Particle-1: Suhail's startup promises 45% fewer tokens than DeepSeek V4.1 Flash
Particle unveiled Particle-1 on October 28, a DeepSeek V4.1 Flash post-trained to handle agent coding tasks with far fewer tokens, at the same per-token price. The claimed gains have not yet been verified by third parties.
Particle, the new company from Suhail Doshi (founder of Mixpanel and Playground), unveiled its first model, Particle-1, on October 28. The goal is not to beat the best models on accuracy, but to do the same coding agent work while spending far fewer tokens, and therefore less money.
A DeepSeek retrained to be brief
Particle-1 is not a model built from scratch. Particle started from DeepSeek V4.1 Flash, an open model, and post-trained it with reinforcement learning (GSPO method) on a small LoRA adapter of rank 64, which modifies only about 1.25% of the parameters. Its reward function values tasks completed with fewer reasoning and tool-call tokens. Training proceeded in successive stages, domain by domain (the “Cascade RL” approach), using RadixArk’s open-source Miles framework.
The per-token price is unchanged: $0.30 per million input tokens, $0.006 for cached input, and $1.20 for output, same as the base model. The savings come solely from the number of tokens consumed. Particle-1 is available through Particle’s API, and Suhail Doshi announced an inference service via OpenRouter.
Claimed gains, measured by Particle
According to Particle, the model consumes on average 45% fewer tokens. Across four coding agent benchmarks (Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1), it would produce 28.4% to 52.2% fewer output tokens and cost 34.4% to 51.6% less per task than DeepSeek V4.1 Flash, with accuracy equivalent to the base model.
On average, Particle reports 72.57% success, versus 71.15% for DeepSeek V4.1 Flash and 66.82% for GLM 5.3 Flash, for an average cost of $0.22 per task, versus $0.34 and $0.65. General capabilities are said to be preserved (GPQA Diamond 90.40% versus 90.91%). Particle also mentions a sign of generalization to law, but on a single task from Harvey’s LAB benchmark.
What remains to be verified
All of these numbers come from Particle. No independent leaderboard has reproduced them yet, and the claimed accuracy gaps are small. The real benefit will also depend on the tasks: a brief agent is only useful if it does not skip steps.
Our take
Particle-1 illustrates a deeper trend: for coding agents, the displayed price per million tokens matters less than the number of tokens needed to finish a task. A model that thinks for less time to reach the same result is also faster. The method is interesting for developers, because it starts from an existing open model and adjusts only a small fraction of the parameters. But as long as the gains are measured only by the vendor, they should be treated as a promise. The useful test will be simple: compare, on your own tasks, the bill and the success rate against the original DeepSeek.
Sources
- Particle — Introducing Particle-1 (launch post) Primary source
- Suhail Doshi on X — Particle-1 announcement Primary source
- Suhail Doshi on X — results and pricing Primary source
- Suhail Doshi on X — inference service on OpenRouter Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
