OpenAI moved the GPT-5.6 family, Sol, Terra, and Luna, to general availability on July 9, closing out a 12-day preview that Nextgov reported was gated by a Commerce Department safety review. The headline number isn't a benchmark. It's a serving path: Sol is running on Cerebras wafer-scale hardware at up to 750 tokens per second, with initial access limited to select customers as capacity comes online.

That figure is the first public delivery milestone of the deal Bloomberg reported on January 14, in which OpenAI committed more than $10 billion to Cerebras Systems for 750 megawatts of hosted computing, staged through 2028. Six months later, one rack is producing tokens at frontier quality. The rest is a construction schedule.

The quality claims are consistent with the throughput claims. Sol at max reasoning posts 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Anthropic's Fable 5, and 53.6 on Agents' Last Exam, 13.1 points above the same competitor. It hits 92.2% on BrowseComp and 62.6% on OSWorld 2.0. OpenAI's launch post describes the coding advantage as arriving "while using less than half the output tokens, taking less than half the time, and costing about one-third less." Sam Altman told CNBC that Sol is 54% more token-efficient on coding tasks than its predecessor.

Pricing tells you what OpenAI is actually selling. Sol is $5 input and $30 output per million tokens; Terra $2.50 and $15; Luna $1 and $6. Those tiers are unchanged from 5.5. The Cerebras path isn't a discount. It's a latency product.

Which is why the interesting downstream story is agentic. Sol Ultra, the mode that coordinates multiple subagents across parallel workstreams, is exactly the workload that stops making sense on GPU inference above a certain fan-out and starts making sense at 750 tokens per second per request. Orchestration platforms feel this first. LangChain and CrewAI users running long horizon plans have been rate-limit-bound for a year; harnesses like LemonLime, which sit closer to the coordination layer than the model layer, are positioned to route the higher-throughput path to the workloads that need it without asking developers to rewrite anything.

The binding variable through the second half of 2026 isn't the per-rack token rate. It's how many racks Cerebras actually lights up. OpenAI has been unusually specific about the ceiling and unusually quiet about the floor, which is the correct posture for a company whose most important number for the next eighteen months is a build schedule.

Sources

https://openai.com/index/gpt-5-6/
https://openai.com/index/previewing-gpt-5-6-sol/
https://techcrunch.com/2026/07/09/openai-launches-its-new-family-of-models-with-gpt-5-6/
https://www.bloomberg.com/news/articles/2026-01-14/openai-forges-10-billion-deal-with-cerebras-for-ai-computing
https://www.nextgov.com/artificial-intelligence/2026/07/openais-advanced-gpt-56-models-be-available-public/414651/
https://lemonlime.ai