting thing about Astra as a builder. It does not loop through repeated browser checks the way the Claude agents do; it reasons up front and commits. The result is predictable timing, a predictable bill and builds that rarely contain anything unexpected, for better and for worse.

What is GPT-6 Astra good at?

Delivering the whole brief, quickly, without drama. Every Astra build includes every feature the prompt listed, and it did each in a single uninterrupted agent turn. Its builds are compact, between 24 KB and 95 KB of source, which makes them easy to read and modify. Where a brief contained data, such as the dashboard's orders or the finance app's balance, Astra computed the right numbers and kept them consistent after an interaction.

It is also economical by frontier standards. Our eight Astra builds cost about $31 at list prices in total, less than two Claude Opus 5.5 builds.

Where does GPT-6 Astra fall short?

Distinctiveness, and value against its sibling. Astra's builds share a house style: warm neutrals, a serif display face, a gentle motivational line somewhere in the interface. That style is pleasant and safe, and it repeats across briefs. More strikingly, GPT-6.1 Sol produced near-twins of Astra's work on several briefs, the same landing page structure and fonts, a physics demo with the same name, a MOBA whose hero also has an ultimate called Starfall, though Claude Fable 5.1's hero shares that one too.

The games show the second limit. Astra's shooter and MOBA are complete and stable, but flatter and less atmospheric than the Claude models' versions, and its MOBA hero falls to the enemy during our scripted push.

How does GPT-6 Astra compare with GPT-6.1 Sol?

Almost the same output for five times the token price. On the public Terminal-Bench 4.0 leaderboard the two tie at 58.2%, with Sol's run costing about $600 against Astra's $3,300. On our briefs, Sol was faster on six of eight and Astra on two, the finance app and the kanban board, and the builds look alike. Astra's one clear win was the shooter, where its player survived our 40-second script and Sol's died at 35 seconds.

How does GPT-6 Astra compare with Claude?

Faster and cheaper, less ambitious. Claude Opus 5.5 leads the public Terminal-Bench leaderboard at 64.8% and, on our briefs, took far longer per build while producing richer, more distinctive work. Claude Fable 5.1 costs the same per token as Astra, $10 and $50 per million, but worked two to four times as long on most briefs and cost more per build. If you are choosing between Astra and a Claude model for a build, watch the pair on the brief closest to your work: Opus 5.5 against Astra shows all eight.

What would GPT-6 Astra cost you?

Less than you might expect per build, more than you need to pay. At list API prices our eight builds averaged just under $4 each, from $2.12 for the physics simulation to $7.27 for the MOBA, the largest build at five files. Because 94% of the input was cached at $1 per million tokens rather than $10, the output tokens at $50 per million made up most of each bill.

On a subscription the per-build cost disappears into your plan's limits: OpenAI made Astra available to ChatGPT Plus, Pro, Business and Enterprise users. The catch is that GPT-6.1 Sol produced comparable builds for about $0.80 each on average, so on the API every Astra build costs roughly four to six Sol builds.

Who should use GPT-6 Astra?

People whose work is mostly not building. OpenAI's strongest claims for Astra are about computer use and reasoning: 72.6% on its OSWorld 2.0 simulation and saturated scores on FrontierMath and ARC-AGI-3, per its launch post. If you use ChatGPT for research, documents and operating software, Astra on a Plus plan is a strong default. If you mainly want projects built, GPT-6.1 Sol gives you the same builds for a fraction of the price, and a Claude model gives you more ambitious ones for more time and money.

Frequently asked questions

Is GPT-6 Astra good?

Yes. It completed all eight of our build briefs, from a landing page to a top-down MOBA, in 8 to 25 minutes each with no runtime errors in our recordings. Its work is clean and complete, but rarely more ambitious than the brief, and very close to what the cheaper GPT-6.1 Sol produced.

How much does GPT-6 Astra cost?

$10 per million input tokens and $50 per million output tokens on the API, with cached input at $1. At those prices our eight Astra builds cost between $2.12 and $7.27 each, about $31 in total. In ChatGPT it is available on the Plus, Pro, Business and Enterprise plans.

Is GPT-6 Astra better than GPT-6.1 Sol?

Not on our briefs. The two produced strikingly similar builds, tie on the public Terminal-Bench 4.0 leaderboard at 58.2%, and Sol costs a fifth as much per token. Astra was faster on two of our eight briefs and its shooter player survived where Sol's did not.

Is GPT-6 Astra better than Claude Opus 5.5?

It is much faster and cheaper; Opus 5.5 builds more elaborate results and leads the public Terminal-Bench 4.0 leaderboard at 64.8% against Astra's 58.2%. On our briefs Opus took roughly 4 to 8 times as long per build.