GPT-5.6 Sol runs 10x faster on Cerebras. Speed is a dial.
OpenAI is serving GPT-5.6 Sol at up to 750 tokens/sec on Cerebras wafer-scale chips. Inference speed is now decoupled from the model — treat it as a choice.
Same weights, ten times the speed. As GPT-5.6 goes generally available, OpenAI says it will serve the flagship Sol model on Cerebras' wafer-scale hardware at up to 750 tokens per second — where a frontier model on a typical GPU cluster streams somewhere in the 40–120 range. Nothing about the model changed. The chip underneath it did. For anyone deploying an AI agent a customer actually waits on, that's the number that matters more than the benchmark.
What actually happened
The context is a deal that's been building all year. In January, OpenAI and Cerebras signed a multi-year agreement to deploy 750 megawatts of Cerebras wafer-scale systems to serve OpenAI customers, rolling out in stages through 2026, per Cerebras' own announcement. Cerebras claims its systems deliver responses "up to 15× faster than GPU-based systems," and its published record speeds back the order of magnitude up: gpt-oss-120B has run at roughly 3,000 tokens/sec on its hardware, and it's shown models like Gemma 4 at over 1,800.
Now that capacity is pointed at GPT-5.6 Sol. The takeaway isn't wafer-scale trivia — it's that inference speed has become a property of where you run the model, not the model itself. The same Sol weights are slow on one substrate and blistering on another. Speed is no longer something you accept; it's something you buy.
Why it matters for your business
Here's the operator's angle. For a batch job — reconciling a month of transactions overnight — tokens per second barely registers. For anything a human sits and waits on, it's the entire experience: a support agent that answers in half a second feels like a product, and one that spins for fifteen feels broken. Latency is not a technical detail. It's the difference between customers using the thing and abandoning it.
The trap is baking one provider's latency into your app as if it were fixed. It isn't. The same logic we push on owning your inference layer applies to speed: route your model calls through a thin abstraction you control, so you can send latency-sensitive traffic to the fastest (and usually priciest) endpoint and dump the background batch work on whatever's cheapest — without rewriting your application. When a wafer-scale endpoint makes your agent feel instant, you want to flip that switch in an afternoon, not a migration. Rent the speed. Own the switch.
Key takeaways
- OpenAI says GPT-5.6 Sol will run at up to 750 tokens/sec on Cerebras wafer-scale chips — roughly an order of magnitude faster than a frontier model on typical GPUs
- The OpenAI–Cerebras deal commits 750 megawatts of wafer-scale capacity, rolling out through 2026; Cerebras claims up to 15× faster responses than GPU systems
- Inference speed is now decoupled from the model — the same weights are slow or fast depending on the hardware you pick
- For customer-facing agents, latency is the product; route model calls through an abstraction you own so speed and cost are dials, not migrations
Building an AI agent your customers wait on? We architect the inference layer so latency-sensitive calls hit the fastest endpoint and batch work hits the cheapest — swappable without touching your app. See how we build systems you own or tell us where it's slow.
Sources: Cerebras — OpenAI partners with Cerebras to bring high-speed inference to the mainstream.
- #ai-automation
- #inference
- #cerebras
- #latency
- #ai-agents
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
ServiceTrade bought an agentic billing agent. Own your last mile.
ServiceTrade acquired Mura to automate field-service order-to-cash with agentic AI. Agentic billing has reached the boring, high-value core of service ops.
Read itPrime Intellect raised $130M so you can train your own agents
Prime Intellect hit a $1B valuation selling infrastructure to train AI agents on your own data instead of renting a frontier lab. Own the optimization loop.
Read it