OpenAI Ships GPT-6 Astra: 63% Lower Cost Per Task

OpenAI has introduced GPT-6 Astra. The company claims a price per task 63% below the nearest alternative, along with two results on the Terminal-Bench benchmark.

Is the gap real or just noise?

On Terminal-Bench 4.0, GPT-6 Astra scores 57.9% against 55.8% for Claude Fable 5.1. A 2.1-point difference is the noise of a single run, not superiority. On Terminal-Bench Science the numbers are 64.6% against 52.6%. Here the gap widens to 12 points, and it holds on domain-specific tasks. The claimed price per task is 63% lower.

What does this change in the budget?

Price per task is the one number in this release that lands in the budget immediately. If agent workflows cost a company a fixed monthly sum, a 63% cut per task changes the whole calculation. A press release proves nothing here. The way to check it is to run a set of typical in-house tasks through the model API. The invoices are then compared across two weeks at identical volume. To size up the order of magnitude, use the model switching calculator.

What has not been announced yet?

Official token pricing, context window limits, the general availability date for the API, and licence terms are all still unannounced. Payment options for Russian companies have not been described either.