On 03.09.2026, OpenAI introduced a new model called Astra, which the company presented as a generational advancement toward AGI.

Capabilities highlighted

OpenAI positions Astra as state-of-the-art across browsing, coding, cybersecurity, scientific research and professional tasks, with 1 050 000 tokens context and outputs up to 128K tokens.

The company reported benchmark results including 100% on ExploitBench, 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, and 96% on GPQA Diamond.

Security profile and release rationale

OpenAI states Astra attained the highest internal cybersecurity classification, able to discover unknown vulnerabilities and craft new exploitation techniques given tooling and access.

According to the company, these capabilities were a central reason for delaying public release until mitigations and guardrails were validated.

Benchmarks, comparisons and reception

By the Artificial Analysis Intelligence Index Astra scored 61.2, compared to 60.9 for GPT-5.6 Sol, a model released less than two months earlier.

OpenAI’s ARC Prize benchmark result that produced Astra’s 98.6% figure does not equate to an official AGI declaration, and independent broad metrics still show competitors ahead on some evaluations.

Pricing and early examples

OpenAI set API pricing at $10 per million input tokens and $50 per million output tokens for Astra access, according to the announcement.

Community members published examples on the day of release, including a multiplayer game demonstrating Astra’s interactive capacities developed within hours of availability.

Context from leadership

«AGI is not a particularly useful term»

Earlier this year, Sam Altman described the AGI concept as limited; in July he said publicly that "we're kind of already in the singularity." The new release follows those comments and company briefings framing Astra as a notable inflection point.