A newly founded startup, Particle, has released its first model: Particle-1, a version of DeepSeek V4.1 Flash trained for a single goal, using fewer tokens to finish a task. The company, founded by Suhail Doshi, co-founder of Mixpanel, says it cut the tokens generated by 45% on average while keeping performance unchanged.
The key is cost. When you pay per token, a model that reaches the same result with fewer tokens is effectively cheaper. Particle reports that across four coding benchmarks, the average cost per task falls from 33.7 to 22.1 US cents, at the same token prices as the base model. The pass rate moves from 71.15% to 72.57%, a difference within the confidence intervals.
Image: particle.ai — https://particle.ai
Where do the saved tokens go? Particle points mainly to internal reasoning, which accounts for about 90% of the savings on Terminal-Bench 2.1, ReactBench and ADE Bench. The company also admits a limit: in one ADE Bench case the model reused an existing macro instead of building a new one, so the saving does not always reflect the same work done more briefly.
The most telling data is the trade-off. Particle pushed the adapter to higher strengths on a contract-review task. At the setting it shipped, all eight reports passed every criterion; beyond that point savings keep rising but answer quality drops. At 8x, tokens fall by 78.7%, but only three of eight reports pass the strict check.
All figures come from the company and have not yet been reproduced by outside labs. The price on OpenRouter is also higher than the original model on DeepSeek: $0.25 per million input tokens and $1.00 output, against $0.15 and $0.60 off-peak. Where you buy therefore shapes much of the real saving.
Technically, Particle-1 is not a new base model but a rank-64 LoRA adapter, about 1.25% of the parameters, trained with reinforcement learning in coding environments the company built itself, with rewards for solving tasks using fewer reasoning tokens. It is available through Particle's API and OpenRouter, with a context window above one million tokens. For now the weights are not public: at publication there was no Hugging Face repository.
Particle presents it as the first step in a series of models for developers. Teams running coding agents will need to test whether the saving holds on their own workloads.
No comments yet. Be the first!