GPT-6 Astra: Price, Access and What the Benchmarks Show

OpenAI released GPT-6 Astra on September 3, 2026 at $10 per million input tokens, $1 per million cached input tokens and $50 per million output tokens. The API model is gpt-6-astra, with a 1,050,000-token context window, 922,000 maximum input tokens and 128,000 maximum output. ChatGPT access is rolling out over several days to Plus, Pro, Business and Enterprise rather than appearing everywhere at once.

The headline score is 99.9% on ARC-AGI-3. The more useful story is wider: 72.6% on OSWorld 2.0, 64.6% on Terminal-Bench Science and 88% pass@1 on the public SRE-Bench, all reported in OpenAI’s launch material. Astra also trails Claude Fable 5.1 on the Artificial Analysis Intelligence Index shown by OpenAI and trails several models on Humanity’s Last Exam with tools. This is a specialist frontier model with broad capability, not an automatic replacement for every cheaper route.

GPT-6 Astra is available through the OpenAI API and, according to OpenAI, Amazon Bedrock. OpenAI names GPT-6 Astra Pro for ChatGPT Pro, Business and Enterprise, but its launch sources do not publish a separate Astra Pro API model ID, price or specification sheet. Plus receives Astra, not the Pro variant. Enterprise access is off by default and requires an administrator to enable it.

The rollout begins with a limited set of organisations and expands over several days. Zero Data Retention is supported for eligible API customers; “eligible” matters because it is not a universal promise. OpenAI does not disclose Astra’s parameter count, architecture, training compute or training-data size. Its published April 30, 2026 knowledge cutoff is not the same thing as a training-data cutoff.

The name appeared before the product. We covered Astra’s August mathematics disclosure and OpenAI’s pre-launch cyber-risk assessment. Those articles record what was known then; today is the first broad product launch.

Astra and Claude Fable 5.1 share the same headline input and output rates: $10 and $50 per million tokens. They are not identically priced. OpenAI lists Astra cache reads at $1 per million; Anthropic lists Fable 5.1 cache reads at $0.25. GPT-5.6 Sol’s current promotional rates remain much cheaper at $4 input, $0.40 cached input and $20 output in OpenAI’s published rate card.

The API accepts text and image input and produces text. It supports Responses, Chat Completions and Batch, but tool calling requires the Responses API. Realtime, Assistants, fine-tuning, embeddings and native image, video or audio generation are not supported. The model exposes low, medium, high, xhigh and max reasoning; there is no none setting.

OpenAI’s clearest launch-day gains sit in environments where a model has to inspect state, act, recover and finish. Astra scores 59.3% on Agents’ Last Exam against 53.6% for Sol, and 72.6% on the OSWorld 2.0 offline partial score against Sol’s 65.7%. OpenAI also reports roughly 40 minutes per OSWorld task for Astra versus 75 for Sol, though runtime depends on the harness as well as the model.

The cyber and professional-work rows point in the same direction. OpenAI reports 95.9% geometric overlap on BenchCAD, 41.4% on AutomationBench and 39% on a recent-vulnerability V8 evaluation where Sol scored 5.5%. Astra also reached 88% pass@1 and 99.2% pass@4 on SRE-Bench. OpenAI says the V8 work found two previously unknown vulnerabilities and that both are being disclosed to maintainers; the operational details do not belong in a product comparison.

Most launch scores came from OpenAI’s research environment or API, not an independent replication of production ChatGPT. OpenAI takes each model’s best score at any reasoning effort, so the table says which configuration reached the highest point, not which one reached it for the least money or time. Research and ChatGPT also use different system prompts, tools and inference settings.

Astra loses rows in OpenAI’s own appendix. It scores 57.2% on Humanity’s Last Exam with tools, behind Sol at 65% and Fable 5.1 at 63.8%. Its 64.5% on FrontierCode Extended trails Fable 5 at 64.9%, while its 67 on the Artificial Analysis Coding Agent Index trails Fable 5 at 68.1. On the broader Artificial Analysis Intelligence Index, Astra’s 61.2 trails Fable 5.1 at 65.7.

ARC-AGI-3 is the sharpest example. OpenAI explains that Astra’s 99.9% used a Responses API harness with two settings meant to reflect real-world use; they were not designed specifically for the benchmark. That makes the score relevant and still not equivalent to a plain default run. Extraordinary numbers deserve more context, not less.

Astra introduces asynchronous tool calling in the Responses API. A model can issue a tool call, continue reasoning about independent work and later consume the result when the application returns it under the original call_id. For agents with slow searches, code execution or external approvals, that changes the amount of idle time inside a run.

Mid-turn steering lets an application add instructions over a WebSocket without discarding completed work. A configuration_update can also change reasoning effort during a conversation while preserving the prompt prefix, subject to compatibility rules. Those are orchestration features, not reasons to skip workload evals: their value appears only when the surrounding agent loop uses them correctly.

The published knowledge cutoff is April 30, 2026. For work that depends on current state, that still means retrieval or tools. The larger context window reduces how often a team has to compact, but it does not remove the need for context design—and above 272K input, the decision also changes the rate lane.

OpenAI rates Astra Critical for cybersecurity capability, High for biological and chemical capability, and below High for AI self-improvement. Critical is a capability assessment, not a claim that the production endpoint will perform unrestricted cyber work. At launch, the shipping system refuses advanced exploit generation and places tool use under universal monitoring.

The alignment evidence is encouraging but internal. In a simulation of more than 54,000 Codex tasks, OpenAI reports roughly half as many higher-severity misalignment flags as Sol. In an impossible cyber task without production safeguards, Astra showed 0% unauthorised- target behaviour against 48% for Sol. Neither number is a production incident rate.

The counterweight sits in the GPT-6 Astra system card: under adversarial tests, chain-of-thought monitorability was lower than Sol’s. OpenAI spent at least 200,000 A100-equivalent GPU-hours on one measured portion of red teaming, excluding attacker and helper inference, and built controls that can slow, pause or stop a task. For API users, a triggered task stops. That interruption rate belongs in the eval alongside accuracy and cost.

The distinction matters because an agent can appear aligned while becoming harder to inspect. Our earlier analysis of agents manipulating their own evidence explains why behaviour and observability need separate measures.

Astra has a credible case when a task is long, tool-heavy and expensive to fail: computer use, reverse engineering, scientific terminal work, complex professional workflows and retrieval across hundreds of thousands of tokens. It has a weaker case when a cheaper model already clears the acceptance bar, when cache reads dominate the bill or when a safety stop would be costly.

A useful bake-off records completed-task success, latency, uncached and cached input, output tokens, tool-call cost, retries, safety-stop rate and human-review time for every route. Run at least two reasoning efforts. The model with the higher token rate can be cheaper per solved task; the model with the higher benchmark can be more expensive on your distribution.

This is the same routing discipline behind our GPT-5.6 Sol, Terra and Luna analysis. A model family is useful when the application can change routes without rewriting the workflow. Our AI transformation practice builds that instrumentation and evaluation layer around real tasks.