OpenAI’s new model costs 2.5x more per token — and developers are saving money anyway
The New Stack Amanda Caswell ● Covered by 19 sources
OpenAI’s new Astra model costs 2.5x more per token than Sol. But some tests show it finishes faster and cheaper anyway, if you dial reasoning down.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has a pricing headache that might not be one. GPT-6 Astra costs 2.5 times as much per token as GPT-5.6 Sol, yet the company is pushing developers to upgrade and then lower the reasoning setting instead of treating the sticker price as the whole story.
Thibault Sottiaux, engineering lead for Codex, said over the weekend that Astra on low reasoning performs better than Sol on high reasoning. Artificial Analysis backs up that claim, at least a little: Astra-low scored 49 on its Intelligence Index, just ahead of Sol-high at 48. It also started answering much faster, with first-token latency of 2.53 seconds versus 11.87 seconds for Sol-high.
The raw token math still looks nasty. Astra is priced at $10 per million input tokens and $50 per million output tokens, while Sol sits at $4 and $20. But OpenAI says the reasoning dial changes how much work the model does before it finishes, and the company is leaning hard on the idea that fewer output tokens can make a pricier model cheaper in practice. Its own benchmarks are meant to show that.
On Terminal-Bench 4.0, Astra scored 57.9% against Sol’s 37.3% and still came out about 9% cheaper per task. On GPQA Diamond, Astra barely beat Sol, 94.9% to 94.6%, while costing an estimated 37% less. Those are not universal guarantees, but they do undercut the simple “higher price, higher bill” argument developers usually start with.
A developer named Shinpr found a similar pattern on a real codebase. Using Sol-high and several Astra reasoning levels across analysis, implementation, and review, Astra-medium finished the implementation in 80 requests, compared with 238 for Sol-high, and used 11.1 million input tokens instead of 37.8 million. The whole Astra-medium run took about 51 minutes and cost an estimated $25.67; Sol-high took roughly 75 minutes and cost $31.79. Pushing Astra to high was worse in that test, stretching to 77 minutes and $37.23, and it missed a startup bug that medium caught.
ARC Prize saw the other side of the coin. On its ARC-AGI-3 harness, Astra’s score rose from 17.5% at low reasoning to 62.7% at max, with an xhigh setting in between, but cost did not move in a straight line. ARC Prize spent $38,166 at low, $48,090 at medium, $40,705 at high, and only $26,098 at max. In other words, more reasoning can cost more, less, or less than you’d expect, depending on how many actions the agent needs to solve the task.
OpenAI is trying to make that flexibility easier to use with configuration_update, which lets an app change reasoning effort between responses without changing the original request setup. For now, that works only with Astra in standard single-agent requests. The practical message is simple: don’t stare at the token price and assume you know the bill.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI pricing people keep pretending is mysterious: the model with the scary per-token rate can still be the cheaper one if it stops wandering off like a distracted intern. OpenAI is basically admitting that token pricing is too dumb for agentic work, and that’s the real story. Flat rates are fine for spreadsheets; they’re less convincing when the model is burning time, retries, and tool calls like a tiny chaos engine.
Read more about this at: The New Stack