Particle-1 review, pricing and limits

Teams running terminal and codebase agents that pay per token are the target, since Particle-1 keeps general-test scores within about two points of its base while writing shorter reasoning. Every figure is self-reported as of 30 October 2026, so run your own 20 tasks before moving a production agent.

Particle-1 is a coding model released on 28 October 2026 by Particle AI, post-trained from DeepSeek V4.1 Flash to use about 45% fewer output tokens on average in the company's own tests. It takes text and images, has a 1,048,576-token context window and is sold through an API; the weights are not public.

Provider: Particle AI · Family: Particle

More about Particle AI on HokAI

Context window: 1,048,576 tokens · Max output: 131,072

Input modalities: text, image, tool-calls · Output: text, tool-calls

About Particle-1

Particle-1 is the first model from Particle AI, released on 28 October 2026 and sold only through an API. It is post-trained from DeepSeek-V4.1-Flash-1017, a Mixture-of-Experts model that DeepSeek released under the MIT licence. Particle did not retrain the base. It trained a rank-64 LoRA adapter over the expert layers, about 1.25% of the base's parameter count, using reinforcement learning with the GSPO algorithm in two stages: coding first, then coding plus data engineering. The training tasks came from Particle's own synthetic-data system, were run through the Terminus 2 agent, and each came with a way to check the result. Particle has not published a total parameter count for the combined model.

The headline claim is efficiency, not higher scores. Across Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1, Particle reports 28.4% to 52.2% fewer output tokens than the base, with the average at about 45%, and a 34.4% to 51.6% lower average cost per task. Pass@1 rates fall inside the base model's confidence intervals, so no accuracy gain is claimed. Each model ran 16 rollouts per task in the same agent with a 512k-token limit. On general tests outside its training domains, Particle reports GPQA Diamond at 90.40% against 90.91% for the base and standard MMLU at 86.08% against 85.98%. All of these are the vendor's own figures. No Artificial Analysis page or independent rerun was found on 30 October 2026.

Particle's trajectory analysis says reasoning tokens account for roughly 90% of the savings on three of the four benchmarks. In its git deployment example, output fell from 43,907 to 14,840 tokens, and the last two confirmation calls alone dropped from 11,445 to 1,765. Expect shorter wrap-up messages and fewer repeated checks. Particle also warns that extra tests and documentation can be useful, and that in one data-engineering example the saving partly reflects a different solution rather than a shorter explanation of the same one.

The cost reading needs one extra step, because the per-token rates for Particle-1 equal those Particle charges for its DeepSeek base. DeepSeek's own API lists the same base model at $0.15 input and $0.60 output per 1M tokens off-peak, and $0.30 and $1.20 at weekday peak hours. Particle's benchmark footnote prices both models at $0.30 and $1.20, which are DeepSeek's peak rates. If the 45% average held on your workload, a million base-equivalent output tokens would cost about $0.55 through Particle, against $0.60 off-peak or $1.20 at peak on DeepSeek direct. That counts output only, and off-peak input also costs more through Particle. The saving over DeepSeek direct therefore depends on when your traffic runs, and the clearest win is against peak-hour rates.

The model takes text and images and returns text, with a 1,048,576-token context window and an output ceiling of 131,072 tokens. The API is OpenAI-compatible, the docs cover LiteLLM as well, the public ID is particle-ai/particle-1, and the request alias is particle-1. Reasoning is on by default at high effort, so set the effort to none for quick calls. Tool calling is supported with a qualification in the docs, and structured output is limited to JSON object mode with no schema guarantee. The model is also listed on OpenRouter. The weights are not public, which puts it under proprietary models in the directory.

Standard inference runs under a zero data retention policy, and Particle says it does not train on customer content unless a customer opts in. On compliance, the company lists no SOC 2, ISO 27001 or HIPAA statement, and its policy says it creates no service-level agreement. Teams that need audited compliance, schema-enforced JSON or self-hosting should look elsewhere. GLM-5.3-Flash is cheaper per token on Particle's own price list, DeepSeek-V4-Pro is the Pro tier from the base's maker DeepSeek, and every model from the vendor sits on its provider page.

Pricing

Billing is prepaid credits from the [Particle AI](/hub/companies/particle-ai) platform, charged per token at $0.25 input, $0.03 cached input and $1.00 output per 1M tokens, in USD. Reasoning tokens are billed as output. The Flex service tier takes 50% off all token prices for requests that can wait in a queue, and hosted web search is a separate $2.50 per 1,000 searches. An empty balance returns a 402 error before anything is processed.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Token-efficiency post-training: An adapter over the expert layers worth about 1.25% of the base's parameters, trained with the GSPO algorithm so tasks end with fewer reasoning and tool-call tokens.
• Million-token context: A 1M-token window shared by prompt and completion, with a 32,000-token default output cap and a 131,072-token ceiling.
• Image input: Accepts base64 data URLs and public HTTPS image URLs, with no image-count cap, up to 16 MiB per image and 100 MiB per request.
• Reasoning dial: reasoning_effort takes none, low, high or max, and chat_template_kwargs accepts an exact integer effort from 1 to 100.
• Prompt caching: Repeated prefixes bill at the cached rate, and prompt_cache_key groups related requests; Particle's docs call cached tokens 4 to 8 times cheaper than uncached input.
• Flex service tier: Requests that can wait in a queue get a discount on token prices, and the X-Particle-Service-Tier header confirms the tier that ran.

Pros

• Charges the same per-token rate as its base, so any reduction in tokens flows straight to the bill without a pricing trade-off.
• In Particle's four general tests the tuned model lands within about two points of the base, so the tuning does not appear to cost much accuracy outside coding.
• Standard inference is zero-retention by default, a stronger privacy default than many hosted APIs publish.

Cons

• Every performance figure comes from Particle, and no independent evaluator had published results at launch.
• Publishes no SOC 2, ISO 27001 or HIPAA claim and no service-level agreement, which rules it out for audited workloads.
• JSON output has no schema guarantee, and tool calling is labelled qualified in the docs.
• Closed weights and a third-party base mean a base-model change or a Particle outage leaves no self-hosted fallback.

Benchmarks

• MMLU: 86.1% vendor-reported · 30 Oct 2026 — General-knowledge exam across 57 subjects, % correct.
• GPQA Diamond: 90.4% vendor-reported · 30 Oct 2026 — PhD-level science questions that are hard to search for, % correct.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Particle-1 Official Page