The mechanism is simple. The model ID is gpt-6-astra. Pricing: ten dollars per million input tokens, fifty dollars per million output; cached input drops to one dollar per million; batch pricing cuts the standard rate in half; Fast mode doubles it. Context: 1,050,000 tokens, with maximum output at 128,000. Knowledge cutoff is April 30, 2026. OpenAI staged the rollout over the week following the September 3 announcement, and by the afternoon of September 6, the API was live for all Pro, Enterprise, and Business Premium users worldwide.

The blast radius is where this becomes a builder's question. The benchmark numbers OpenAI published are for coding task evaluations: GPT-6 Astra solved 88.0 percent of tasks in a single attempt and 99.2 percent within four. Its predecessor, GPT-5.6 Sol, solved 55.9 percent single-attempt and 68.7 percent within four. That is not a marginal improvement. The gap between Sol and Astra on these metrics is larger than the gap between GPT-4 and GPT-4o was on most practical benchmarks. If you are running a coding agent at scale, you have a cost-versus-performance question on your desk this week. At ten dollars per million input tokens, Astra is not cheap relative to Sol. But the performance delta is large enough that the math often closes quickly on complex tasks, where single-attempt success rate directly determines how many API calls you are paying for.

The safety note from OpenAI's announcement is worth reading carefully. The company added safeguards following the Hugging Face breach earlier this summer. Sam Altman told CNBC those measures "sufficiently minimize the risk of severe harm for release." Minimize, not eliminate. Sufficient, not certain. That is the considered public posture at this capability level, and it is the right one to name precisely rather than round up.

The pattern is what makes this interesting beyond the model itself. OpenAI has been systematically compressing the time between major model announcement and broad API availability. GPT-4 took weeks to reach wide API access. GPT-4o took days. GPT-6 Astra took three. The trend is deliberate. The faster a model is in the API, the faster the independent evaluation community runs it, benchmarks it, and publishes comparisons. OpenAI is betting on speed-to-ecosystem as a structural advantage: if your model is billed before competitors have published their response, you set the framing for the conversation.

The read: Anthropic shipped Mythos 5.1 on September 1. Anthropic has not published a direct benchmark comparison against Astra as of this writing. OpenAI has. That asymmetry is a choice, and both choices make sense given each company's positioning, but one of them will look better in approximately thirty days when the independent eval papers come in. The question is whether Mythos 5.1 outperforms Astra on the tasks that matter to builders, and right now only one lab is on the record with numbers.

The builder's move: run gpt-6-astra on the specific task types you care about before committing to a stack change. The 88 percent single-attempt number is a ceiling, not a floor. What it looks like on your domain is a separate experiment. Start with a batch evaluation, watch the cost curve, and pay attention to cached-input pricing if your agent repeatedly reads a large system context.