OpenAI's GPT-6 Astra Reportedly Automates AI Engineering Tasks at Under $6 an Hour, According to Latent Space Testing

What happened

AI-focused blog Latent Space published a report describing a model it calls "GPT-6 Astra," which it says OpenAI released and which the blog tested extensively—burning what it describes as more than 20 billion tokens across a range of practical tasks. According to the report, Astra can perform work the author compares to hiring a junior AI engineer: selecting and training models, labeling data, monitoring pipelines, reading logs, deploying and debugging systems, and coordinating fleets of subagents (including agents running other models) over what the blog says are threads spanning billions of tokens.

It's important to note this article is based entirely on a single third-party blog's early-access testing. There is no independent OpenAI announcement text included in the source material, and no system card details beyond a reference to one existing. Benchmark numbers, pricing, and capability claims below should be read as reported by Latent Space, not independently verified by us.

The numbers, as reported

Latent Space claims Astra achieves 97.6% on the hardest version of FrontierMath and 99.9% on ARC-AGI-3, which the blog describes as "completely saturating" both benchmarks. The blog also claims Astra beats a competing model it calls "Fable 5.1" on unspecified metrics, and that Artificial Analysis independently confirmed Astra is more token-efficient than models it refers to as "Sol" and "Fable." None of these comparison models appear in confirmed OpenAI or third-party model registries at this time, and the naming suggests these may be code names or placeholders rather than confirmed product names.

On pricing, the blog states it observed throughput of 33 tokens per second at a maximum rate of $50 per million tokens, and extrapolates a real-world cost of under $6 per hour for typical engineering workloads. The report does not break out separate input and output token pricing, nor does it specify context window size. Both figures should be treated as pricing not yet disclosed in verifiable form.

What OpenAI has confirmed

The source material contains no direct OpenAI statement, press release, or pricing page. Quotes attributed to OpenAI figures Greg (likely Greg Brockman) and Jakub (likely Jakub Pachocki) are secondhand, with Greg reportedly saying "AGI is here" and Jakub reportedly calling Astra the "Automated AI Research Intern" he wanted. Neither quote is sourced to an official transcript in the material provided.

What this means

If accurate, a model capable of autonomously running engineering workflows—data labeling, subagent orchestration, deployment, debugging—at single-digit hourly cost would represent a meaningful drop in the cost of AI-assisted software and research work. But every figure here originates from one blog's self-reported, unaudited testing, not from OpenAI's own materials, a published system card, or a benchmark leaderboard we can verify independently. Readers should treat the FrontierMath and ARC-AGI-3 scores, the pricing model, and the capability claims as unconfirmed until OpenAI publishes official documentation or a system card that independent evaluators can check against. We will update this article once verifiable pricing, context window specifications, and benchmark methodology become available.