Celeris-1-decision: 2x faster than Jev
Celeris released its first decision model on the 8th of October 2026. By its own tests it answers in 67 milliseconds, and a million typical questions cost about 8 dollars.
Celeris launches Celeris-1-decisionCeleris says its new Celeris-1-decision answers a typed question in 67 milliseconds, about half the time of TypeSafe AI's Jev. It was released on the 8th of October 2026 in a post by Celeris chief executive Tom Hamer.
A decision model answers with yes or no, a pick from a list, or a score, plus how sure it is. Celeris-1-decision reads text, structured data or images, the post says.
Below we break down Celeris's "60% faster than Jev" with its own timings
None of the post's timings gives the 60% in its headline, by our arithmetic.
For one question, Celeris timed itself at 67 milliseconds and Jev at 135, from Amazon's US East region. That is 50% less time, or twice as fast.
For 50 questions at once, the figures were 81 and 153 milliseconds, or 47% less time. Network travel is included, Celeris wrote.
On the jev-bench test, Celeris scored itself 0.800 for accuracy, against 0.790 for Perplexity's decider and 0.731 for Jev.
Prices: about 8 dollars per million answers, against 20 at OpenAI
Celeris charges 0.04 dollars per million input tokens, the pieces of text a model reads, its price page says. Answers are free.
A million questions of 200 tokens each then cost 8 dollars, by our arithmetic. Jev would charge 8.40 dollars and OpenAI's gpt-6-luna 20.
Celeris lists Perplexity at 0.04 dollars too. Perplexity's own price page says 0.02, half the price of Celeris.
What is confirmed: every score comes from Celeris's own tests
BenchLM, a site that tracks model results, wrote that it has not rerun Celeris's results. No independent test was published by the 12th of October.
Celeris called its 3,210-item test "full jev-bench". The jev-bench page on Hugging Face lists 22,773 test records, and Celeris did not say how it chose.
How to try it: a waitlist for new accounts and prepaid credit
Celeris activates a limited number of new workspaces each day, its docs say. Credit is paid in advance, expires after 30 days, and there is no free tier.
The model runs only on Celeris's online service. Code written for Jev works after a change of address and key, the docs say.
Key facts
Released by Celeris on the 8th of October 2026.
67 milliseconds per question against 135 for TypeSafe AI's Jev, in Celeris's own tests.
0.04 dollars per million input tokens, answers free.
Source: Celeris blog post by Tom Hamer, the 8th of October 2026