What GPT-6.2 changes about the cost of a finished task
Astra’s token prices fell 20%. The bigger swing in the release data comes from how many tokens a job takes.

OpenAI priced a million GPT-6.2 Astra input tokens at $8 and a million output tokens at $40, down 20% from GPT-6 Astra’s standard rates. On Agents’ Last Exam, the company says the new model scored 5.2 points higher at under a third of its predecessor’s cost. Those two discounts measure different things.1
What happens between calls
ARC Prize shows how much the surrounding software can matter. On ARC-AGI-3, Astra reached 70.3% at maximum reasoning with the standard setup. With an adapter that preserves internal reasoning between requests, the same model reached 100% at the xhigh setting. The configurations differ in both the adapter and the reasoning setting.2
What a buyer can measure
Run the workflow customers actually use, count retries and include the reviewer’s time. A cheaper token helps each attempt. Fewer abandoned attempts can change the economics of the whole job. The published benchmarks identify places to test that proposition.