GPT-6 Astra: The Jump Is in Action, Not Knowledge

OpenAI announced GPT-6 Astra on September 3, 2026. The announcement itself follows a familiar pattern: a new generation, higher scores, a new name. But the published benchmark results don’t paint the usual “better at everything” picture. The results split in two, and that split determines exactly where the model belongs.

On one set of benchmarks, the gain is a point or two. On another, scores double or nearly triple. The difference isn’t random: the benchmarks that measure knowledge are saturated; the ones that measure action are not.

Where the jump actually is

Table 1: Benchmark results as reported by OpenAI, sorted by the size of the gap.

The two ends of the table tell two different stories. The bottom end — GPQA Diamond, BrowseComp, DeepSWE — measures what the model knows and how accurately it answers in a single step. Scores above 90% here already create a ceiling effect; the remaining few points are both hard to win and of limited practical value.

The top end measures something else entirely: whether the model can carry a multi-step job all the way through. Going from 22.4% to 64.6% on Terminal-Bench Science doesn’t mean knowing the right answer better. It means completing a long technical chain without breaking it. AutomationBench’s more-than-doubling and SRE-Bench’s 32-point gap measure the same capability from different angles.

For enterprise evaluation, the implication is clear: a pilot that tests the model on question-and-answer quality will not see what changed in this generation. The difference only shows up when the model is handed a job end to end.

Computer use: from 75 minutes to 40

The metric OpenAI leads with isn’t just success rate — it’s time. In latency simulations, Astra completes OSWorld 2.0 tasks in an average of 40 minutes; the same tasks took GPT-5.6 Sol an average of 75 minutes. That works out to roughly 47% less time per task.

The jump on ScreenSpot-Pro, from 76.9% to 92.7%, is the technical explanation behind that number. This benchmark measures how accurately a model locates an interface element on screen. In a computer-using agent, most failures aren’t reasoning failures — they’re clicks in the wrong place; every misclick produces a new screen state, and the model spends extra steps recovering from it. When targeting accuracy improves, success rate and completion time improve together.

OpenAI says the model can work directly inside applications like KiCad, Excel, Blender, and Power BI. On the Codex side, context management changes as well: the model can now keep searchable notes across context windows. The feature is experimental for now and is enabled through Codex configuration. It’s an attempt to stop a full context window from being a hard ceiling on long tasks.

Astra doesn’t retire the 5.6 family

When a new flagship is announced, the assumption is that the previous generation is obsolete. The price table says otherwise.

Table 2: Astra and the GPT-5.6 family. Sol’s $4.00/$20.00 is promotional pricing available at least through November 21, 2026 — a floor, not an expiry date; list pricing is $5.00/$30.00. Batch processing runs at half price across all models, Astra included.

There’s one more option that doesn’t get its own row: GPT-5.6 Sol Ultra. It isn’t a separate model but an operating mode in which Sol runs parallel subagents that communicate with each other mid-task and merge their work into a single answer. It lifts Sol’s Terminal-Bench 2.1 score from 88.8% to 91.9% — note that this is a different and older version of the benchmark than the Terminal-Bench 4.0 in the table above, so the two numbers aren’t directly comparable. Ultra is priced the same as Sol, but a single call burns several times the tokens. The cost doesn’t rise on the rate card; it rises in consumption — a distinction that budget planning routinely misses.

Between Astra and Luna there is a 50x price gap on input and roughly 42x on output. A range of that size makes it economically indefensible for one model to serve every workload. Running Astra to classify a customer support ticket is technically possible and financially pointless.

So the enterprise takeaway from this launch isn’t a model migration — it’s a routing decision. The question to ask isn’t “which model should we switch to” but “which work goes to which tier.” Routine classification stays on Luna, high-volume production work on Terra, long-horizon agent runs on Astra. Teams that skip this distinction end up paying back the capability gain on the invoice.

One warning: requests exceeding 272,000 input tokens are billed at 2x input and 1.5x output across all tiers, Astra included, and the higher rate applies to the entire request. In long-context workflows, the listed price is not the price you pay.

The competitive picture: two flagships, two days apart

Astra’s announcement came just two days after Anthropic launched Claude Fable 5.1 and Mythos 5.1 on September 1, 2026. OpenAI’s own table contains several comparable results.

Table 3: Self-reported results from both vendors. Because they may have been produced under different conditions, direct comparison has limits.

On the 2.1-point gap in Terminal-Bench 4.0, Anthropic reports a standard error of ±1.6 to 2.0 points per model and publishes no paired confidence interval for the difference between them; it’s too early to call that a meaningful separation. On Terminal-Bench Science, the reported margin is ±3.5 to 4.5 points, and the 12-point gap there clears it comfortably.

Fable 5.1 leads on the last two rows — both of them numbers OpenAI published itself. One more detail: Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguard layers. Mythos 5.1 scores 60.9% on Terminal-Bench 4.0, which puts it above Astra.

Safety: a Critical threshold and a monitorability problem

OpenAI reports strong numbers on alignment. In scenarios testing whether the model exceeds its authorized scope, Astra goes outside that scope in 0% of cases; for Sol — tested without production safeguards — the figure was 48%. On an internal computer-use safety benchmark the error rate is 2.4% (Sol: 22.0%), and on a hallucination benchmark 4.2% (Sol: 12.2%).

Two further points in the same announcement deserve attention.

First, OpenAI reports that the model meets the “Critical” threshold for cybersecurity under its Preparedness Framework. During evaluation, the model identified two previously unknown vulnerabilities in the V8 engine. Less restrictive safeguards will be rolled out in stages through the OpenAI Daybreak program.

Second, and far less discussed: in tests that explicitly asked the model to evade monitoring, OpenAI found Astra’s written reasoning harder to monitor than Sol’s. This is a finding that moves in the opposite direction from the alignment metrics. A model committing fewer boundary violations and a model whose violations remain visible are two separate properties. When the second one regresses, audit and compliance teams give back part of what the first one gained.

Read together, these two items point the same way as Anthropic’s launch two days earlier: the industry is trying to balance capability gains with access and monitoring layers. Different vendors, different names, the same problem.

The “AGI era” claim

The most-quoted line from the launch didn’t come from a benchmark. In a press briefing ahead of the announcement, OpenAI President Greg Brockman said it wasn’t unreasonable to feel that we are now in the AGI era, and — referring to the trigger clause in the company’s old Microsoft agreement — described AGI not as a contractual threshold but as “a mission concept or spiritual concept.”

That definition puts the claim beyond verification. A threshold without a metric is a threshold you can never show was crossed or missed. The published results do document real and substantial progress on agentic workloads. But in the same result set, the gain on knowledge benchmarks stays around one to one and a half points, and the monitorability of the model’s reasoning gets worse. This is the data of a generational leap, not of a threshold crossing.

The practical conclusion for enterprise decision-makers: base the buying decision on measurements from your own workload, not on the framing in the announcement. What Astra does for you is a measurable question, independent of any definition of AGI.

Conclusion

GPT-6 Astra is not a model that pushes the knowledge frontier. It’s a model that makes action more reliable. The 42-point gap on Terminal-Bench Science, 23 on AutomationBench, and 32 on SRE-Bench all point to the same capability: finishing long, multi-step work without falling apart along the way. By contrast, gains of 1.4 points on GPQA Diamond and 1.1 on BrowseComp show that knowledge benchmarks no longer discriminate between models in this generation.

Three concrete steps stand out for enterprise teams. First, build your pilot around end-to-end task completion rate rather than question-and-answer quality; that’s the only place this generation’s difference is visible. Second, position Astra above the 5.6 family, not in place of it: a price gap of up to 50x makes routing architecture more important than model selection, and the surcharge above 272,000 tokens pushes long-context workflows past the number on the rate card. Third, read the monitorability note before you raise autonomy levels; a system that violates fewer boundaries but makes those violations harder to see carries a different risk profile in any process that requires oversight.

Two flagships were announced two days apart, and both point the same direction: the competition is no longer about what a model knows, but about how much of the work it can finish on its own, under supervision.