OpenAI launches GPT-6 Astra focused on computer use and coding

OpenAI announced GPT-6 Astra, a model focused on computer use, coding, professional workflows, science, and cybersecurity. According to InfoQ, it starts out available to a limited set of organizations and is rolling out to ChatGPT Plus, Pro, Business, and Enterprise, as well as the OpenAI API, Microsoft Azure, and AWS Bedrock.

The central message isn't about generating better text: Astra is positioned to execute multi-step tasks directly inside software. OpenAI says the model interacts with graphical interfaces to fill out forms, update CRM records, do research, build websites, analyze data, install and test software, and diagnose problems visible on screen. It's agent logic moving out of the experiment and into what the company sells as a product.

The numbers OpenAI put on the table

The source compiles the benchmarks released by OpenAI itself. It's worth reading with the usual caveat: these are results reported by the vendor, not independent measurements.

| Evaluation | GPT-6 Astra | Comparison | |---|---|---| | OSWorld 2.0 (computer use) | 72.6% | 65.7% (GPT-5.6 Sol) | | Terminal-Bench 4.0 (code) | 57.9% | — | | DeepSWE v1.1 (code) | 74.1% | — | | MRCR 512K-1M tokens (long context) | 96.3% | — | | Hallucination rate (internal) | 4.2% | 12.2% (GPT-5.6 Sol) |

The jump in OSWorld 2.0 and the sharp drop in the internal hallucination rate are the figures that matter most to anyone considering delegating real tasks to an agent. Less hallucination, in theory, means less review rework on what the agent produced, although "internal evaluation" is exactly the kind of metric that calls for third-party validation before it becomes an architecture decision.

Persistent memory in Codex is the detail for developers

For developers who already use agents in long-running tasks, the most concrete news is in Codex. Astra introduces an experimental context mechanism that lets the agent keep notes across context windows, instead of relying solely on compaction.

In practice, as described by OpenAI: previous context windows remain searchable, and the model can retrieve requirements, test results, and tool outputs from past steps during a long-running coding task. Anyone who has watched an agent "forget" a spec agreed on 40 minutes earlier, or reintroduce a bug it had just fixed, understands the problem this tries to address. It's the difference between an assistant that loses track of things and one that keeps the history of the work session.

The model also supports context of up to 1 million tokens, with a result of 96.3% in the 512K to 1M range on the MRCR evaluations. A large context matters less as a marketing number and more for what it enables: dropping an entire repository, long logs, or extensive documentation into a single session without chopping everything into pieces.

The critical point: cybersecurity

Here's the part that changes the tone of the launch. Astra is the first OpenAI model classified at the critical capability level for cybersecurity within the company's Preparedness Framework.

In tests without production safeguards, OpenAI states that the model discovered and used two previously unknown vulnerabilities and demonstrated the ability to develop exploits against hardened browsers and operating systems. The production version restricts advanced offensive tasks, and the company says it will offer broader defensive capabilities through the Daybreak program.

For those working in security in Brazil, this cuts both ways. On one hand, a tool with this kind of defensive capability can speed up the discovery of flaws in in-house code. On the other, the existence of a commercial model that, unrestricted, finds zero-days and writes exploits changes the threat baseline, even if the public version ships limited.

The counterpoint OpenAI itself admits

Not everything is a gain. OpenAI acknowledges that Astra's written reasoning is harder to monitor than its predecessor's, in tests designed to measure whether a model can obscure its own reasoning. The company classifies improving this monitorability as an active area of research.

In practice, this means a meaningful trade-off for anyone who needs to audit what the agent did and why: less hallucination, but a less transparent reasoning chain. In regulated sectors, such as fintechs and healthcare, traceability of the model's decision isn't a luxury, it's a requirement.

How it stacks up in the competition

Astra competes with Claude Fable 5.1, from Anthropic, and Gemini 3.8 Flash, from Google. Based on the evaluations published by OpenAI, the scoreboard is uneven depending on the task:

• Astra leads on Terminal-Bench 4.0 and on several computer-use evaluations;
• Claude Fable 5.1 scores higher on Humanity's Last Exam;
• Gemini 3.8 Flash supports native video and audio input, something Astra doesn't do.

In other words: there's no single winner. For a pipeline centered on an agent that operates a terminal and browser, Astra's numbers stand out. For heavy multimodality, Gemini remains ahead.

On the infrastructure side, Nvidia CEO Jensen Huang commented on the model's training:

GPT-6 Astra, trained on ~100k+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations, OpenAI team. 400k GPUs coming online next.

>

-- Jensen Huang, CEO of Nvidia

The use of the term AGI also appears in other community reactions cited by InfoQ. Here's where the editorial filter matters: "AGI has arrived" is the framing of someone who sells GPUs and someone celebrating the launch, not a consensual technical classification. For those building software, what matters isn't the label, but the behavior in production.

What changes for those running agents in Brazil

The practical point for the Brazilian developer lies in the combination of availability and cost. Astra arrives via OpenAI API, Azure, and AWS Bedrock, the three channels most teams here already use, which reduces adoption friction: you can plug it in without switching cloud providers.

What remains open and deserves attention before moving real workloads:

• Pricing wasn't detailed in the source. A 1M token context and multi-step tasks tend to consume a lot of tokens, and cost per execution is what decides whether agent automation pencils out or not;
• Benchmarks come from the vendor. It's worth waiting for independent measurements before assuming the 72.6% on OSWorld as guaranteed for your use case;
• Codex's persistent memory is experimental, so it's not something to bet a critical workflow on yet.

The sensible way to test without surprises: run Astra on an agent task you already perform today with another model, measure cost per execution and accuracy rate in your context, and compare it against your current baseline before any migration. The announcement is big; the decision is still an engineering one.