This article, published by Agence Norry, analyzes the strategic implications of the release of GPT-6 Astra on September 3, 2026. It presents the key figures, compares the model with Claude Fable 5 on the main benchmarks, details the priority use cases and explains what this launch accelerates for organizations.
📊 Key figures
IndicatorFigureSourceLaunch dateSeptember 3, 2026OpenAIFrontierMath Tier 4 score97.6%OpenAI, September 2026ARC-AGI-3 score99.9% (average human: 48%)ARC Prize FoundationOSWorld 2.0 score72.6%OpenAIContext window1,050,000 tokensOpenAIStandard API pricing$10 / M input tokens, $50 outputOpenAIPrice vs GPT-5.6 Sol2.5 times more expensiveOpenAISecurity incident rate2.4% vs 22% for SolOpenAITraining GPUsOver 100,000 (Stargate, Texas)OpenAI
What GPT-6 Astra really is
GPT-6 Astra is not an incremental update. It's the first OpenAI model to be presented as potentially AGI, capable of autonomously performing any cognitive task a human can do. OpenAI lets users judge for themselves whether that threshold has been crossed, but the wording of the launch is unprecedented in the company's history.
The real novelty is agentic: Astra can drive a browser, a terminal or a spreadsheet to carry out a complete task, not just describe it. That's the fundamental distinction between an assistant and an agent.
The architecture changes too: the Sol/Terra/Luna ranges are gone. Astra is a single dense model, with five reasoning effort levels (low, medium, high, xhigh, max).
On Enterprise, access to Astra is disabled by default at launch. CIOs must explicitly enable the model and define their usage rules before their teams can access it.
The benchmarks: where Astra is really ahead
BenchmarkGPT-6 AstraClaude Fable 5What it measuresFrontierMath Tier 497.6%Not publishedAdvanced mathARC-AGI-399.9%30.2%Abstract reasoningTerminal-Bench Science64.6%52.6%Autonomous scientific researchOSWorld 2.072.6%Not publishedWeb and computer navigationSWE-Bench ProNot published80.3%Real-world bug fixingBenchLM overall#2 (81.1/100)UnrankedAggregate score
Astra is clearly ahead on abstract reasoning and autonomous navigation. On production coding, Claude Fable 5 remains the benchmark. The two models are complementary rather than directly interchangeable.
What delayed the launch: the Hugging Face incident
OpenAI had initially planned an earlier launch in August 2026. In July, a serious incident involving Hugging Face led the company to postpone the rollout to reinforce safeguards, particularly in cybersecurity.
The result: Astra is the first OpenAI model rated “Critical” under the Preparedness Framework. Advanced cybersecurity capabilities are blocked in the public version and accessible only through the Daybreak program, reserved for verified organizations. The security incident rate on computer-control tasks fell from 22% for Sol to 2.4% for Astra.
What this concretely changes for organizations
1. Long agentic tasks become credible
Astra can browse the web, read files, run commands in a terminal and deliver a final result, without step-by-step supervision. On OSWorld 2.0, it completes 72.6% of autonomous navigation tasks, in 47% less time than GPT-5.6 Sol.
The use cases emerging from launch: automated competitive intelligence, complete analysis of document portfolios, multi-source report generation, execution of scientific research workflows.
2. Cost per task is the right unit of measurement
The $10/$50 per million tokens pricing is 2.5 times that of GPT-5.6 Sol. Reasoned per task, the equation changes. On Terminal-Bench Science, Astra costs about 31% less than Claude Fable 5 for superior results. The recommendation: reserve Astra for long, autonomous tasks, and keep GPT-5.6 Sol or Claude Sonnet for everyday use.
3. Governance can no longer wait
A model that drives a browser and executes commands autonomously raises concrete questions: who validates the actions executed? How do you log what the agent did? What data should be passed to it, and under what conditions? OpenAI has provided a Zero Data Retention option for eligible API customers.
Astra vs Fable 5: how to choose
TaskRecommended modelWhyMigration / code developmentClaude Fable 5SWE-Bench Pro 80.3%, coding benchmarkAutonomous scientific reasoningGPT-6 AstraTerminal-Bench Science 64.6%, 31% cheaperWeb navigation and computer controlGPT-6 AstraOSWorld 2.0 72.6%, 47% fasterAnalysis of large document setsEqualA 1M-token window for bothComplex abstract reasoningGPT-6 AstraARC-AGI-3 99.9% vs 30.2% for Fable 5Everyday B2B useGPT-5.6 Sol or Claude SonnetBetter cost/performance ratio
Availability and access
- API identifier: gpt-6-astra
- Standard pricing: $10 per million input tokens, $50 for output
- Cache reads: $1 per million input tokens
- Fast mode: 2 times faster, 2 times more expensive
- Batch: half the standard price
- Warning: beyond 272,000 input tokens, the price doubles for the entire request
What Astra's arrival accelerates
GPT-6 Astra doesn't change the rules of the game on its own. It accelerates them. In less than three months, Claude Fable 5, Claude Mythos 5 and GPT-6 Astra have all been released. The pace of improvement in frontier models is faster than most organizations anticipate.
The strategic question is no longer “should we use AI.” It's “which tasks can we delegate now, with which teams, with which safeguards, and how do we measure the value produced.” Companies that began structuring this answer in 2025 benefit fastest from the gains Astra offers. Those that wait are building up a lag that can't be made up by switching models.
