Particle-1

ExplainX summarizes the Particle-1 launch and separates claimed from verified results: Particle says the model completes the same work as DeepSeek V4.1 Flash with about 45% fewer output tokens on average, ranging 28.4% to 52.2% across four benchmarks, at 34.4% to 51.6% lower cost per task, with pass@1 inside the base model's confidence intervals. It found no independent verification at publication and notes weights were not announced, so access is API-only. The post describes the training setup: internally developed reinforcement-learning environments with an autonomous synthetic data system called Datamaker, tasks in Harbor format verified automatically and run under the Terminus 2 agent harness, and training in successive stages following the Cascade RL approach, with the first stage on coding and a second stage shifted mainly to data engineering.

CellCog details Particle-1 as DeepSeek V4.1 Flash with a rank-64 LoRA adapter covering roughly 1.25% of its parameters, trained by Suhail Doshi's new company Particle and released on October 28, 2026. The model is listed as particle-ai/particle-1 with a 1,048,576-token context, and is reachable through Particle's own API and OpenRouter, with public weights not released. The write-up notes that every benchmark number in circulation comes from Particle's launch post, and that as of October 29 no outside lab had rerun them. It frames the release around agent workloads, citing Box CEO Aaron Levie's remark that agents will consume far more tokens than people do today, and flags token savings as a price cut for per-token billing.

Plotly benchmarked Particle-1 on its own data-analytics exam and reports the model passed 40 of 43 questions, compared with 36 for the DeepSeek V4.1 Flash base, with all four extra points coming from the hardest categories such as statistics traps and hidden needles. The full run cost $0.35 versus $0.46 for the base model and $1.31 for GPT-6.1 Terra, making Particle-1 the cheapest model on Plotly's leaderboard to score 40 or more. Median time per question fell to 16 seconds for Particle-1 from 21 seconds for DeepSeek V4.1 Flash, and Plotly places the model on the cost-accuracy frontier between GPT-6 Luna and GPT-6.1 Sol. Plotly also records the adapter covering about 1.25% of parameters, non-open weights, and OpenRouter pricing of $0.05 per million input tokens, $0.50 per million output tokens and $0.003 per million cached input tokens.

Particle, the startup launched by Mixpanel co-founder Suhail Doshi, has published the launch post for Particle-1, a model post-trained on DeepSeek V4.1 Flash and focused first on software engineering. The company states the model uses roughly 45% fewer output tokens on average, with 28.4% to 52.2% fewer output tokens and 34.4% to 51.6% lower average cost per task than the base model across Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1, while Pass@1 stays inside the base model's confidence intervals. The post gives per-task figures including $0.2211 average cost and 59,580 mean output tokens for Particle-1 versus $0.3368 and 83,246 for DeepSeek V4.1 Flash and $0.6523 and 98,155 for GLM 5.3 Flash, with task pass rates of 72.57%, 71.15% and 66.82%. Outside its training domains, Particle-1 stays close to the base on GPQA Diamond (90.40% vs 90.91%), Tau Airlines (86.00% vs 88.00%), MMLU (86.08% vs 85.98%) and BabyVision with tools (89.18% vs 89.60%).