Suhail launches Particle and its first model, Particle-1: a DeepSeek V4.1 Flash post-train that uses about 45% fewer tokens on coding tasks

Background and pain points

Agentic coding gets expensive mostly because of token volume, not unit price: repeated reasoning, rebuilds and long final confirmations push a single task to tens of thousands of output tokens. Mixpanel founder Suhail Doshi, who previously founded Playground, announced his new company Particle on X. Particle positions itself as an inference layer for open-weight models, and its first model, Particle-1, aims to use fewer tokens without losing accuracy.

How it works

According to Particle's blog, Particle-1 is post-trained from DeepSeek V4.1 Flash in internal reinforcement-learning environments, with a reward that values both task success and fewer tokens spent on reasoning and tool calls. Training adds only a rank-64 LoRA adapter over the experts in all layers, about 1.25% of the parameters. Policy updates use GSPO, which weights and clips whole sequences; Particle says a token-level loss oscillated and diverged. Tasks come from its synthetic data system Datamaker in Harbor format and are solved and verified through the Terminus 2 agent, in a coding stage followed by a data engineering stage. Particle attributes the savings to less deliberation, stopping once the task is done, and fewer trial-and-repair rounds; reasoning tokens account for about 90% of the savings on three of the benchmarks.

Benchmarks and numbers

All figures are vendor-reported, using the Terminus 2 agent, a 512k-token limit and max reasoning: 28.4% to 52.2% fewer output tokens and 34.4% to 51.6% lower cost per task across four benchmarks; $0.2211 per task versus $0.3368 for DeepSeek V4.1 Flash and $0.6523 for GLM 5.3 Flash; and a 72.57% pass rate versus 71.15% for the base. Outside its training domains, GPQA Diamond is 90.40% (base 90.91%) and Tau Airlines 86.00% (base 88.00%), slightly lower. The blog's cost math uses $0.30 input and $1.20 output per million tokens, while the docs now list $0.25 and $1.00. No independent re-test has been published yet.

Getting started

The Particle API is OpenAI-compatible; call the model as particle-1 (public ID particle-ai/particle-1). It accepts text and images, supports tool calling and reasoning_effort levels, and defaults to high reasoning. It is also selectable on OpenRouter and Vercel AI Gateway. A good first test is to run the same batch of tasks in your coding agent on the base model and on Particle-1 and compare pass rates and token bills. Particle has optimized only for software engineering so far, so results in other domains are unproven.