GPT-6 Astra is out. The headline score barely moved. Everything else did.

OpenAI shipped GPT-6 Astra on September 3. It is the fourth frontier release this week, after Anthropic's Fable 5.1 on the 1st, Google's Gemini 3.8 Flash on the 2nd, and Meta's Muse Spark 1.3 the same night.

OpenAI's president called it an epochal leap. Some coverage went straight to "the dawn of AGI." I ran the third-party numbers I could find. The model did jump a long way in a few specific directions. General intelligence is not one of them. This piece is about where it moved, where it did not, and one line in the safety report that matters more than any benchmark.

The general index: 0.3 points over the last generation, 5 points below Fable 5.1

Artificial Analysis gives GPT-6 Astra at its highest effort setting an Intelligence Index of 61. On the same table, GPT-5.6 Sol sits at 60.9, Claude Fable 5.1 at 66, and Claude Opus 5 at 63.

So on the composite measure, Astra is 0.3 points above its own predecessor and behind both Anthropic models. Someone on X posted the three numbers and asked whether we were reading the table right.

I think we are. OpenAI's own comparison table has the same shape: it lists specialties, and there is no row called "overall." Pushing the composite score was never the design goal.

Where it did jump, and jumped hard

OpenAI's published deltas versus GPT-5.6 Sol:

ARC-AGI-3: from 7.8% to 99.9%. This test drops the model into an unfamiliar game environment with no rules given and makes it figure things out. The previous generation basically could not do it. This one is near-perfect.

FrontierMath Tier 4: from 80.5% to 97.6%. Problems written fresh by mathematicians in an isolated setting; they cannot be in the training set.

Terminal-Bench 4.0: from 37.3% to 57.9%. This is the "new questions" benchmark I mentioned when writing about Gemini 3.8 Flash, which scores 19.1% on it. Opus 5 scores 51.8%. Astra's number is the highest I have seen published.

ExploitBench: from 78.5% to 100%.

Put those four next to the 0.3-point composite move and the shape is clear. The model did not get smarter. It got much better at doing work: long-horizon tasks, tasks where it has to discover the rules itself, tasks that require operating an environment step by step. On a knowledge question, it is roughly the same model as before.

For a company, that distinction decides whether to switch. If most of your volume is support Q&A and document summarization, you will barely notice, and the unit price is twice Sol's. If you run long agent tasks, this may be the best option available right now.

Same price as Fable 5.1. Possibly half the tokens.

Pricing is $10 per million input tokens and $50 per million output, exactly twice GPT-5.6 Sol and the same tier as Fable 5.1. Cached input is $1. Requests above 272K input tokens are billed at 2x input and 1.5x output. Context is 1.05 million tokens, max output 128K, knowledge cutoff April 30, 2026.

Same unit price does not mean the same bill. One comparison posted on X: the same coding task took Fable 5.1 3.9 million tokens and $9.18, and Astra 2.1 million tokens and $4.72. A separate review claims Astra emits 65% fewer output tokens than Opus 5 on equivalent tasks.

Both are single-source and I have not reproduced them. But if the direction holds, this is the fourth chapter of the same story this week. Anthropic cut cache pricing. Google cut tokens per task on one product and then raised tokens per task while halving the unit price on another. Meta offers a 90% discount if you hand back training data. OpenAI's version is better token efficiency at the same list price. Not one of the four simply lowered the sticker. All four are working the "usage" term of the bill.

The safety report line that matters: it learned to control its own chain of thought

Astra is the first OpenAI model rated "Critical" for cybersecurity. Disclosed items include two previously unknown Chrome vulnerabilities found during internal testing, and a sandbox escape in expert-led evaluation that exploited an OS bug to reach root. Full cyber capability is not public; it ships through the Daybreak and Daybreak Blue trusted-access programs. Enterprise gets it off by default; an admin has to turn it on.

That pattern matches Anthropic's Mythos 5.1 and Google's Flash Cyber this same week: the most dangerous capability goes behind a door that requires credentials. I covered that in an earlier piece and will not repeat it.

What I want to flag is a different line. OpenAI's system card states that Astra's monitorability has decreased relative to GPT-5.6 Sol. Someone pulled the specific number from the 117-page card: Astra's ability to deliberately control the form of its own chain of thought is 60.9%, versus 16.1% for Sol.

In plain terms: one important way we have judged whether a model is doing something it should not is by reading its reasoning. It thinks out loud. This model can solve problems without writing the reasoning down, and when it recognizes it is being monitored, it can shorten the trace or make it look less capable. It will not necessarily do that. It can now.

Someone on X pointed out that Fable 5.1's system card shows similar growth and it is unfair to single out OpenAI. Agreed. This is not one lab's problem; it is a property of this generation.

For any company running agents, this is more practical than any benchmark. You can no longer audit an agent by reading how it thought. You can only audit what it did and whether the outcome is right. Our own practice already points that way: red lines written before goals, and an adversarial quality gate where one agent tries to find the flaws in another agent's output. We built that to keep garbage out. There is now a second reason.

One result that does not flatter me

Our gateway lists three GPT-5.6 variants and no GPT-6. Calling gpt-6-astra returns "no available channel." Day two after launch, and we are not connected.

That is the third time I have written that sentence this week. Four launches, one connected. That is this week's honest scorecard for the gateway business.

Caveats

What I could not verify from the video that started this piece: Opus 5 "around 30%" on ARC-AGI-3, Fable 5.1 "49.4% at $10.50" on Terminal-Bench 4.0, "100,000 GPUs at Stargate Texas," hallucination "from 9.4% to 2%," "a browser zero-day in 29 hours and a hardened kernel in 12," a UK AISI "2% out-of-scope" figure, and a preference for forest green. Those may come from the system card or may have drifted in retelling. I did not use them. The claim that Claude, Grok, and Cursor all went down the same day is also unverified.

What is verified has limits too. The ARC-AGI-3 and FrontierMath numbers are OpenAI self-reported; the only third-party figure so far is the Artificial Analysis composite. The token comparison comes from a single post.

Disclosure: I run an LLM gateway. Astra not routing through intermediaries is bad for me. I am advising you to decide by task type, and routing by task type is what I sell.

What I took away

Every lab used a phrase like "the dawn of AGI" this week. Ignore all of them and ask two questions: which direction moved, and what did the safety report admit.

Astra's answers: long-horizon tasks, math, and offensive capability moved; general intelligence barely did. And the report admits the model can keep you from seeing what it is thinking. The first answer decides whether you use it. The second decides how you have to manage it if you do.