Introducing Celeris-1 Magnus

The Celeris team · August 31, 2026

Today we're releasing Celeris-1 Magnus, our second model and our most capable. Magnus is a larger hybrid diffusion model with native reasoning, built for multi-step tool use, code that has to run, and tasks where the model must go find things out. It is available on the API now.

The headline: τ³-Banking

τ³-Banking is the hardest benchmark we run: a simulated customer-service conversation at a bank, with the policies hidden in a knowledge base behind retrieval tools, and credit only for finding and executing the exact action sequence each task requires. We ran Magnus against the newest agents we could get API access to - OpenAI's gpt-5.6-sol and gpt-5.6-luna, and Google's gemini-3.7-flash - on the same harness, tools and customer simulator. sol, luna and magnus ran at the same low reasoning setting; gemini-3.7-flash at its API defaults.

Figure 1. tau2-bench banking knowledge domain, all 97 tasks, one trial per agent, 200-step cap. gpt-5.6-sol and gpt-5.6-luna ran at reasoning_effort low, the same setting as the magnus · low row; gemini-3.7-flash ran at its API defaults. Identical retrieval tools (BM25 + dense embeddings + sandboxed shell) and the same gemini-3.5-flash customer simulator for every agent. Full credit only for the complete required action sequence.

Magnus solves 40 of 97 - the most of any agent we measured, ahead of gpt-5.6-sol at the same reasoning setting and well ahead of gemini-3.7-flash. And it does so while finishing first. An agentic task is twenty-odd short tool-calling turns, and the customer is on the line for all of them; the latency that matters is how long a finished task takes.

Figure 2. Tasks solved against median wall time per completed conversation. Up and to the left wins. Durations include the customer simulator and retrieval tools, identical for every agent.

Magnus finishes the median conversation in 57 seconds. gpt-5.6-sol takes 79 to solve slightly fewer - so magnus is three tasks ahead, nearly thirty percent faster per task, from an API bill roughly a tenth of the size. gemini-3.7-flash takes a minute and a half to solve nine fewer. The most accurate agent on the chart and the fastest finishing agent are the same model.

Availability

Celeris-1 Magnus is live for all API customers today at inference.celeris.ai/celeris-1-magnus/v1. It shares celeris-1's OpenAI-compatible spec, so switching is a one-line change. Reasoning is controlled with reasoning_effort and is on by default; set it to none for celeris-1-style direct answers, or low for the agentic configuration we report here. celeris-1 isn't going anywhere - use it where latency is the product, and magnus where the task is.

Notes on the measurement

τ³-Banking is the tau2-bench banking knowledge domain, unmodified; grading is the harness's own action-sequence checker. gpt-5.6 tool use with reasoning requires OpenAI's Responses API and was driven through it.
Reasoning settings: gpt-5.6-sol and gpt-5.6-luna at reasoning_effort: low, matching magnus · low; magnus · off with reasoning disabled; gemini-3.7-flash at its API defaults (dynamic thinking).
One trial of 97 binary tasks per agent, so run-to-run variance is material. Treat single-run gaps under about seven points - including magnus against sol - as suggestive rather than settled; multi-trial runs are queued.
magnus-low's 57 s is from a fully serial rerun; the other agents were timed at concurrency 2, which adds some self-contention (magnus-low itself measured 65 s under those conditions).
Magnus was not trained or finetuned on this benchmark, and the banking knowledge base was not in its training data.

Celeris-1 Magnus is live.

celeris-1 for the page load, magnus for the job.

Get started