GPT-6 Astra ships at $10/$50 per million tokens

OpenAI shipped GPT-6 Astra on September 3 at $10 per million input tokens and $50 per million output tokens. That is 2.5x the current promotional rate on GPT-5.6 Sol and roughly eight times what Gemini 3.8 Flash costs. Before you write it off on the GPT-6 Astra pricing alone, run the arithmetic OpenAI is quietly betting on: the per-token price went up, and the per-task price may not have.

What actually happened

Astra is rolling out to a limited set of organizations first, then to ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API as gpt-6-astra, and Amazon Bedrock (OpenAI). Enterprise access is off by default at launch — an admin has to turn it on. Fast mode runs at up to 2x the speed of Standard for 2x the price. Zero Data Retention is supported for eligible API customers.

The benchmark numbers OpenAI leads with are cost-adjusted, not raw. On Terminal-Bench 4.0, Astra scores 57.9% against Claude Fable 5.1's 55.8% at an estimated 9% lower cost per task. On Agents' Last Exam it hits 59.3% versus Claude Opus 5's 55.5% while using about 65% fewer output tokens. On Terminal-Bench Science it beats Fable 5.1 by 12 points at ~31% lower estimated cost. Per The New Stack, the $10/$50 rate matches Anthropic's Fable 5.1 pricing and sits far above Meta Muse's $1.25/$4.25 and Gemini 3.8 Flash's $0.75/$3.75 introductory rates.

There is a second detail buried in the launch post that matters more than the price. Astra is the first OpenAI model to meet the Critical cybersecurity threshold under its Preparedness Framework, so OpenAI is running misalignment classifiers in production that halt unauthorized activity. In ChatGPT and Codex, a paused task asks you to review. In the API, the task stops. OpenAI says these checks can interrupt legitimate work.

Why per-task pricing matters for your business

Token price is a unit cost for something nobody buys. Nobody buys a million tokens. You buy a resolved support ticket, a migrated table, a reconciled invoice. If a $50/M model finishes in one pass and a $20/M model needs three attempts plus a human correction, the expensive model is the cheap one.

So measure the unit you actually sell. Log cost per completed task, per successful task, and the human-minutes spent fixing failures. Most teams we work with have never split those three, which is why they argue about token rates instead of margins. OpenAI's own launch data is too thin to prove the savings hold on your workload — the only benchmark that settles it is yours.

Then handle the stop. A model that halts mid-run on a safety classifier is not a timeout and not a 500. Your retry logic almost certainly treats it as one, which means you will burn budget re-running a task that will be stopped again. Catch that condition specifically, route it to a human queue, and count it.

Sources: OpenAI, The New Stack.