GPT-6 Astra finishes its rollout at thirteen times Gemini Flash's price

OpenAI finished the GPT-6 Astra rollout on September 8, five days after announcing the model. Plus, Pro, Business and Enterprise, 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 up to 2.5 times quicker for double that.

OpenAI@OpenAI

Astra is fully rolled out to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work.

Go build!

And if you need inspiration, watch Astra in action, live: t.co/VmBLksCNZW

Now hold that rate card next to the one Google published six days earlier. Gemini 3.8 Flash sells for $0.75 per million input tokens and $3.75 per million output. Astra costs 13.3 times as much on both sides of the meter.

And on the single coding benchmark both labs chose to report, DeepSWE v1.1, Astra scores 74.1 percent against 3.8 Flash's 73.7.

Four tenths of a point. Thirteen times the price.

So what is the money actually buying you?

The clock, mostly. Vercel's DeepsecBench, which scores models on finding real vulnerabilities in application code, put Astra at extra-high effort in first place with 37.79 — reached in 49 minutes and 47 seconds. GPT-5.6 Sol took three hours and 39 minutes to finish second at 35.44. Claude Opus 5 at max effort came third on 32.44 and spent $127.93 doing it, roughly twice Astra's $63.70. (It also logged ten false positives to Astra's two, which is the number a security team would care about first.)

Work out the cost per point and the premium stops looking like a premium. Astra runs $1.69 per point of DeepsecBench score, Sol $1.58. You are not paying for a cheaper answer. You are paying for the same answer in a quarter of the wall-clock time, with 97.9 percent precision instead of 96.3 and two false positives instead of three.

OpenAI's own launch numbers point the same way: 99.9 percent on ARC-AGI-3, 97.6 percent on FrontierMath Tier 4, and task completion on Mind2Web 1.9 times faster than the GPT-5.6 Sol experience it replaces. Speed is the headline the benchmarks agree on. Accuracy is probably not.

Our read is that Astra is priced as a long-horizon model and evaluated as a short-horizon one, which is why the comparison flatters Flash. Give a model six hours and the difference compounds. Theo Browne says he had Astra compile Super Smash Bros Melee for macOS at 120 frames per second, in a loop, in about six hours — we have not reproduced it, and it is one person's run, but it is the shape of task where a 0.4-point benchmark gap is beside the point.

Theo - t3.gg@theo

Astra was able to get Super Smash Bros Melee compiled for Mac OS, running at 120 FPS with higher resolutions and upscalable textures. Took about 6 hours in a loop.

We live in wild times. t.co/9eH6GRhr8B

The rollout itself was messier than the announcement suggests. On September 5, Tibor Blaho mapped an availability matrix in which Plus users got 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 Thibault Sottiaux spent that Friday handing out full banked usage resets. Shipping ahead of schedule has a price too. It is usually paid in support threads.

We would expect Flash-class models to stay the default inside coding harnesses through October, with Astra reserved for the runs measured in hours rather than minutes. A five-point Astra lead on a shared long-horizon benchmark would tell us we had this backwards. So would a Flash price cut.

There is a date on the other side of this, though. Gemini 3.8 Flash's $0.75 is an introductory price that expires on December 31. On January 1 it becomes $1.50 and $7.50, and the thirteen-fold gap quietly halves to under seven. Unless OpenAI moves first (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.