Particle-1: DeepSeek Flash With 45% Fewer Tokens

§ 01Where the tokens go#

Particle compared runs on the same task and seed where both models passed, then read the trajectories side by side. The cleanest case is a calendar task from Terminal-Bench 2.1: “Both models produced the correct calendar event using four model calls and six commands”. The actions matched. What differed was the talk around them. The final completion and confirmation fell from 5,193 tokens to 1,078, and total output fell from 11,023 to 4,977.

Particle names four patterns: less reasoning for the same actions, stopping once a verified task is done, fewer rounds of trial and repair, and narrower exploration. By its count, “Reasoning tokens account for roughly 90% of the token savings on Terminal-Bench 2.1, ReactBench, and ADE.” On DeepSWE 1.1, tool-call text accounts for about 30 percent of the savings.

The post also flags its own caveat. In one ADE Bench case Particle-1 reused an existing key macro where the base model built a new one, and Particle writes that “the savings also reflect a different solution rather than merely a shorter explanation of identical work.” A shorter run is not always the same run, and Particle says so itself.

§ 02The numbers, averaged over four coding benchmarks#

The four benchmarks are Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1, all run through the Terminus 2 agent with 16 rollouts per task. On each one, Particle reports 28.4 to 52.2 percent fewer output tokens and 34.4 to 51.6 percent lower cost per task than DeepSeek V4.1 Flash, and “Pass@1 rates are within the confidence intervals of the base model.” The cost column prices Particle-1 and its base at identical rates ($0.30 per million uncached input, $0.006 cached, $1.20 output, at a fixed 95 percent cache hit rate), so the gap comes from tokens alone.

Two readings of that table. First, pooled across the four benchmarks, mean output falls from 83,246 to 59,580 tokens, a cut of 28.4 percent. The 45 percent headline is an average taken across tasks, so short tasks with large percentage cuts count as much as long ones (our reading of Particle’s wording). Second, accuracy outside coding barely moved: GPQA Diamond 90.40 percent against 90.91 for the base, Tau Airlines 86.00 against 88.00, MMLU 86.08 against 85.98, and BabyVision with tools 89.18 against 89.60.

§ 03What Particle-1 is#

Particle-1 is not a new base model. Particle trained a rank-64 LoRA adapter over the experts in every layer of DeepSeek V4.1 Flash, about 1.25 percent of the model’s parameters. The training was reinforcement learning in coding and data-engineering environments that Particle built itself, with a reward that pays for solving the task in fewer reasoning and tool-call tokens. The stack is Miles, RadixArk’s open-source RL framework, with Group Sequence Policy Optimization; Particle writes that a token-level loss “oscillated and eventually diverged.” In its own words: “This first release focuses on software engineering.”

OpenRouter lists particle-ai/particle-1 with a 1,048,576-token context. As of October 29, 13:05 UTC, there was no Particle-1 repository on Hugging Face, so the weights are not public.

§ 04The dial: how far the squeeze goes#

Particle also ran an experiment that most launch posts would leave out. It took a contract-review task from Harvey LAB (read four legal documents, check an arbitration clause against the institution’s rules, write a Word report) and scaled the adapter’s strength from 0.25x to 8x.

At the shipped 1x setting, all eight reports passed every criterion. The base model scored 62.5 percent on the same task while writing the most tokens. Past 1x the savings keep growing but answers start losing content, and at 8x only three of eight reports pass. Particle reads it plainly: aggressive scaling “compresses useful qualifications along with redundant analysis.” This is one task, and it is the most honest chart in the launch: token efficiency has a floor, and Particle shipped the setting just above it.

§ 05What it costs to run#

OpenRouter lists Particle-1 at $0.25 per million input tokens and $1.00 per million output tokens (read October 29). DeepSeek’s own API charges $0.15 input and $0.60 output for V4.1 Flash off-peak, double that at peak, and several hosts charge $0.30 and $1.20, the rates Particle used in its table. Particle-1 costs more per token than DeepSeek’s first-party price and uses fewer tokens, so the two effects partly cancel.

On output alone, using Particle’s mean output per task, that comes to $0.060 a task for Particle-1 at the OpenRouter rate, against $0.050 for DeepSeek off-peak and $0.100 at DeepSeek’s peak or host rate (our arithmetic from the listed prices). If you buy DeepSeek direct at night, the saving on output is gone. If you pay host rates or run in daytime, it is real. Particle makes a second argument that holds either way: “tasks finish sooner which is good for end-users.”

§ 06Our read#

CellCog does not route to Particle-1. Every CellCog tier runs on Claude Opus 5.5. We still follow this closely, because the bill for agent work is mostly tokens spent thinking, and Particle shows that a small, cheap adapter can remove a large share of them without moving accuracy, as long as you stop turning the dial before answers start to thin. On CellCog you pay for the work, not the hire, and the cost depends purely on how much work you assign. A model that does the same work in fewer tokens is the same goal approached from the model side.

§ 07What we are watching#

• An independent rerun. Every number here is Particle’s own. Third-party Terminal-Bench or DeepSWE runs on Particle-1 are what turn a launch table into a comparison.
• Weights and license. The base is MIT-licensed DeepSeek V4.1 Flash; whether Particle opens its adapter is not stated.
• Particle’s own API price. OpenRouter’s listing is the only public price as of October 29.
• The next domains. Particle says it plans to bring the method to other domains, and the Harvey LAB result is its first look outside coding.

§ 08Update log#

As of October 29, 2026, 13:20 UTC: animated explainers added under Where the tokens go and The dial.

As of October 29, 2026, 13:12 UTC: page opened from Particle’s launch post (October 28), Suhail Doshi’s announcement at 17:36 UTC, OpenRouter’s listing and DeepSeek’s price page.