DeepSeek-Based Particle-1 Cuts Cost per Coding Task by 34%

Particle, a new San Francisco company founded by Suhail Doshi, launched its first model on 28 October 2026. According to Particle’s launch post, Particle-1 is post-trained on DeepSeek V4.1 Flash and uses “around ~45% fewer tokens on average across a wide variety of tasks”, with a cost per task in its summary table of $0.2211 against $0.3368 for the DeepSeek base model.

By our arithmetic, that is 34.4% lower cost per task. Perplexity chief executive Aravind Srinivas posted the same day that his company looks forward to working with Particle “on tools like search and model orchestration”.

Business Pill · Usage-Based Pricing

An evergreen Business Engineer explainer on usage-based pricing. It explains the general idea only and does not discuss the companies in this piece.

What Particle Launched

Particle’s site describes the company as “building a new AI inference layer for open weight models”. Its about page states the mission as making “intelligence as accessible and abundant as electricity” and names Suhail Doshi as founder. Doshi’s X profile lists him as past chief executive and founder of Mixpanel.

The launch post says the first release “focuses on software engineering”. Particle trained the model “in reinforcement learning environments developed internally, with the goal of solving tasks with fewer tokens while maintaining accuracy”. It describes a rank-64 LoRA adapter whose trainable parameters are equivalent to “about 1.25% of the model’s parameter count”.

Particle’s documentation lists three models on its API, all with a 1,048,576-token context window: Particle-1, DeepSeek V4.1 Flash and GLM 5.3 Flash. It lists Particle-1 at $0.25 per million input tokens, $0.03 per million cached input tokens and $1.00 per million output tokens.

What Was Posted on X

Perplexity CEO Aravind Srinivas’s post says Particle is a new company focused on making open-weight models more token-efficient and cost-efficient to serve, and that Perplexity looks forward to working with it on tools like search and model orchestration.

The Numbers Particle Published

The launch post’s summary table compares three models across four benchmarks: DeepSWE 1.1, Terminal-Bench 2.1, ReactBench and ADE Bench. Particle-1 shows a task pass rate of 72.57%, against 71.15% for DeepSeek V4.1 Flash and 66.82% for GLM 5.3 Flash.

Mean output tokens per task are 59,580 for Particle-1, 83,246 for DeepSeek V4.1 Flash and 98,155 for GLM 5.3 Flash. Cost per task is $0.2211, $0.3368 and $0.6523. Particle says that across the benchmarks Particle-1 uses “28.4%–52.2% fewer output tokens” than DeepSeek V4.1 Flash, and that pass rates are “within the confidence intervals of the base model”.

On general benchmarks outside its training domains, the post lists GPQA Diamond at 90.40% against 90.91% for the base model, and Standard MMLU at 86.08% against 85.98%. It says reasoning tokens account for roughly 90% of the savings on three of the four benchmarks.

Cost per Solved Task

Dividing cost per task by pass rate gives a cost per solved task. By our arithmetic, that is about $0.305 for Particle-1, $0.473 for DeepSeek V4.1 Flash and $0.976 for GLM 5.3 Flash. On that basis Particle-1 is about 36% cheaper than its base model per solved task.

Particle also reports an early test outside coding: on one Harvey LAB contract-review task, Particle-1 used 24.6% fewer output tokens than the base model. The post presents this as an illustration, not a benchmark.

The Structural Read

As we read the launch post, Particle is competing on the cost of serving open-weight models rather than on a new base model: Particle-1 is a post-trained version of DeepSeek V4.1 Flash, served alongside DeepSeek V4.1 Flash and GLM 5.3 Flash on its own API.

As we read the benchmark table, the saving comes mostly from shorter reasoning: Particle says reasoning tokens account for roughly 90% of the token savings on three of its four benchmarks.

Aravind Srinivas posted that Perplexity looks forward to working with Particle on search and model orchestration; his post does not describe the terms.

Three Implications

FOR DEVELOPERS Particle’s documentation lists Particle-1 at $0.25 per million input tokens and $1.00 per million output tokens, with a 1,048,576-token context window.

FOR MODEL MAKERS Particle says the model was trained with a rank-64 LoRA adapter whose trainable parameters equal about 1.25% of the model’s parameter count.

FOR PERPLEXITY Srinivas’s post names search and model orchestration as the areas where it expects to work with Particle.

The Business Engineer Lens

This story maps onto the Business Engineer framework Tokenomics: The Economics of AI.

The framework describes the token this way: “It is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”

Particle-1 targets the number of tokens a model spends per task, not the price per token. The framework does not discuss Particle.

What Is Not Established

Every benchmark figure comes from Particle’s own post; we have not seen an independent evaluation. Particle’s about page lists backers including Y Combinator, Caffeinated Capital and 8VC, but we found no funding amount stated by the company.

Srinivas’s post does not describe the terms of any arrangement between Perplexity and Particle, and does not say Perplexity invested. Our cost-per-solved-task figures divide the published averages and are not a measure Particle reports.

The Bottom Line

Particle’s first model keeps DeepSeek V4.1 Flash’s pass rate within the confidence interval while using fewer output tokens, according to its own benchmarks. Its published cost per task is $0.2211 against $0.3368 for the base model, 34.4% lower by our arithmetic.