OpenAI's pricier Astra model saves developers money despite 2.5x token costs

OpenAI's new GPT-6 Astra model carries a significant sticker shock: tokens cost 2.5 times as much as its predecessor GPT-5.6 Sol, at $10 per million input tokens and $50 per million output tokens versus $4 and $20 for Sol. Yet developers upgrading are finding their actual bills shrink, not grow.

The counterintuitive math hinges on how much reasoning the model must expend to complete a task. OpenAI's Thibault Sottiaux, engineering lead for Codex, recently noted that Astra running at its lowest reasoning setting outperforms Sol at its highest — and does so faster, delivering the first token in 2.53 seconds compared to 11.87 seconds.

Reasoning effort drives real-world cost#

The pricing puzzle deepens when you consider that adjusting the reasoning dial doesn't change per-token rates but does alter how efficiently the model solves problems. OpenAI benchmarks show the pattern clearly: on Terminal-Bench 4.0, Astra scored 57.9% to Sol's 37.3% while spending approximately 9% less per task. On GPQA Diamond, Astra nearly matched Sol's 94.6% score with 94.9% accuracy — but at an estimated 37% lower cost.

This suggests that raw token pricing alone masks the true operating expense of a model. The company is exploring this disconnect through outcome-based pricing models, which charge based on results rather than computation volume.

When medium reasoning wins#

A developer named Shinpr tested the migration on actual code generation tasks, comparing Sol at maximum reasoning against Astra at multiple levels. Astra set to medium reasoning handled implementation work in just 80 requests versus Sol's 238, processing 11.1 million input tokens against Sol's 37.8 million. The entire workflow — analysis, implementation, and review — took roughly 51 minutes and cost $25.67, while Sol required 75 minutes and $31.79.

Pushing Astra to its highest reasoning setting backfired, stretching execution to 77 minutes and $37.23 while missing a bug that medium-effort reasoning had caught.

When maximum reasoning saves#

The ARC Prize team discovered the opposite dynamic when evaluating Astra on the ARC-AGI-3 benchmark. Performance improved with more reasoning — jumping from 17.5% at low to 62.7% at maximum — but spending didn't follow a simple linear path. Low reasoning runs cost $38,166, medium $48,090, and high $40,705, but maximum reasoning came in at just $26,098.