Mixpanel Founder Launches Particle-1: Same Work, ~45% Fewer Tokens

Coding with AI is cheap until an agent starts chewing through tokens. Suhail Doshi, Mixpanel's founder, has a new company called Particle, and its first model, Particle-1, is aimed straight at that bill: finish the job, say less on the way.

Teaching a strong base model to shut up sooner

Particle-1 is not a from-scratch pretrained model. It is DeepSeek V4.1 Flash after post-training. The company says it used internal reinforcement-learning environments with a reward that pushes two things at once: solve the task, and spend fewer tokens on reasoning and tool calls. The first release focuses on software engineering; other domains are planned later.

Official comparisons span four suites: Terminal-Bench 2.1 for terminal workflows, ReactBench for building and fixing React apps, ADE Bench for analytics and data engineering, and DeepSWE 1.1 for longer software jobs. Versus base DeepSeek V4.1 Flash, Particle says Particle-1 uses about 45% fewer output tokens on average, with a range of roughly 28.4% to 52.2%, and 34.4% to 51.6% lower average cost per task, while Pass@1 stays within the base model's confidence intervals. Shorter trajectories also tend to finish sooner for end users.

Most of the savings are less overthinking

Looking at paired successful rollouts, Particle describes familiar patterns: less deliberation before the same actions; stopping once the solution is verified instead of re-checking; fewer repair loops; tighter exploration. The company says reasoning tokens account for about 90% of the savings on Terminal-Bench, ReactBench, and ADE, while on DeepSWE tool-call text also contributes around 30%.

In other words, the pitch is not a brand-new architecture. It is taking a capable model and training the rambling out of it—useful when you pay by the token and agents run for many steps.

Available in agent workflows today

Particle-1 is live through Particle's own API, OpenRouter, and Vercel AI Gateway, with about a million tokens of context, aimed at terminal work, codebase edits, and data-engineering agents. Particle has also shown early signs of token savings outside coding, such as a contract-review example, but that is illustrative rather than the main launch claim.

The practical test for teams is simple: run Particle-1 and DeepSeek V4.1 Flash side by side in your own harness. Watch pass rate, total output tokens, and end-to-end time. If the savings come from less chatter rather than less work, this kind of post-training is doing its job.