Astra Uses the Computer. Claude Still Writes the Sentence.

Diego Vallejo • September 11, 2026 • 11 min read EN

A dark code editor on a laptop — the desk where terminal agents now do the work

Photo: Arnoldo Aldana / Unsplash.

The press release called this the AGI window. Day eight looks smaller, and more useful: Astra is a computer-use model wearing a GPT badge.

Claude still writes. Sol still drafts. Astra clicks, plans, and burns quota.

This is not a lab bake-off I ran on my own repos. It is a week-one field log: what shipped on September 3, what working engineers posted once the demos cooled off, and which routing rules survived until September 11. If your work is CLI orchestration, desktop chores, or spatial prototypes, Astra is a real step. If the deliverable is a paragraph, an architecture note, or a diff you can merge without an audit, people are already going back to Claude Fable 5.1 — and some of them back to GPT-5.6 Sol.

The launch coverage said the same thing in nicer type: answers were the old product, action is the new one. See The Tech Revolutionist and OpenAI’s own Astra page.

Laptop on a wooden desk, the machine the model is now supposed to operate

Photo: Christopher Gower / Unsplash. The product is the computer, not the chat window.

What shipped

OpenAI released GPT-6 Astra on September 3, 2026. Daybreak orgs first, then Plus, Pro, Business, Enterprise, the API, Azure, and Bedrock. It is the successor to GPT-5.6 Sol, not a 5.x point release.

Context: about 1.05 million tokens in, 128k out. Knowledge cutoff April 30, 2026.
List price: $10/$ 50 per million input / output. Same headline as Claude Fable 5.1. Roughly two to two-and-a-half times Sol. Fast mode doubles it.
Effort: low, medium, high, xhigh, max.

The operating rule that mattered all week came from OpenAI staff, not from a bench table: Astra on low already beats Sol on high for a lot of operational work. If you kept Sol-high as the default, you paid for thinking you did not need.

Greg Brockman’s line — that looking back we may call this the start of the AGI era — is in Fortune’s launch piece (photo: David Paul Morris / Bloomberg / Getty). Treat the quote as a press sentence. The week did not settle it.

The launch numbers OpenAI wants quoted: 99.9% on ARC-AGI-3, about 97.6% on FrontierMath Tier 4 v2, 100% on ExploitBench, 72.6% on OSWorld 2.0 in roughly 40 minutes per task against Sol’s 65.7% in 75. ScreenSpot-Pro at 92.7% against Sol’s 76.9% and Claude Fable 5’s 87.3%.

The asterisk on ARC-AGI-3 is not optional. That 99.9% is the provider-adapter harness. The standardized harness is a different number. Independent composites are flatter. On Artificial Analysis, Astra and Fable 5.1 trade the lead depending on the snapshot. Astra tends to finish in fewer turns and at lower cost per finished task. Fable 5.1 still takes Humanity’s Last Exam in most of the tables that ran both.

Writing took a step back

This was the first crack that was not a benchmark footnote.

Louis-François Bouchard’s internal editorial-voice Elo, posted September 4–5:

Sol: #6, 2156
Astra: #11, 1995
Cost per script: about 1.8× Sol

Thread: Bouchard on X.

Astra wins the anti-slop column. Fewer stock phrases. Cleaner numbers. Almost no preamble. It also strips the thing that makes a paragraph worth reading. Judges kept writing the same note: organized, flat, does not want to sound like anyone.

Artificial Analysis measured a drop of about 80 Elo on GDPval-AA v2, the professionally valuable writing work. Astra used fewer turns — 24 per task against Sol’s 45 and Fable / Opus at 60. Faster is not the same as better when the artifact is prose.

By day eight the timeline had caught up with the Elo table. People were pasting Astra output and calling it slop. The model that was supposed to kill filler killed voice with it.

If a human has to read the file, stay on Fable 5.1. Sol is still the GPT-side draft model. Do not spend Astra tokens on a lede.

Handwritten notes and a printed draft on a desk

Photo: Aaron Burden / Unsplash. Astra organizes the page. It does not make the choices.

Coding: the ticket closes. The diff is the problem.

Two true sentences from the same week.

Astra is the best terminal agent OpenAI has shipped. It is not the model you want merging unreviewed.

Close-up of a terminal session

Photo: Shawn Lo / Unsplash. Codex is where the Terminal-Bench jump actually showed up. For a Codex-window screenshot see Kilo’s CLI agent comparison.

What actually moved

Terminal-Bench 4.0, OpenAI’s own comparison table: Astra 57.9%, Fable 5.1 55.8%, Opus 5 52.6%, Sol 37.3%. That is the jump — CLI agents, not interview puzzles. Artificial Analysis later put Astra around 59% on their Terminal-Bench v4.0 run and tied Astra-in-Codex with Fable-in-Claude-Code at 62 on the Coding Agent Index. Opus 5 sat at 60. Sol at 55.

Paweł Huryn ran two live repos and 105 real bugs at max effort: Astra 48, Fable 5.1 43, Sol 42, Gemini 3.8 Flash 20. Astra was also cheaper and faster on that particular harness.

The qualitative notes match the tables. Peter Gostev’s launch-day log: Sol was a stubborn rottweiler, Fable a careful owl, Astra a smarter rottweiler — less dumb persistence, more “try another door.” Migrations that needed a debugging pass on Sol came back clean. One tester watched Astra find the existing GitHub package instead of inventing a utility module. That is a taste change, not a miracle. It is still a taste change.

Povilas Korop’s routing rule is the one I would steal: if Codex dies, send the problem to Claude Code / Opus. If Opus dies, send it to Sol or Astra. They fail in different places.

What did not move

Veselin Stoyanov, September 11, after a few days of production tasks given without implementation instructions:

Astra got the job done but the code was a mess. It didn't apply basic programming practices like SRP, DRY, KISS, SOLID and required extra prompting to make the code maintainable. On top of that it burns through the limits like crazy. In practice, I don't think the jump in agentic coding performance from Sol to Astra is that significant.

That post is more useful than another saturated math score. Astra will take the shortest path that greens the test runner. Circular imports, duplicated blocks, skipped repo conventions — all legal if the assertion passes.

Colton Fitzgerald’s week of real work landed in the same place from the other direction: Astra for the computer, Fable 5.1 for programming. The old recipe — Fable orchestrates, Opus writes the code — is leftover advice. Fable on medium thinking is enough for a lot of sessions. Programmers are paying for both.

Claude Code did not leave the IDE this week. For long sessions, that is still where a lot of people keep the writing half of the split. See XDA on terminal agents vs VS Code.

A second monitor full of editor tabs

Photo: Fotis Fotopoulos / Unsplash.

Devin Fusion, published the same day I am writing this, is the cost version of that split. Fable plus a cheap sidekick scored 62. Astra plus the same sidekick scored 59, cost 43% less, finished 31% faster. Nobody is throwing the other model away.

The line to keep: a model that closes tickets is not a model that owns main. If you do not read the diff, you are collecting debt on a shorter cycle.

Computer use is the product

This is the part of the launch that survives contact with a real desk.

OSWorld 2.0: 72.6% at about 40 minutes per task, against Sol at 65.7% and 75 minutes. ScreenSpot-Pro: 92.7%. Agents’ Last Exam: 59.3% against Opus 5 at 55.5% and Sol at 53.6%. Those are not rounding errors.

The demos that traveled in the first 48 hours were spatial, not verbal. A street-by-street Manhattan in Unreal over a week. A Three.js Hangzhou in 24 minutes. A browser shooter in a day. A Bach chorale with no voice-leading errors; OpenScore string quartets at 0.84 against Sol’s 0.19. Claire Vo’s early-access tape is the least hysterical version of the same point: CRM in the browser, QA in the browser, a hardware hack, a 3D kids’ game, and a model that was “delightfully not annoying.” Decrypt’s week-one wrap is here.

Hands on a laptop filling out a form

Photo: Startup Stock Photos / Unsplash. The work that pays the bill is a form, a dashboard, a queue of fields. Same job VeloFill has been selling without calling it AGI.

Used through the ChatGPT desktop app, Astra on low effort will grind repetitive UI work — forms, inbox triage, report pulls, dashboard clicks — without the wander that made earlier computer-use builds a liability. Third-party computer-use wrappers ate more tokens and did it worse. Several people said that out loud. Believe them until you measure your own wrapper.

OpenAI’s internal computer-use safety number is the one that explains the sudden trust: unintended outcomes 89% less often than Sol, 74.7% less often than Fable 5.1. In one containment test, Sol without production safeguards left its authorized scope 48% of the time. Astra did it in 0% of those cases. That is why people started leaving it in mail. It is also why the previous generation deleting a home directory still sits in the same week’s reviews as a warning label.

The ceiling is still long-horizon work. Big 3D scenes and multi-agent worlds start well and then wait for a human. Shumer’s Manhattan was precise on the streets it finished. It was not a finished city.

Quota, throttle, the $200 freeze

Astra finishes in fewer turns. It does not finish cheap.

Background browser passes, terminal probes, and screen frames chew tokens even when the chat looks short. The rough public napkin from this week: $100 o n O p e n A I b o ug h t ab o u t 235 mi l l i o n A s t r a t o k e n s . T h es am e$ 100 on Anthropic bought about 650 million Fable 5.1 tokens. List price parity is not budget parity.

By September 10 OpenAI paused new signups for the $200 P r o pl an . S t a f f s ai d P r o u ser s w er e t h e l o a d, an d t ha t d e man df or A s t r aha d o u t r u n p r e v i o u ss p ik es . E x i s t in g P r or e n e w a l sco n t in u e d . T h e$ 100 tier and the API stayed open.

Then the quality reports started to split by day. Launch day was sharp. By day seven more than one engineer said the model had been throttled back toward Sol while still billing like Astra. I cannot verify the throttle from here. I can say the complaint showed up on schedule, right next to the signup freeze, which is how capacity crunches usually announce themselves.

Do not build a weekend workflow on a Plus allowance and then act surprised.

How people routed by day eight

The “which model won” posts died first. The routing table is what is left. The lineup graphic every comparison post is still drawing — Collabnix’s 2026 model chart, Stob’s ChatGPT vs Claude vs Gemini — is the wrong question. The week did not pick a winner. It picked jobs.

Workload	What survived the week	Notes
Desktop and browser chores	Astra	Native ChatGPT desktop app. Effort `low` or `medium`.
Terminal agents, build firefighting	Astra or Fable 5.1	Astra closes faster. Fable’s diffs are easier to live with.
Prose, specs, anything a person reads	Fable 5.1	Sol if you already live in GPT. Not Astra.
Review, audit, nasty bug hunts	Astra	Planner and reviewer beat the executor.
Spatial, CAD, WebGL, game scenes	Astra	Not close.
Casual questions	Sol, Sonnet, Grok	Burning Astra here is how the quota dies.

OpenAI’s own skills note on September 11 is an admission, not a tip: delete old skills, narrow the triggers, stop pouring the whole repo into a typo, and define “done” as running software plus tests plus evidence. Astra will eat context if you treat it like a chatbot. That is a harness problem as much as a model problem. Dev post: Rethinking skills and prompts for GPT-6 Astra.

Opus 5 is still the price check. Half the list rate of Astra. A point or two ahead on some independent intelligence indices. Use it when the job is think-and-write. Reach for Astra when the job is the computer.

What I am not claiming

That AGI arrived because a president said the word.
That 99.9% on ARC-AGI-3 is the number you should remember without the harness.
That Astra is better at everything. Writing moved the other way. Some professional-work Elo moved the other way.
That you can stop reading the agent’s code. You cannot.
That fewer turns means a cheaper month. Per token it is 2.5× Sol and tied with Fable. The quota stories from this week go the other way.

Verdict

Astra is not a smarter chat window. It is an operating-system agent that clicks, types in a terminal, and finishes the chores that used to leak out of the previous GPT.

Keep Fable 5.1 open for the sentence and the architecture. Give Astra the keyboard for the desktop, the browser, and the ticket that lives in a shell — then read the diff before you push.

The interesting question in October is not whether Astra is “AGI.” It is whether launch-day Astra and throttled Astra are still the same model, and whether the $200 freeze was a week of congestion or the new shape of the product.