OpenAI Launches GPT-6 Astra with Near-Perfect Benchmark Scores

OpenAI claims its newly released Astra model, referred to as GPT-6, is the product of years of research and heavy computational investment. [3] The system achieved a 98% score on the FrontierMath Tier 4 benchmark. [3] Additionally, the model saturated the ARC-AGI 3 benchmark with a 99.9% accuracy rate and scored 100% on ExploitBench. [3] Despite these high scores, independent experts noted that the launch messaging serves marketing goals alongside real progress. [3]

The Astra model currently powers the GPT-6 Pro offering in normal chat for users on the Pro, Business, and Enterprise tiers. [1] In contrast, standard Plus users currently only access Astra through the Work and Codex features. [1] A community member noted that while one announcement suggested Astra would be available to all Plus, Pro, Business, and Enterprise users, current documentation shows Plus users only receive access through Work and Codex rather than the standard Chat interface. [1]

OpenAI is pushing the limits of benchmark testing with Astra, specifically targeting advanced reasoning and exploit discovery. By limiting full chat access to premium tiers, the company is likely prioritizing high-revenue enterprise customers for its most computationally expensive models. The selective availability implies a tiered rollout strategy for bleeding-edge capabilities, diverging from the broader access models seen in earlier GPT-4 iterations. What the sources don't address: Whether OpenAI intends to eventually bring the fully powered GPT-6 Pro chat experience to standard Plus subscribers.