Cerebras says it cut an AI assistant’s dinner-booking task to 22 seconds, compared with several minutes for rival assistants, by combining faster inference with better task instructions and parallel tool use. The more interesting result is that speed did not come from silicon alone. Much of the delay came from agents repeatedly discovering how websites worked.
Cerebras Watch analysis
What happened
In a test published on 24 September, Cerebras used Qwen 3.8 27B on its own hardware with Pi as the agent harness, then compared the setup with major consumer assistants on the same restaurant-booking request. Cerebras says its assistant completed the task in 22 seconds, while Meta Muse took 4 minutes 36 seconds, Claude Cowork 6 minutes 25 seconds and Grok Bot 7 minutes 40 seconds. Each system successfully made a reservation at one of three restaurants.
Cerebras attributes the improvement to several changes working together. Its optimised run checked independent restaurant options in parallel, used saved instructions for the booking process and reduced tool calls by more than 80 per cent. The company says the browser and API portion took 6.8 seconds in its optimised run, compared with 4 minutes 31 seconds in the earlier Grok trace.
That is not a clean apples-to-apples measure of processor speed. The comparison also changes the model, harness, instructions and execution path, and Cerebras explicitly says the figures do not isolate the effect of any one improvement. Its practical point is nevertheless useful: an agent can waste more time deciding how to navigate a website than doing the actual booking.
Why it matters
Consumer AI assistants are increasingly judged by whether they complete chores, not whether they produce a fluent paragraph about them. If every task begins with an agent poking around a site, waiting for a model response and repeating the same discovery work, the promised convenience can become a seven-minute hostage situation involving a restaurant booking.
The test suggests that reusable procedures and parallel work may matter as much as a larger model. Faster inference shortens the pauses between seeing, deciding and acting, but it cannot make a slow website load instantly or remove a phone-verification step.
Our read
Cerebras has published a persuasive engineering demonstration, not an independent benchmark of the whole assistant market. The useful lesson is that agent performance is a systems problem. Better chips help, but so do sensible workflows and remembering what the agent has already learnt.
The next serious test is whether these gains survive unfamiliar websites, changing menus, failed payments and tasks where the agent cannot quietly rely on a pre-written route. Dinner is a friendly benchmark. The internet, as ever, is less house-trained.
What to watch
- Generalisation
Whether saved skills work across new sites and changing layouts. - End-to-end timing
Whether future tests publish comparable model, harness and tool configurations. - Failure recovery
How the assistants behave when availability, payments or verification steps go wrong. - Inference economics
Whether faster responses reduce the cost of agents enough to make routine automation practical. Source: Cerebras’ test and methodology.
Discussion spark: Should AI assistant benchmarks reward the fastest successful result, or should they give equal weight to how much prior task-specific preparation made that result possible?
Sources and evidence
- Why AI Assistants Are Slow-and How to Make Them Faster – Cerebras (24 September 2026, 21:41 UTC)
Independent WittyWires tracker for public updates about Cerebras. Not affiliated with or endorsed by Cerebras; this is not an official account.