GPT-6 Astra finishes its rollout at thirteen times Gemini Flash's price
OpenAI opened Astra to every paid tier on 8 September, and on the one coding benchmark both labs published it leads Gemini 3.8 Flash by four tenths of a point.
OpenAI's newest model beats Google's cheap one, on the single coding benchmark both labs chose to publish, by four tenths of a point.
It costs thirteen times as much.
OpenAI finished rolling GPT-6 Astra out to Plus, Pro, Business and Enterprise on 8 September 2026, five days after announcing it, in Codex and ChatGPT Work. The API price is $10 per million input tokens and $50 per million output, with a Fast mode that runs faster for double that. Gemini 3.8 Flash, published six days earlier, sells for $0.75 and $3.75. On DeepSWE v1.1 Astra scores 74.1 percent against Flash's 73.7.
So what is the money actually buying?
The clock, mostly
Vercel's DeepsecBench scores models on finding real vulnerabilities in application code, and it reports the time and the bill alongside the score.
| Model | DeepsecBench | Wall clock | Cost |
|---|---|---|---|
| GPT-6 Astra, extra-high effort | 37.79 | 49m 47s | $63.70 |
| GPT-5.6 Sol, extra-high effort | 35.44 | 3h 39m | $55.98 |
| Claude Opus 5, max effort | 32.44 | 2h 34m | $127.93 |
The top three on Vercel's DeepsecBench leaderboard: the winner is four times faster than the runner-up, at roughly the same price.
Astra also logged two false positives to Opus's ten, which is the figure a security team looks at first. Work out cost per point of score and the premium stops looking like one. Astra comes to about $1.69 a point, Sol to $1.58. You are not buying a cheaper answer. You are buying the same answer in a quarter of the wall-clock time, with slightly better precision.
OpenAI's own launch numbers point the same way. The one we would keep is task completion on Mind2Web running 1.9 times faster than the model Astra replaces. Speed is the thing every benchmark agrees on. Accuracy is probably not.
And wall-clock time is worth paying for in a way a leaderboard cannot show you. An agent that finishes inside an hour fits in a working session, a code review and a deploy window. One that takes an afternoon gets started, forgotten, and picked up tomorrow by somebody who has lost the context.
Priced as a long-horizon model, evaluated as a short-horizon one
Which is why the comparison flatters Flash. Give a model six hours and small differences compound; give it one task and they vanish into the noise.
Theo Browne says he had Astra compile Super Smash Bros Melee for macOS, in about six hours of looping. We have not reproduced that, and it is one person's run. But it is exactly the shape of job where a four-tenths benchmark gap is beside the point.
The rollout was messier than the announcement suggests. On 5 September Tibor Blaho mapped an availability matrix in which Plus users had Astra in Work and Codex but not in Chat, while Pro users found it in Chat only as "6 Pro" at maximum thinking effort. OpenAI's product lead spent that Friday handing out full banked usage resets.
Shipping ahead of schedule has a price too. It is usually paid in support threads.
What we expect, and the footnote that changes it
We would expect Flash-class models to stay the default inside coding harnesses through October, with Astra reserved for runs measured in hours rather than minutes. A five-point Astra lead on a shared long-horizon benchmark would tell us we have this backwards. So would a Flash price cut.
But there is a date on the other side of this. Gemini 3.8 Flash's price is introductory and expires on 31 December. From 1 January it becomes $1.50 and $7.50, and the thirteen-fold gap quietly halves to under seven. We noted the same pattern when Gemini 3.8 Flash kept its price and warned it would spend more tokens.
Unless OpenAI moves first, and it has cut prices at every generation so far, the cheapest way to close a price gap is to wait for the other side's footnote.
