# The 158 ms call boundary

Sent July 29, 2026

Sponsored by Celeris. The benchmark results in this issue were produced and originally published by the Celeris team.

Agentic workflows do not wait once; they wait at every model call. Celeris-1 is a diffusion language model built for repeated, structured steps such as routing, extraction, scoring, and query rewriting. In Celeris’s provider-run MMLU-Pro test, it scored 75.9% at a 158 ms p50 server response time. A separate long-form run measured 1,664 output tokens per second and 0.58 seconds for a complete response.

## Provider-run measurements

- MMLU-Pro, full test set, five-shot chain-of-thought, strict scoring, reasoning budget set to zero: 75.9%
- p50 server response time in the same provider-run MMLU-Pro test: 158 ms

## The limit

These measurements were designed and run by Celeris, not an independent lab, and they do not rank every workload. Speed changes with the prompt, output length, region, and network. Retain a slower, more capable model for very long-form generation or extended multi-step reasoning. Test a labeled fixture at each node and keep a fallback.

## Where it earns a trial

- Route to a fixed label: Reject any output outside the allowed tool or intent set.
- Extract a fixed schema: Ask for JSON, then validate types and required fields before the next step runs.
- Score a bounded decision: Calibrate the result against human-labeled cases instead of trusting fluent reasons.
- Rewrite a retrieval query: Preserve names, dates, and constraints, then judge the retrieved result rather than the query’s prose.

## The swap is deliberately small

- Compatible request shape: The API supports OpenAI-compatible chat completions and Responses at https://inference.celeris.ai/celeris-1/v1.
- Current limits and pricing: The model documentation set an 8,192-token total for prompt plus requested output at publication. Current pricing is $0.20 per million prompt tokens and $0.70 per million completion tokens.

## Decision note

Put Celeris-1 on one repeated, checkable call boundary. If the task still passes and the trace gets shorter, keep it there.

## Partner links

The links below are part of this sponsored edition.

- [Test Celeris-1](https://celeris.ai/?utm_source=benchlm&utm_medium=web&utm_campaign=celeris_1_archive&utm_content=final_cta): Read the model documentation and try the API on a bounded workload.
- [Read the speed methodology](https://celeris.ai/?utm_source=benchlm&utm_medium=web&utm_campaign=celeris_1_archive&utm_content=speed_methodology): Review how Celeris measured the provider-run latency results.

Archive copy preserves the sponsor disclosure, provider attribution, limits, and pricing stated when this partner brief was sent.

Canonical page: https://benchlm.ai/newsletter/issues/2026-07-29-celeris-1-sponsored