particle-1 model Cuts Coding Tokens ~45% on Particle API

The particle-1 model is Particle's first public release, a software-engineering-focused system post-trained on DeepSeek V4.1 Flash and launched on October 28. Particle's blog says it uses around 45% fewer tokens on average across a wide variety of tasks, finishes work sooner for end users, and is available through an API at api.particle.ai while matching the base model's per-token prices.

What actually happened

Particle published "Introducing Particle-1" on October 28 and framed the company around lowering the cost of capable models by training them to solve problems with fewer tokens. The particle-1 model is described as a LoRA post-train on DeepSeek V4.1 Flash, with a rank-64 adapter covering experts in all layers and trainable parameters equivalent to about 1.25% of the model parameter count.

The headline efficiency table on the blog compares Particle-1 with DeepSeek V4.1 Flash and GLM 5.3 Flash on cost per task, mean tokens, and accuracy. On DeepSWE 1.1, Terminal-Bench 2.1, ReactBench, and ADE Bench, Particle lists mean output tokens of 59,580 versus 83,246 for DeepSeek V4.1 Flash and 98,155 for GLM 5.3 Flash, with task pass rates of 72.57%, 71.15%, and 66.82% respectively on that row's accuracy line. Treat those as company-reported figures from the published table, not third-party audits.

RuntimeWire's same-day piece stressed the business angle: Particle sells efficiency gained through post-training while charging the same per-token rates as its base model. Doshi's path runs through Mixpanel and Playground before this inference-focused company. Pricing footnotes on the Particle blog list uncached input, cached input, and output at $0.30, $0.006, and $1.20 per million tokens for both Particle-1 and DeepSeek V4.1 Flash.

What the benches actually test

Terminal-Bench 2.1 covers terminal workflows such as compiling, debugging, and configuring systems. ReactBench hits agents building and fixing React apps. ADE Bench uses real dbt projects and databases for analytics and data engineering. DeepSWE 1.1 stretches into longer software-engineering tasks that need codebase understanding and verification. Particle says all four used the Terminus 2 agent, a 512k-token limit, temperature 1, top_p 0.95, and max reasoning, with Terminal-Bench excluding the fastText task.

Across those benches the company claims 28.4%–52.2% fewer output tokens and 34.4%–51.6% lower average cost per task than DeepSeek V4.1 Flash, with Pass@1 rates inside the base model's confidence intervals. On general benches outside the training domains it reports GPQA Diamond 90.40% versus base 90.91%, Tau Airlines 86.00% versus 88.00%, Standard MMLU 86.08% versus 85.98%, and BabyVision with tools 89.18% versus 89.60%.

Trajectory notes on the blog say reasoning tokens account for roughly 90% of savings on Terminal-Bench, ReactBench, and ADE, while tool-call text contributes around 30% of savings on DeepSWE. Patterns include less deliberation for the same actions, stopping once verification passes, fewer repair loops, and tighter exploration. That is useful color, still company-authored.

How to read a launch without independent evals

Independent third-party evals are thin at launch, which Particle itself does not hide by pointing readers to its own tables. Compare the particle-1 model claim set with other efficiency pitches in our AI lane such as Liquid AI LFM3 and decision-focused stacks like Microsoft Decision-1 model or OpenAI Decisions API. Different labs optimize different bottlenecks; Particle is explicitly hunting tokens and wall-clock on coding agents.

Also note the generalization teaser. Particle shows a Harvey LAB contract-review illustration where the standard adapter used 24.6% fewer output tokens than the base with all eight reports passing both judges, while stronger adapter settings cut tokens further but missed criteria. That is a demo, not a legal product claim.

Fewer tokens at the same sticker price is a margin story for buyers only if quality holds on your repo, not just theirs.

What to do as a reader

If you evaluate the particle-1 model, run it on your own harness with the same agent settings Particle documents, and log token counts beside pass/fail. Read the Particle blog footnotes on pricing and cache assumptions before you spreadsheet a cost win.

Do not treat company benches as gospel until outside evals land. Sign up for API access only if you already have a coding-agent workload worth metering. Nothing here is investment advice, and an efficiency launch is not a promise of cheaper production forever.

Sources: Particle blog, RuntimeWire on Doshi and Particle.