Particle-1, a post-trained version of DeepSeek V4.1 Flash, says it matches accuracy while using roughly 45% fewer output tokens
Suhail, the founder and former CEO of Mixpanel, has a new company. It is called Particle, and its pitch fits in one sentence: same answers, fewer words.
On October 28, 2026, Suhail announced the San Francisco-based AI startup alongside its first model, Particle-1. The company says the model uses about 45% fewer output tokens on average during evaluations while keeping its accuracy intact. In an industry that bills by the token, that is less a footnote and more the whole business model.
What Particle-1 actually is
Particle-1 is not built from scratch. It is a post-trained version of DeepSeek V4.1 Flash, an existing model that Particle took and refined.
The tuning came through reinforcement learning, according to the research findings on the launch. Reinforcement learning is a training method where a model gets rewarded for desirable behavior. In this case, the desired behavior was reaching the right answer without rambling on the way there.
The model is aimed at developers first. Particle says its initial focus is software engineering workloads, with plans to expand into other domains later.
Particle-1 is publicly available at particle.ai.
The numbers behind the efficiency claim
The headline figure is roughly 45% fewer output tokens on average. The benchmark breakdown shows a wider spread.
Across its evaluations, Particle-1 reportedly cut output tokens by between 28.4% and 52.2% compared with its base model. The cost figures follow a similar pattern. Particle says average cost per task came in 34.4% to 51.6% lower than DeepSeek V4.1 Flash.
Particle says the token reduction translates into faster end-to-end task execution.
Who is backing it
Particle is backed by Y Combinator and 8VC, along with a group of angel investors that includes Patrick Collison and Adam D’Angelo.
No funding amount was disclosed in the announcement materials.
The launch also came without a warm-up act. There was no prior public information about Particle or Particle-1 before the October 28 announcement.
Why token efficiency is becoming the battleground
Particle describes its mission as making optimized intelligence accessible and pushing the cost of intelligence toward zero. The research findings on the launch tie it to heightened industry interest in models that improve efficiency and cut development costs without sacrificing performance.
What this means for developers and the AI market
For developers, if the claims hold up in real-world use, teams running heavy coding workloads could see meaningfully lower bills and quicker turnaround on tasks. The cost-per-task reduction of 34.4% to 51.6% is the number procurement teams will care about most.
The caveat is that these are company-reported benchmark results. The wide range in token savings, from 28.4% to 52.2%, is itself a hint that results will vary depending on what you ask the model to do.
There is also a strategic question in Particle’s approach. Building on top of DeepSeek V4.1 Flash means the company did not need to train a frontier model from the ground up. Particle’s value proposition rests on how much better its tuning is than the base model, rather than on owning the foundation itself.
