Cerebras CS-4: OpenAI's Sol tier now runs up to 14x faster
Three wafer-scale chips hit 30x GPU speed; AMD does prefill, Cerebras the decode. Plus Replit's free Luna tier and Vercel comments hit the terminal.
Copy markdown
OpenAI's Sol Ultrafast, now on Cerebras silicon
OpenAI's GPT-5.6 Sol now has an Ultrafast tier running on Cerebras hardware at up to 14x its standard speed — it churned through 2,500 Humanity's Last Exam questions in ~11 hours versus three-plus days elsewhere. It's limited to select customers for now, but it's the clearest sign fast-inference silicon is going mainstream.
One box, three wafers, 30x a GPU
The CS-4 fuses three Wafer Scale Engine 3 Turbo chips with 43 PB/s of memory bandwidth per wafer and wafer-to-wafer latency as low as 2 microseconds, for up to 30x faster token generation than GPUs and 10x more tokens per watt than the CS-3. General availability is slated for Q3 2026.
AMD does the reading, Cerebras does the writing
Cerebras and AMD split inference: AMD GPUs handle the prefill — chewing through your input context — while Cerebras streams the token-by-token decode, together claimed at 10x GPU speed and 5x Cerebras alone. For latency-bound agent loops, that division of labor is the interesting part.
Elsewhere: Replit's Free Mode runs on OpenAI's Luna
Replit added a Free Mode that routes lightweight work — checkpoints, small edits, Q&A — to OpenAI's low-cost GPT-5.6 Luna without spending your monthly credits on paid plans. Heavier builds still draw credits, so it mainly stretches how far a plan goes.
Also: manage Vercel comments from the terminal
Vercel shipped a `vercel comments` command to list, inspect, reply to, and resolve Vercel Toolbar comments without leaving your terminal — handy if you review preview deployments from the command line instead of the dashboard.