1300 166 138 Tell us what you need
News / OpenAI

OpenAI previews Ultrafast GPT 5.6 Sol at 14 times the speed

By Monika Tantau • Published 3 months ago
OpenAI previews Ultrafast GPT 5.6 Sol at 14 times the speed

Key takeaways

• Ultrafast runs GPT 5.6 Sol up to 14 times faster than the standard API tier
• Output reaches 750 tokens per second on Cerebras wafer scale hardware
• Each Cerebras chip holds 44 GB of memory so model weights stay on the silicon
• Limited preview opened 13 August with a waitlist and no pricing announced

Most Read

OpenAI is previewing Ultrafast, a new service tier for its API that runs GPT 5.6 Sol up to 14 times faster than standard processing and generates up to 750 output tokens per second, according to OpenAI. Access opened on 13 August as a limited preview for a select group of customers, with a waitlist for everyone else.

No pricing has been announced and there is no date for general availability. The announcement still matters, because it shows where the model race heads once raw capability stops being the only axis worth competing on. Speed is becoming a product in its own right.

What OpenAI actually announced

Ultrafast applies to one model only, GPT 5.6 Sol, and sits alongside the existing standard tier rather than replacing it. OpenAI describes the tier as built for time sensitive work and names coding, commerce, financial research and customer support among the first use cases being tested with an initial group of companies. The framing is telling: as models get larger and reasoning chains get longer, keeping interactive applications feeling interactive becomes its own engineering problem.

The company is taking sign ups from businesses that want to be notified as access expands. One caveat appears in OpenAI’s own framing: for incident response work, engineers remain responsible for judgment and deployment. A faster answer does not shift accountability from the person who acts on it.

Cerebras supplies the hardware

The speed comes from outside OpenAI’s usual stack. Cerebras confirmed the same day that Ultrafast runs on its wafer scale processors, chips the size of an entire silicon wafer holding 44 GB of memory on the chip itself. Model weights stay on the silicon, which removes the memory bandwidth bottleneck that limits GPU based serving, the company explained.

Cerebras also published its own comparisons: a 5.6 times end to end speedup on the GDP Val benchmark with no quality loss, 2,500 Humanity’s Last Exam questions finished in 11 hours 11 minutes against more than 78 hours for Claude Fable 5, a claim of 11 times faster output than Anthropic’s flagship, and five times faster than Claude Opus 4.8 running in its Fast mode. Treat those as vendor numbers rather than independent benchmarks until third parties measure the tier.

Why Australian businesses should care

Speed changes which automations are practical. A support agent that streams an answer at 750 tokens per second feels instant to a customer on chat or voice, where the standard tier can leave visible pauses. Long report generation and batch document processing shrink from background jobs into interactive ones. That matters most in the sort of conversational AI work where response time decides whether people keep using the tool at all.

The missing number is price. Sol is already OpenAI’s premium model, and a premium speed tier on top of it will not be cheap, so there is no business case to build yet. The sensible move now is to join the waitlist, measure where latency actually costs you money, and check how Sol compares with its cheaper siblings before assuming the fast lane is worth paying for.

Why Australian businesses should care

Speed becomes the next battleground

The preview landed the same day Google launched Gemini 3.7 Flash, its cut price workhorse for agents. One vendor is competing on cost per token, the other on tokens per second, and both are aiming at the same production workloads. Specialised inference hardware has been a niche story for years. Watching OpenAI put its premium model on wafer scale silicon suggests the niche just went mainstream.

Scroll to Top