Particle
Particle-1 uses 45% fewer tokens than DeepSeek V4.1 Flash
Particle post-trained DeepSeek V4.1 Flash to use fewer tokens. In Particle's own tests, cost per task fell 34.4% to 51.6% at similar pass rates.
Particle-1, a software engineering model from the new inference company Particle, uses about 45% fewer tokens on average than the DeepSeek model it is built on. Because the API bills per token, Particle says a task costs 34.4% to 51.6% less in its own tests, with pass rates level with the base model.
Particle launched Particle-1 on Oct. 28, 2026. “Today, we’re launching Particle to make highly optimized intelligence accessible to everyone,” the Particle team wrote in Introducing Particle-1.
Reinforcement Learning Rewards Shorter Work
Particle post-trained Particle-1 on DeepSeek V4.1 Flash with reinforcement learning. The reward favors solving a task with fewer tokens for reasoning and tool calls. Particle trained a rank-64 LoRA adapter, a small add-on layer of weights, across the experts in all layers. It is about 1.25% of the model’s parameter count, trained with Group Sequence Policy Optimization.
The first release focuses on software engineering. Particle plans other domains in future releases.
Particle’s Tests Show Lower Cost at Similar Pass Rates
On four benchmarks, Particle-1 used 28.4% to 52.2% fewer output tokens and cost 34.4% to 51.6% less per task than DeepSeek V4.1 Flash. Particle reported its own results on Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1 in the Oct. 28, 2026 post, and also compared against GLM 5.3 Flash.
Pass@1 rates, the share of tasks solved on the first try, fall within the confidence intervals of the base model. On DeepSWE 1.1, Particle-1 cost $0.2211 per task, averaged 59,580 output tokens and passed 72.57% of tasks.
Outside its training domain, Particle reports accuracy close to the base model. Particle-1 scored 90.40% on GPQA Diamond against 90.91% for the base, 86.00% on Tau Airlines against 88.00%, 86.08% on Standard MMLU against 85.98% and 89.18% on BabyVision with tools against 89.60%.
Reasoning Accounts for Most of the Savings
Reasoning tokens make up roughly 90% of the savings on Terminal-Bench 2.1, ReactBench and ADE Bench, Particle said, and tasks finish sooner as a result. The company describes four patterns: less reasoning for the same actions, stopping once the task is complete, fewer rounds of trial and repair, and more focused exploration and verification.
Particle-1 Is Live on a Paid API
Particle-1 is generally available through Particle’s paid API, which runs on prepaid credits topped up by card. The model ID is particle-1, and the API is OpenAI-compatible. The context window is 1,048,576 tokens, and output can reach 131,072 tokens. The model takes text and images and returns text. It supports tool calling and reasons by default, at high effort.
Particle’s model catalog listed prices on Nov. 1, 2026 of $0.25 per million input tokens, $0.03 per million cached input tokens and $1.00 per million output tokens. Reasoning tokens are not charged separately. The Flex service tier takes 50% off token prices for work that can wait. Standard inference runs under zero data retention.
The same API also serves DeepSeek V4.1 Flash and GLM 5.3 Flash.
Our View
We think reasoning is where developers gain most, because thinking time is waiting time. Cutting it shortens each task as well as the bill. The API is OpenAI-compatible, so a developer running coding agents can try it by switching the model ID. That payoff holds only if the shorter reasoning keeps solving the problems in the developer’s own repositories, which four benchmarks can only sample.