Developers who pay per token for coding agents are the audience: Particle AI trains open models to stop reasoning once a task is done, so the same job costs less. The company is new, with no published certifications, no disclosed funding round and four open roles, so treat it as a young API option to test.
Particle AI is a San Francisco inference company that launched on 28 October 2026 with Particle-1, a coding model post-trained on DeepSeek V4.1 Flash. Particle reports that the model uses about 45% fewer output tokens across its tests. It sells API access to three models, and its standard inference runs under a zero data retention policy.
HQ: San Francisco, CA, USA · CEO: Suhail Doshi · Funding: Amount not disclosed. Backers named on its About page: Y Combinator, Caffeinated Capital, 8VC and a group of angel investors
About Particle AI
Particle AI is the trade name of Unmetered Machines, Inc., according to the company's zero data retention policy, and it works out of San Francisco. It launched publicly on 28 October 2026 with a model called Particle-1. Its founder is Suhail Doshi, who co-founded the analytics company Mixpanel and later founded the creative-tools company Playground, as RuntimeWire reported at the launch. The company is unrelated to Particle, the AI news reader app that shares part of its name. Its About page ties the company to a cost argument: intelligence should get cheaper as the systems behind it get more efficient, and developers come first.
The product is a developer API with three models: Particle-1, DeepSeek V4.1 Flash and GLM 5.3 Flash. Particle-1 is Particle's own work. It is post-trained from DeepSeek-V4.1-Flash-1017 with a small LoRA adapter and reinforcement learning, and the goal is to solve coding and data engineering tasks with fewer reasoning and tool-call tokens. The other two are third-party models served on the same platform: the DeepSeek base, from DeepSeek (see all DeepSeek models), and GLM-5.3-Flash, from Z.ai. The catalogue stops at those three, and larger siblings such as DeepSeek-V4-Pro are not on it. The full write-up of the flagship is on the Particle-1 model page.
The amount raised has not been published. The About page names Y Combinator, Caffeinated Capital and 8VC as backers, plus twelve named individuals, among them people tied to Stripe, Box, Superhuman, Mercury, Windsurf, Rippling, Lambda, Instacart and Factory. No valuation, round date or lead investor appears on the company's own pages, so none is stated here.
The team is small and still forming. The About page lists four open roles, all in San Francisco: post-training research engineer, inference performance engineer, research engineer for synthetic data, and an account executive. No headcount is published.
Particle is one of the AI startups in the directory. Against hosted-inference providers such as Together AI and Groq, its argument is not a faster chip or a cheaper token. Its per-token prices for Particle-1 match the ones it charges for the DeepSeek base, so the saving has to come from the model finishing sooner. That is a testable claim, and the launch numbers are Particle's own. Particle-1 is also listed on OpenRouter. Gaps to weigh before relying on it: no SOC 2, ISO 27001 or HIPAA statement is published, the model weights are not public, and the zero data retention policy says it does not create a service-level agreement.
Mission
To make intelligence as accessible and abundant as electricity, starting with a radically lower cost per unit of intelligence for developers.
Products
- Particle API (Developer inference API): https://platform.particle.ai/
- Particle-1 (Token-efficient coding model): https://particle.ai/blog/introducing-particle-1
- Hosted web search (API add-on): https://particle.ai/pricing
Particle AI Models on HokAI
Links
Frequently Asked Questions
Has Particle AI disclosed how much funding it has raised?
No. The company has not published a round size, a valuation or a lead investor. The backers it lists are Y Combinator, Caffeinated Capital and 8VC, plus twelve named individuals tied to Stripe, Box, Mercury, Instacart and other companies. Funding figures you may see for a company called Particle usually belong to the unrelated AI news app.
What does Particle AI sell in 2026?
A prepaid-credit API with three models. Particle-1 and its DeepSeek base are both priced at $0.25 per 1M input tokens, $0.03 per 1M cached input tokens and $1.00 per 1M output tokens, while GLM 5.3 Flash is $0.15, $0.03 and $0.50. Hosted web search costs $2.50 per 1,000 searches, and a Flex service tier takes 50% off token prices for work that can wait.
How does Particle AI handle data privacy and compliance?
Standard inference runs under a published zero data retention policy: raw prompts and outputs are not intentionally stored, and customer content is not used for training without a separate opt-in. It still keeps operational records such as token counts, timestamps and IP addresses. It publishes no SOC 2, ISO 27001 or HIPAA claim, and the policy says it creates no service-level agreement.
Which companies compete with Particle AI?
RuntimeWire lists Together AI and Azure among other hosts of the DeepSeek base model, so a team can buy the base model elsewhere; Particle's pitch is that its tuned version finishes tasks in fewer tokens at the same rate. Z.ai's GLM Flash model is cheaper per token on Particle's own list, so it wins when the rate matters more than tokens per task. Particle loses on breadth: it lists three models.
How do you start using Particle AI?
Sign in at platform.particle.ai with Google, create an API key, and add prepaid credits under Billing. Then point any OpenAI SDK at https://api.particle.ai/v1 and set the model to particle-1. The docs also show how to connect the OpenCode coding agent, and Particle-1 is listed on OpenRouter as an alternative route.
Top Alternatives
- Together AI: Pick Together AI if you want a wider hosted catalogue; pick Particle AI if you want a coding model trained to use fewer tokens.
- DeepSeek: Pick DeepSeek's own API for the lowest token price on the base model; pick Particle AI if tokens per task matter more than the rate.
- Z.ai: Pick Z.ai's GLM Flash line for cheaper tokens and native vision; pick Particle AI for a model tuned on coding and data engineering.