New coding model aims to save money by using fewer words

Particle, a new company in San Francisco founded by Suhail Doshi, announced itself and its first model, Particle-1, on October 28, 2026. The model is built for software development, meaning writing and fixing code.

Particle-1 was not built from scratch. It was post-trained on DeepSeek V4.1 Flash (version 1017). According to Particle, this was done with a so-called LoRA adapter, a kind of small add-on that adjusts an existing model. Here it covers about 1.25 percent of the base model's parameters, the numbers that make up the model.

Rewarding short answers

According to Particle, the model was trained in internal environments where it is rewarded for finding correct solutions with as few tokens as possible. Tokens are the small chunks of text an AI model reads and writes, and they are typically what you pay for. Fewer tokens therefore usually means a lower bill.

Particle says the model uses about 45 percent fewer tokens on average than the DeepSeek base model across a wide range of tasks. In the company's tests on four benchmarks (Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1), it used 28.4-52.2 percent fewer output tokens. The average cost per task was 34.4-51.6 percent lower.

According to Particle's summary, a task costs about $0.22 with Particle-1 versus about $0.34 with the base model. That is roughly 34 percent lower. It is not clearly stated what the figures are based on.

Does it cost accuracy?

Particle says the share of solved tasks falls within the base model's confidence intervals. That means the difference does not appear to be larger than the normal uncertainty in the measurements. The company therefore does not believe the savings come at the expense of quality. According to Particle, about 90 percent of the savings on most benchmarks come from reasoning tokens, the tokens the model uses to think its way to the answer.

Particle also points to early signs that the effect extends beyond coding: on review of legal documents, the model used 24.6 percent fewer tokens.

Caveats and plans

All figures come from Particle's own tests, and there is no indication that anyone else has verified them. The 34-52 percent range also means the savings vary between task types.

Particle-1 is available through Particle's API, which developers can use to connect the model to their own tools. The company plans to expand the model to more areas than software development. Aravind Srinivas, Garry Tan and Aaron Levie have responded to the announcement on social media.