OpenAI ships GPT-6.1 Sol with 80% lower inference cost than Astra baseline

OpenAI posted the GPT-6.1 Sol weights and API endpoints on 12 October. Internal evaluation logs list 1.2 trillion active parameters with a 128k context window. The release notes cite a 4.8x throughput increase on H100 clusters versus the Astra production stack.

Benchmark tables show GPT-6.1 Sol trailing Astra by 2.3 points on MMLU and 1.9 points on HumanEval while recording 19.8% of Astra's measured FLOPs per token. Latency traces from the public API endpoint average 42 ms per token at batch size 1.

The cost reduction stems from a distilled mixture-of-experts router that activates 12% of parameters per forward pass. Production logs from early partners indicate 3.1x higher tokens processed per dollar on financial document extraction workloads compared with the prior model tier.

Deployment data through week three shows 14 enterprise tenants running sustained loads above 10 million tokens per day. No public roadmap for further price tiers has been issued.