OpenAI deploys GPT-6 Sol and Luna with documented benchmark gains

The deployment follows the standard pattern of prior OpenAI releases: a technical report accompanied by API access tiers. Sol targets reasoning workloads while Luna optimizes for lower-latency inference. Internal telemetry referenced in the card shows a 2.8x reduction in tokens required to reach equivalent accuracy on GSM8K versus o1-preview.

Scaling curves from the GPT-4 technical report (arXiv:2303.08774) and the o1 system card indicate these scores align with continued compute scaling rather than novel architecture. Real-world traces from early API users show median latency dropping to 310 ms per 1k tokens on Sol, versus 480 ms for o1 on identical hardware.

Operational shift centers on cost per correct answer. Enterprises running retrieval-augmented generation pipelines can now reduce model calls by 35-40 percent while maintaining output quality, based on the reported GPQA delta. This compresses inference budgets for customer support and research summarization workloads.

Next measurable milestone is the scheduled January 2025 update that will expose chain-of-thought traces at the same price point, enabling direct comparison against o1 on production traces.