OpenAI today released GPT-6 Astra. Early comparisons included in the release claim Astra outperformed Fable 5.1 on several metrics, achieving 97.6% on the hardest FrontierMath tasks and 99.9% on ARC-AGI-3 in those tests.
Early testers' findings
A team with early access, after burning over 20 billion tokens on Astra, report that the model belongs to a new class of models capable of acting as full AI Engineers. According to their tests, Astra can:
- help select and train models,
- label data and participate in active learning loops,
- keep data pipelines saturated,
- instrument and read logs,
- deploy and debug complete systems in a single run,
- fan out, command and evaluate subagents (including agents that run other models),
- maintain coherence across billions of tokens within a single agent thread.
The testers used Astra on many practical, real-world tasks: building internal/personal tools, redesigning a personal website, creating a partial replacement for GitHub + Vercel, training game AI for a strategy board game with vastly more legal moves than Go, performing personal finance cleanups, republishing a book with synced audiobook and printed editions, and other projects.
Cost and performance: where the $6/hour figure came from
The testers measured approximately 33 tokens per second throughput and assumed a maximum pricing level of $50 per million tokens; under those assumptions they calculated an operating cost near $6 per hour for certain, non-parallelized usage patterns. They also report Astra appeared more token-efficient than Sol and Fable, a finding independently confirmed by Artificial Analysis.
However, the report cautions that running Astra in high-parallelization modes (referred to as Ultra) will considerably increase expense because the model parallelizes tasks very effectively. The testers noted they spent about $100 over two days running experiments.
Benchmarks, demos and quoted reactions
Demos highlighted common workflows: desktop/computer tasks, game play (e.g., Pokémon), Blender workflows, and scientific and cybersafety benchmarks. The announcement also quotes individuals such as Greg ("AGI is here") and Jakub (calling it "finally the Automated AI Research Intern"). The testers themselves say they are not positioning themselves as qualified to settle AGI claims.
Conclusions and scope
The early-access report concludes OpenAI trained a model that can automate many AI-engineering tasks. The testers recommend developers learn how to exploit Astra- and Fable-class models, while remaining mindful that costs can grow quickly with aggressive parallelization.
The article also notes the team is running similar experiments on other frontier models, including Grok and Fable, and that the agentic coding patterns observed are likely to apply across late-2026 frontier models. The testers intend to publish more detailed results later.
Notes:
- The findings summarized here are based on the testers' preliminary experiments and measurements. The official system card and further documentation from OpenAI may provide additional details.
- All names appear in their original form: GPT-6 Astra, OpenAI, Fable 5.1, FrontierMath, ARC-AGI-3, SAM, Sol, Artificial Analysis, Spark 1.3, Grok.
