OpenAI launches GPT-6 Astra days after Claude Fable 5.1 in escalating model race

OpenAI has released GPT-6 Astra, its newest frontier model, less than a week after Anthropic shipped Claude Fable 5.1. According to a dev.to writeup on the launch, OpenAI presents Astra as its most intelligent and best-aligned model to date, with the pitch centred on completing work rather than describing how to do it.

A model meant to operate the machine

The headline capability is end-to-end computer use. According to the writeup, Astra can fill out forms, update CRM records, run frontend QA checks, and troubleshoot software by watching what happens on screen, all without step-by-step hand-holding. OpenAI's self-reported numbers put it at 72.6% on OSWorld 2.0, narrowly ahead of Claude Opus 5 at 70.2% and its own predecessor GPT-5.6 Sol at 65.7%. One example from the launch material has the model extracting figures from a W-2 and entering them into a Form 1040, automating routine tax preparation while leaving the final return for human review.

A related change is better judgment about when to interrupt. In a side-by-side demo, GPT-5.6 Sol built a personal career website unassisted in roughly 13 minutes, while Astra paused after 20 seconds to ask what career the user was actually moving into. The point of the demo is a model that weighs whether acting autonomously risks the wrong outcome.

Persistent notes for long coding sessions

In Codex, Astra can keep searchable notes across context windows instead of repeatedly compressing long debugging sessions into a single summary, preserving details that compaction tends to lose, such as why an earlier fix failed. The feature is opt-in through Codex's config file for now, with OpenAI saying it will become the default in the coming weeks.

Cybersecurity capability reaches OpenAI's highest risk tier

The sharpest change is not a productivity feature. OpenAI reports that Astra reaches the Critical threshold for cybersecurity under its own Preparedness Framework, its highest risk tier, meaning the model can independently identify and develop working exploits for previously unknown vulnerabilities. The company reports a perfect score on ExploitBench and an 88% first-attempt success rate on SRE-Bench reverse-engineering tasks. Because of that rating, the riskier parts are gated at launch: Astra will help with defensive work such as secure code review and patch validation, but it refuses to create proof-of-concept exploits until OpenAI expands access through its Daybreak program.

The benchmark picture is uneven

All figures come from OpenAI's own launch materials, and they do not add up to a clean sweep. Astra posts 57.7% on Terminal-Bench 4.0, ahead of Claude Fable 5.1's 55.8% but only narrowly, and 96.0% on GPQA Diamond, just above Gemini 3.8 Flash's 95.3%. On FrontierMath Tier 4 it reports 97.6%, up sharply from GPT-5.6 Sol's 83.0%. Yet on Humanity's Last Exam with tools it scores 57.2%, behind both Claude Fable 5.1 at 65.0% and Claude Opus 5 at 63.6%.

One headline number also carries an asterisk. Astra's marketed 99.9% on ARC-AGI-3 depends on an expensive, stateful evaluation harness, and the writeup reports that independent testing by the ARC Prize Foundation found a standard stateless API call scores somewhere between 17% and 63% depending on the reasoning tier used.

Pricing and availability

Astra is rolling out in stages: limited organizations first, then ChatGPT Plus, Pro, Business and Enterprise users over the following days. Developers can access it as gpt-6-astra through the OpenAI API, Microsoft Azure and Amazon Bedrock, at $10 per million input tokens and $50 per million output tokens, with a fast mode offering roughly 2.5x the speed at 2x the price. That sits well above GPT-5.6 Terra's $2/$12 rates and Claude Opus 5's $5/$25, a signal, the writeup suggests, that the model targets high-value autonomous work rather than bulk text generation. Zero data retention is available for eligible API customers. Early comparisons from users with access, including a Blender villa scene posted by developer Karan Kendre, have generally favoured Astra, though neither model is broadly available yet.

Why it matters

Two things stand out. First, the frontier-model race is compressing: Astra landed within a week of Anthropic's release, and competition is now measured in agentic reliability — computer use, session memory, and knowing when to stop and ask — rather than chat quality. Second, Astra's cybersecurity rating shows capability outpacing deployment policy: OpenAI has built a model it classifies at its own highest risk tier and is shipping it anyway, with the dangerous portions fenced off behind a separate access program. Between contested evaluation setups and pricing aimed at autonomous workloads, the real verdict will come from whether organisations can hand Astra messy, multi-step tasks and trust the result — a test no leaderboard captures.