One production finding sharpens Part 6. Replit reported on September 29 that Astra is the first model they have seen routinely delegate to general workers without being told to, where Fable models keep implementation for themselves. A model that spawns its own workers makes the fleet problem arrive sooner.15
On September 3 OpenAI shipped a model whose entire pitch is that it can operate a computer for hours without losing the thread. It is the best computer-use model anyone has released, it is flat on independent reasoning, and it costs exactly what Claude Fable costs.
Those three facts pull in different directions, which is why this is a guide rather than a review. Here is the whole thing in one picture.
Part 1. GPT-6 Astra Costs $1 or $100 per Million Tokens. Same Model.
The spec that makes long runs possible, and the pricing table nobody read past the first line. Cached input is a tenth of standard, there is a cliff at 272K, and batch is half price.
Part 2. GPT-6 Astra Scored 99.9% and 62.7% on the Same Benchmark.
The same model scored 62.7% and 99.9% on the same benchmark. The difference was what it carried between calls, and the better run was cheaper. The three mechanisms are yours to copy.
Part 3. GPT-6 Astra and Claude Fable Now Cost the Same. Which Task Goes Where.
Identical pricing turns a procurement decision into a routing decision. What to send each way, and three rules that survive the next model release.
Part 4. Use GPT-6 Astra to Test Your Product, Not to Build It.
The best operating guidance of launch week, why the interface is the wrong door for building and the only honest one for verifying, and the argument that UI is back.
Part 5. GPT-6 Astra Turned 10,000s of Emails Into a Knowledge Base.
A company-shaped job wearing a personal disguise. What changed is not reading comprehension, it is that reconstructing a decade of decisions became an overnight run.
Part 6. GPT-6 Astra Runs Unattended for Hours. Answer Five Questions First.
Why the OpenAI agent swarm story is a fleet problem your install already has, why scoped keys do not help when the agent drives a screen, and the five answers to write down.
Where to start
What actually shipped
A model with a 1,050,000-token context window, of which up to 922,000 can be input and 128,000 output, a knowledge cutoff of April 30 2026, five reasoning-effort levels from low to max, and ten built-in tools including computer use, a hosted shell, apply-patch, skills and MCP. Text and image in, text only out.
Note what is absent, because it constrains architecture: no fine-tuning, no embeddings, no realtime, no audio. If your plan involved tuning a model on your own corpus, that plan is not on the menu. The intended path is context and tools, not weights.
Pricing is $10 per million input and $50 per million output, which is 2.5x the GPT-5.6 Sol it replaces and identical to Claude Fable. The second comparison is the one that should drive decisions, and Part 1 covers why the headline rate is the least useful number in the table.
The one thing to take from the launch
Not "AGI is here" and not "the benchmarks are fake." Both readings skip the finding.
The most valuable number OpenAI published was the gap between 62.7% and 99.9% on ARC-AGI-3, because that gap was produced entirely by scaffolding: reasoning state carried between calls, and compaction across a long context. It made the model 37 points better, 3.66x faster and 27% cheaper on the same task set.
If a harness change is worth 37 points to a frontier lab on a benchmark, it is worth more than that to you on your own workflows, where nobody has tuned anything yet.
What actually goes wrong
It overcomplicates. This is the consistent complaint from the people who liked it most. Dan Shipper's read is that Fable still has better instincts for building a product, and the failure mode is not wrong output, it is elaborate output. Watch for this on anything where a simple answer was correct.
The bill scales with the thing you bought. Hours-long autonomous runs plus per-call tool billing is a combination that surprises people in month two, not week one. Set a budget per workflow before the first unattended run, not after.
The headline number is not your number. OpenAI's best benchmark result came from OpenAI's own harness. Your results will come from your harness. Assume the neutral figure is the honest starting point and that the gap is yours to close.
Offensive security work is gated. Astra supports defensive work, secure code review and patching, and refuses advanced offensive tasks such as writing proof-of-concept exploits. If your security team was planning red-team automation, plan around this.
What to do Monday
Do not switch your default model. Add Astra as a route, not a replacement. The independent index says your general reasoning gets no better and your bill gets 2.5x worse if you move everything.
Pick one long-horizon job and run it overnight. Nightly QA against your real UI is the highest-signal first test, because you will know by morning whether it found something your tests missed. That is a falsifiable pilot, which is rarer than it should be.
Audit what your agents carry between calls. This is the 37-point question applied to your own stack. Are you restarting context every call? Truncating instead of compacting? Making the model re-derive what it already worked out? Fix that before you shop for a better model.
Turn on prompt caching before you tune anything else. Cached input is a tenth of standard. On a long-horizon workload that re-reads the same corpus every turn, this is the largest single cost lever available and it requires no change to what the agent does.
Answer the five fleet questions in section 7 in writing. Identity per agent, what they share, what survives a shutdown, where the network boundary is, and who is made whole. An hour of work, and it is the difference between a pilot and a liability.
Write the blast-radius rule before the capability tempts you. Decide now what an unattended run is allowed to touch. A model that can operate any application for hours is only as safe as the credentials you handed it.
The launch argument was about whether this is the AGI era. The operator question is smaller and more useful: OpenAI found 37 points in the scaffolding around a model that was already trained.
Yours is untouched.
Keep reading.
Teardowns, benchmarks and playbooks on building AI-native products.
Browse the Playbook →The launch argument was about whether this is the AGI era. The operator question is smaller and more useful: OpenAI found 37 points in the scaffolding around a model that was already trained.
Yours is untouched.
Keep reading.
Teardowns, benchmarks and playbooks on building AI-native products.
Browse the Playbook →