OpenAI's GPT-6 Astra raises the bar on AI capabilities—but claims of AGI require careful scrutiny

OpenAI unveiled GPT-6 Astra on Thursday, positioning it as a watershed moment for artificial intelligence. President Greg Brockman declared the industry now lives in the "AGI era," suggesting future observers might pinpoint Astra as the model marking that transition—though he stopped short of a formal AGI declaration, describing the term instead as a "mission concept or spiritual concept."

The model does deliver measurable improvements across several benchmarks. Yet many of those gains come wrapped in important caveats that complicate the narrative of unchecked progress.

Training scale and architecture shifts#

Astra represents OpenAI's largest training run yet. According to Aidan Clark, the company conducted the pre-training process using more than 100,000 GPUs at its Stargate facility in Texas. Notably, earlier OpenAI models played a meaningful role supervising this training—a structural difference from previous releases.

The rollout begins with enterprise customers enrolled in OpenAI's Daybreak program, expanding to Plus, Pro, Business, and Enterprise subscribers as well as API users and AWS integrations "in the coming days." Standard API pricing will run $10 per million input tokens and $50 per million output tokens—2.5 times higher than the current promotional rate for earlier models but aligned with Anthropic's pricing for Fable 5.1.

Benchmark gains—with context#

Astra's performance highlights vary considerably in their weight as evidence of breakthrough capability.

Coding work shows meaningful but contested improvement. The model scored 74.1% on DeepSWE v1.1, a 113-task agentic coding benchmark, compared with 70.8% for the prior model. However, Meta's Muse Spark 1.3 reported 75.4% at maximum reasoning settings—a slim nominal advantage, though safety review has kept that configuration unavailable to general users. The reported uncertainty ranges across leading models overlap significantly, leaving no clear winner on this measure.

Outside coding, larger gaps emerge. Astra achieved 98.6% on ARC-AGI-3, though that score reflects both the model and OpenAI's agent system combined through specialized reasoning harnesses and context management—not the underlying model in isolation. On FrontierMath Tier 4 (97.6%) and BenchCAD's Vision2Code subset (95.9%), Astra posts commanding leads, though the FrontierMath test involved OpenAI-funded development with exclusive access arrangements worth noting.

On Terminal-Bench Science, measuring 70 command-line research tasks, Astra reached 64.6%—well above Fable 5.1's 52.6% and the public leaderboard's 30% ceiling. Both OpenAI and Anthropic are now pushing models into active research workflows rather than mere question-answering.