Most AI model launches promise to be smarter. Particle’s first model promises to say less. The new San Francisco company, founded by Mixpanel founder Suhail Doshi, has released Particle-1, a coding model post-trained on DeepSeek V4.1 Flash. Particle says it uses “around ~45% fewer tokens on average” than the model it was built on, while keeping accuracy within the margin of error.
That matters because agentic coding tools are billed by the token, and long reasoning traces make up much of the bill. Particle is pitching efficiency rather than raw capability. It sells Particle-1 through its own API at exactly the same list price as the DeepSeek model underneath, so every token it saves is a direct cost saving. All the performance figures below come from Particle’s own launch post and have not been independently replicated.
What Particle-1 is
Particle-1 starts from DeepSeek V4.1 Flash 1017, the open-weight model DeepSeek released under an MIT licence in October. Particle trained it further with reinforcement learning in its own internal environments, rewarding the model for solving tasks with fewer tokens while keeping accuracy.
The company says the change is small in weight terms. It trained a rank-64 LoRA adapter, about 1.25% of the model’s parameter count, in two reinforcement-learning stages. The first covered coding only. The second shifted towards data engineering. Particle says that after the first stage alone the model was already using about 20% fewer tokens on data-engineering tasks it had not been trained on.
“This first release focuses on software engineering,” the post says. Particle plans to take the approach into other domains later. In his launch thread on X, Doshi wrote that the team “managed to keep parity in performance while using ~45% fewer tokens”. In a follow-up post he said one major goal is “continuous post-training: a new way to ship improvements to models as fast as we ship software.”
The benchmark claims
Particle tested the model on four agentic coding benchmarks: Terminal-Bench 2.1 for terminal work, ReactBench for building React apps, ADE Bench for analytics and data engineering on real dbt projects, and DeepSWE 1.1 for longer software-engineering tasks. “Across these benchmarks, Particle-1 uses 28.4%–52.2% fewer output tokens and achieves 34.4%–51.6% lower average cost per task than DeepSeek V4.1 Flash,” the post says. “Pass@1 rates are within the confidence intervals of the base model.”
The runs used the Terminus 2 agent, a 512,000-token limit, maximum reasoning and 16 attempts per task. Particle says reasoning tokens account for “roughly 90% of the token savings” on three of the four benchmarks. It describes four habits the training produced: less reasoning before the same actions, stopping once a task is complete, fewer rounds of trial and repair, and more focused exploration. In one example Git deployment task, Particle-1 used 15,892 output tokens against the base model’s 47,006.
On general benchmarks the scores stay close to the base model. Particle reports 90.40% on GPQA Diamond (base 90.91%), 86.00% on Tau Airlines (88.00%) and 86.08% on MMLU (85.98%).
Particle also tried the model on a single contract-review task from Harvey’s legal benchmark, with eight attempts per setting. At normal adapter strength, Particle-1 used 24.6% fewer output tokens and all eight reports passed every criterion. Pushing the adapter harder saved up to 78.7% of tokens, but then only three of eight passed the strict check.
Price and availability
Particle’s API documentation lists Particle-1 at $0.25 per million uncached input tokens, $0.03 per million cached input tokens and $1.00 per million output tokens. That is identical to the price it charges for DeepSeek V4.1 Flash 1017. It also serves GLM 5.3 Flash. All three have a context window of about 1 million tokens, and off-peak use between 20:30 and 04:00 New York time is half price. The documentation says Particle “operates zero-data-retention for inference traffic: prompts and completions are processed in memory and never persisted.”
The model is also available through Vercel’s AI Gateway and on OpenRouter, both announced on launch day.
Efficiency as the product
TSN’s analysis: Particle is betting that the next gains in coding agents will come from waste, not intelligence. That fits the wider push to cut inference bills, from Together AI’s tools for custom RL loops to Notion’s auto-router that sends work to cheaper models. Particle’s twist is to sell the tuned model at the same price as its base. Buyers need not trust a new price sheet, only the token counts in their own logs.
That makes the claim easy to test. Any team already running DeepSeek V4.1 Flash on agentic work can switch the model name and compare output tokens and pass rates on its own tasks. Results will vary with the agent harness, because Particle’s figures all come from Terminus 2.
Questions over the baseline
The numbers are Particle’s own, from Particle-built test runs. On X, one researcher questioned whether the comparison used an older DeepSeek checkpoint. Doshi replied: “We are comparing to 1017 baseline. I tried very hard to not over claim anything and credit.” There is also a gap in the published costs. The blog’s cost footnote assumes higher prices than the API now charges, so the dollar figures per task will not match a real bill, although the percentage saving still holds because both models share one price.
The legal result is one task with eight attempts. It shows the trade-off clearly but proves little about general legal work. Safety is a separate question: researchers at Fangcun recently flagged oversight gaps in the base model, and Particle has not published a safety evaluation of its own. The first independent runs on public leaderboards will show whether the 45% survives outside Particle’s own harness.
Sources
- Particle, “Introducing Particle”, Particle team, 28 October 2026
- Particle API documentation: models, pricing and data retention
- Particle, About page
- Suhail Doshi (@Suhail) on X, launch thread, 28 October 2026
- Suhail Doshi (@Suhail) on X, continuous post-training, 28 October 2026
- Suhail Doshi (@Suhail) on X, reply on baseline, 29 October 2026
- Vercel changelog, “Particle-1 is now available on AI Gateway”
Related stories
- DeepSeek Releases DeepSeek-V4.1-Flash-1017 Open-Weight Update
- Fangcun Finds DeepSeek V4.1 Flash-1017 Safer but Flags Oversight Gaps
- Together AI Lets Teams Write Their Own RL Loops on Open Models
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
