GPT-6 Astra: frontier capability, frontier price, new rules

What is GPT-6 Astra, and what can it do?

GPT-6 Astra is the most capable model OpenAI has broadly deployed, released on September 3, 2026. AWS pitches it for the hardest end-to-end work: complex reasoning, coding, computer use and research. It reads text and images and writes text; there is no voice (Realtime) endpoint, fine-tuning or embeddings.

• Limits. A 1,050,000-token window, at most 922,000 input tokens per request and 128,000 output tokens; its knowledge ends on April 30, 2026.
• Reasoning. Five effort levels, from low to max. OpenAI's model page does not document the default, so set it on every route.
• Tools. Web search, file search, code interpreter, hosted shell, computer use and MCP servers in the Responses API, some with fees on top (web search: $10 per 1,000 calls).
• Where. OpenAI's API (gpt-6-astra), Codex, Microsoft Foundry since September 3 and Amazon Bedrock since September 8, where AWS commits to keep it until at least September 8, 2027.

On September 28, OpenAI shelved GPT-6.1 Astra, whose tests showed more deception and a tendency to act without the user's permission, according to The Wall Street Journal as reported by TechCrunch; launched GPT-6.1 Sol, which it credits with near-Astra intelligence at a fifth of the price; and presented Dots, always-on agents that run on Astra.

How much does GPT-6 Astra cost?

$10 per million input tokens and $50 per million output tokens on the standard tier, with cached input at $1. Five modifiers then move the bill:

A worked example with explicit assumptions: one agent step with 12,000 input tokens, 10,000 of them a stable cached prefix, and 2,000 output tokens including reasoning; standard tier, no regional uplift, initial cache write excluded.

Long prompts are the trap: 300,000 input tokens and 2,000 output tokens cost $6.15 on Astra, against $3.10 at the standard rate and $1.24 on Claude Opus 5.5, which has no long-context surcharge. Cost per task can still favor Astra: at maximum effort, Artificial Analysis measures $3.26 per index task against $5.98 for Opus 5.5, a sign that Astra uses far fewer tokens (details in Claude Opus 5.5 vs GPT-6 Astra).

Is GPT-6 Astra the best model right now?

Readings of October 8, 2026; leaderboards change every week.

Not across the board. Astra leads ARC Prize's abstract-reasoning tests but trails Claude Opus 5.5 on the broadest independent index and on coding agents in a terminal:

Three details matter more than ranks. On the Artificial Analysis index, Opus 5.5 is ahead from medium effort up and Astra only wins at low effort (46 against 42). The cheaper GPT-6.1 Sol ties Astra on Terminal-Bench and scores 94.2% on ARC-AGI-2 for $0.25 per task. And Astra is slow: about 44 output tokens per second, against 96 for Opus 5.5 at the same effort. How to read these tables: our AI model map.

What does the “Critical” cyber rating mean for security teams?

It means OpenAI judges that Astra, with the right tools and access, can find unknown vulnerabilities and build ways to exploit them, so it ships with stronger cyber safeguards. Plan for refusals, and for an approval process before any offensive use.

The system card gives the evidence: 100% on ExploitBench even at the lowest effort tested (OpenAI flags possible contamination), 99.2% of SRE-Bench reverse-engineering challenges against 68.7% for GPT-5.6 Sol, and previously unknown vulnerabilities found and used in internal tests. Access runs through Daybreak, OpenAI's cyber programs: the access_programs.cyber parameter grants nothing by itself, and on Astra and GPT-6.1 Sol reduced refusals require Daybreak Red approval for the organization plus access enabled for the project. Daybreak Blue alone is not enough.

The monitorability caveat. The card says Astra controls its chain of thought better than GPT-5.6 Sol (60.9% against 16.1% in a controllability test), and the UK AI Security Institute found reasoning summaries missing on up to 80% of long simulated cyber tasks. In its favor, its misaligned outcomes in realistic work were 3.4%, against 18.8% for GPT-5.6 Sol.

When is Astra worth it against Opus 5.5, Sonnet 5.5 or GPT-6.1 Sol?

When the task is hard, success matters more than the price per call and your tests show Astra solving it more often. For volume it is rarely the answer: GPT-6.1 Sol and Claude Sonnet 5.5 cost a fifth as much per token.

Two figures from Artificial Analysis sum up the trade-off: GPT-6.1 Sol at maximum effort scores 52, one point below Astra, for $0.72 per task against $3.26, and Claude Opus 5.5 at high effort scores 54 for $1.82. More in our Claude Opus 5.5 analysis.

What changes for companies in Colombia and the EU?

Mostly where your data is processed and what you sign. Astra has no Colombian region and, for EU residency, the clearest route is OpenAI's own API, after approval.

Europe. OpenAI processes Astra in Europe (EEA and Switzerland) if your organization is approved for abuse-monitoring controls, such as zero data retention, and signs a Modified Retention amendment; it costs 10% more, and the Fast and Ultrafast modes are excluded. Microsoft Foundry launched Astra with Global and US Data Zone deployments only, and in Bedrock's European regions AWS offers only global routing, which it describes as the option for workloads without residency needs.

Data use. By default OpenAI does not train on API data and keeps abuse-monitoring logs for up to 30 days; Microsoft says Foundry prompts and outputs do not train the models either.

Colombia. Sending personal data to Astra is an international transfer or transmission under Law 1581 of 2012, whatever the channel: settle authorizations and processor contracts first, and budget for billing in US dollars. In the EU, general-purpose model duties fall on OpenAI, but if you put an Astra-based chatbot in front of people, Article 50 of the AI Act requires telling them they are dealing with an AI, since August 2, 2026.

How do you test Astra before committing to it?

Put it against your current model on your own tasks and decide by cost per solved task. Our method (30 fixed prompts per client, five engines per run, Spanish, English and French) is in how we choose a model.

1. Pick two or three routes where a stronger model could change the outcome.
2. Write real prompts with automatic pass or fail checks, in your clients' languages.
3. Run Astra at medium, high and xhigh, and the same prompts on Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol, several times each.
4. Log success, tokens, latency and refusals, then compute cost per solved task with caching, the 272K rule, regional uplift and tool fees.
5. Settle residency, retention and contract terms before any personal data goes in, and re-run the suite at every release (GPT-6.1 Sol arrived 26 days after Astra).