Your signal. Your price.
Microsoft's custom post-training of its small MAI Code 1 Flash model yielded a 10% higher code acceptance rate than GPT-5.4 Mini and Haiku-4.5. Fine-tuning also increased the model's SweeBench verified score from 72% to 86%.
GPT-6 Astra initially scored 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Soul and trailing Fable 5.1. This prompted Artificial Analysis to release index version 4.2, which prioritized agentic tasks and placed Astra second overall.
Astra achieved a 100 percent score on Exploit Bench across all effort levels. On an internal OpenAI benchmark of recently disclosed vulnerabilities, Astra successfully constructed exploits for 39 percent of cases, compared to GPT-5.6 Soul's 5.5 percent.
Only 3 results for these filters — try broadening your search