OpenAI confirmed it will deploy GPT-5.6 Sol on Cerebras wafer-scale hardware this July, targeting up to 750 tokens per second for select customers. For context, standard GPU serving of a frontier model runs around 50 tokens a second. This is roughly 15 times faster. That is not a spec-sheet brag — it is a different product.
Why wafer-scale wins here
Cerebras puts entire model layers on a single wafer instead of splitting them across a rack of GPUs stitched together with networking. The bottleneck on standard clusters is not raw compute, it is the constant memory shuffle between chips. Kill that shuffle and generation speed jumps. That is the whole trick, and it is why the number is 15x and not 15%.
The tiers
GPT-5.6 comes in three sizes: Sol at $5 in / $30 out, Terra at $2.50 / $15, and Luna at $1 / $6. Sol is the flagship, and it is the one getting the Cerebras speed treatment first.
The AI tool stack actually worth paying for
One email a week. The models, tools and moves that matter, stripped of hype and filtered so you don’t have to drink from the firehose. Free, and you can bail anytime.
What 750 tokens a second actually buys
Three things stop being painful. Voice apps where the model answers before the silence turns awkward. Coding agents that iterate at human pace instead of making you watch a spinner. And multi-step agent chains where latency, not cost, was the thing killing the experience. If you have shelved a real-time idea because the model was too slow to feel alive, this is the unlock.
The catch
“Select customers.” This is a gated July deployment, not general availability, and there is no public date for when the rest of us get in. So admire the number, sketch the product, but do not rewrite your roadmap around throughput you cannot buy yet. Speed this good has a way of arriving late and metered.