Skip to main content
Trending
Corporate earningsAi infrastructureEnergy marketsElection litigationAviation safetyCrypto securityAi governanceEnterprise aiPlatform liabilityAvian influenzaClinical trialsFederal debtPresidential runoffRpg developmentSpace debrisAir defenseAutonomous vehiclesBudget crisisBug bountiesEnergy m&aForeign investmentFranchise restructuringGene therapyPandemic treatyTranslation pathwaysUn peacekeepingWeapons testingWeight-space programmingAi copyright lawArchaeological toxicology

full timeline

OpenAI–GPT-6 Astra

OpenAI's new flagship model scored 99.9% on a major reasoning test using a custom memory harness, but dropped to 62.7% under standard testing conditions. The discrepancy highlights how software scaffolding now heavily influences artificial intelligence performance metrics.

Factlen Newsroom· Updated

Follow OpenAI–GPT-6 Astra in the Factlen app and get an alert when there is news. Get the app

Independent testing reveals that GPT-6 Astra's benchmark performance relies heavily on its surrounding software harness.From: Independent Testers Find 37-Point Gap in GPT-6 Astra's Key Intelligence Benchmark

Share this story

Charts, maps and images

GPT-6 Astra's performance on the ARC-AGI-3 benchmark varies by 37 percentage points depending on the memory harness used.From: Independent Testers Find 37-Point Gap in GPT-6 Astra's Key Intelligence Benchmark ·
Artificial Analysis Intelligence Index scores place GPT-6 Astra closely in line with its predecessor.From: Independent Testers Find 37-Point Gap in GPT-6 Astra's Key Intelligence Benchmark ·
The recurrent depth architecture routes tokens through the same layers multiple times to increase reasoning capacity.From: How GPT-6 Astra Actually Works and What Shipped in OpenAI's Rollout ·
API developers access GPT-6 Astra at a standard rate of $10 per million input tokens.From: How GPT-6 Astra Actually Works and What Shipped in OpenAI's Rollout ·

Timeline