As of: 2026-10-25 23:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/m/celeris-1/

# Celeris-1

Celeris. Status: current. Type: language model. Released: 2026-07-24. Training cutoff: not published.

## How to call it

| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Celeris API (OpenAI-compatible chat completions) | `celeris-1` | https://inference.celeris.ai/celeris-1/v1/chat/completions | https://docs.celeris.ai/ |

## Pricing

- Input: $0.20 per 1M tokens
- Output: $0.70 per 1M tokens
- Unit as published: per 1M tokens (USD, input and output metered separately)
- Source: https://celeris.ai/pricing.html
- Last checked: 2026-10-25

## Specs

- Input: text
- Output: text
- Context window: not published
- Max output: not published
- Open weights: no
- Licence: proprietary

## What it has not seen

- No published training cutoff. Counted from its release (2026-07-24): at least 1,209 AI events since (259 major, 38 historic), as of 2026-10-25.
- Everything since June 2026: https://postcutoff.com/since/2026-06/

## What stands out

- Diffusion language model built for speed: Celeris describes it as 'the world's fastest language models, powered by diffusion': it 'starts with a rough representation of the response and progressively refines it into language' instead of generating one token at a time. Source: https://celeris.ai/
- #1 output speed on Artificial Analysis (found after launch): Celeris reported 2,038 output tokens/s on Artificial Analysis (Aug 4, 2026 X post, ~2.3M views); RuntimeWire later cited 2,157.9 tok/s and 0.64 s TTFT; AA's model page showed 1,605 tok/s, #1 of 59, on 2026-10-25. Figures differ by snapshot. Source: https://artificialanalysis.ai/models/celeris-1
- 158 ms complete answers: Vendor benchmark: 75.9% on MMLU-Pro (5-shot CoT) at 158 ms p50 server-reported latency with reasoning budget 0 (up to 81.4% with reasoning enabled, per homepage), vs Inception Mercury 2 at 63.7% / 257 ms. Source: https://celeris.ai/blog-benchmarks

## Notes

Made by Celeris (Celeris Labs), led by Marqo co-founders Tom Hamer and Jesse Clark (RuntimeWire, press summaries); company site celeris.ai. Release date 2026-07-24 per Artificial Analysis and launch coverage (BenchLM says Jul 22; not settled). Artificial Analysis (read 2026-10-25, https://artificialanalysis.ai/models/celeris-1): #1 of 59 non-reasoning models on output speed (1,605.4 tok/s, TTFT 0.56 s), Intelligence Index 6 (#58 of 71) — far below frontier models; AA lists $0.20/$0.70 and a ~131k context window, but Celeris's own docs give no context window or max output, so those fields are left out (one press summary said 8,192). Some third-party listings show older $2/$6 pricing; the Celeris pricing page shows $0.20/$0.70. Image input is listed by AA but not documented by Celeris, so modality_in is text only. Sibling products: Celeris-1 Decision (Oct 8, 2026; decision/scoring model, $0.04 per 1M input, output free) and Celeris-1 Magnus (named in navigation, no details). No timeline entry yet.

Celeris-1 is a proprietary diffusion LLM aimed at latency-critical uses (voice agents, coding tools, agent loops) where speed matters more than peak intelligence. Comparable to Inception's Mercury line.

```bash
curl https://inference.celeris.ai/celeris-1/v1/chat/completions -H "Authorization: Bearer $CELERIS_API_KEY" \
  -H "Content-Type: application/json" -d '{"model":"celeris-1","messages":[{"role":"user","content":"Hi"}],"max_tokens":256}'
```

Sources: [Celeris](https://celeris.ai/) · [Pricing](https://celeris.ai/pricing.html) · [Docs](https://docs.celeris.ai/) ·
[Benchmarks blog](https://celeris.ai/blog-benchmarks) · [Artificial Analysis](https://artificialanalysis.ai/models/celeris-1) ·
[RuntimeWire](https://runtimewire.com/article/celeris-1-tops-ai-speed-ranking-diffusion-model) ·
[Celeris on X (Aug 4, 2026)](https://x.com/Celeris_ai/status/2084429307422872026)

Model page: https://postcutoff.com/m/celeris-1/