What happened
Vercel’s AI Gateway model page describes Particle-1 as a model from Particle AI for agentic coding and software engineering. It says the model was post-trained on DeepSeek V4.1 Flash to complete tasks with fewer tokens. The page lists text and image inputs, reasoning, tool use, and implicit prompt caching among its capabilities. These are product-page descriptions; the listing does not provide an independent evaluation of the claimed task efficiency or establish that it will use fewer tokens on a particular workload.
The model identifier for requests is particle/particle-1. Vercel says it can be called through the AI SDK’s generateText and streamText functions, or through OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages APIs by changing the base URL. The gateway authenticates a request and routes it to an available provider. The page shows a TypeScript example using streamText; its getting-started instructions also describe installing the AI SDK, creating an API key, and setting AI_GATEWAY_API_KEY. This is a documented gateway integration, not a claim that the model is bundled with those client APIs or available through every third-party tool.
The listing specifies a 1,048,576-token context window and a maximum output of 131,072 tokens. The output limit includes reasoning tokens, according to the model parameters. Text prompts and responses share the context window; image input is supported by URL, base64, or Uint8Array and counts as input tokens. Vercel lists pricing of $0.13 per million input tokens and $0.50 per million output tokens. The detailed provider row gives the input price as $0.13/M and output price as $0.50/M, alongside a read-cache price of $0.02/M. The page’s summary price display does not separately explain when cached reads apply.
The page lists Particle.AI as the provider and gives a release date of 28 October 2026. It also describes gateway routing options, including the ability to restrict provider routing to selected providers. The supplied material does not state a separate subscription price, a free allocation for this model, or a performance guarantee. Users’ model use is subject to Particle.AI’s terms and privacy policies, linked from the listing.
What the archive adds
This is a new access detail on a known story, not a new model launch. Particle-1’s launch was already covered as a coding model post-trained on DeepSeek V4.1 Flash. The gateway listing now specifies a callable model ID, API routes, limits, and prices. A separate DeepSeek V4.1 Flash listing is related background, but does not establish how Particle-1 performs through this gateway.
Trend signal
This listing is consistent with harness architecture outliving model choice: a model is exposed through a gateway and familiar request formats, so teams can evaluate it within an existing application path. One listing is a small signal, however; it does not show that the model is interchangeable in practice, or that routing abstraction removes differences in behavior, cost, or reliability.
Why now
The gateway page lists Particle-1’s release date as 28 October 2026 and provides a model-specific route and pricing. The supplied source does not explain the timing of the listing or describe a launch campaign, so the update is best understood simply as a newly documented availability path.
Why it matters
For a team already using Vercel AI Gateway, the listing reduces the work needed to test Particle-1: the model ID, API options, limits, and prices are visible in one place. Existing code built around the AI SDK or the documented OpenAI and Anthropic API formats may need less integration work than a provider-specific client would, although the page does not guarantee compatibility with every feature or application.
The numbers make a first-pass budget possible, but not a cost comparison by themselves. Input and output prices differ, and the model’s stated goal of using fewer tokens is a claim to test against the team’s own tasks. Context size and maximum output are ceilings, not evidence that a workload needs them or that long runs will be economical. Teams should compare task completion, token use, latency, and review effort on representative coding jobs before changing defaults. The listing provides a route for that evaluation; it does not establish that Particle-1 beats the underlying model or another coding model.
Limits and uncertainty
The source is a Vercel model listing and integration page. It does not present benchmark results, reproducible task comparisons, or evidence that Particle-1 consistently completes work with fewer tokens. It gives no workload-specific latency or reliability guarantee in the supplied material. Pricing and limits are listed, but terms for cached reads and any other provider-side conditions are not fully explained here. The page also does not establish that all listed API formats expose identical behavior, or that the model is available outside the gateway. Treat capabilities and efficiency as provider descriptions until tested against real tasks.
Skill to build now
Run a small, controlled coding evaluation before routing production work. Select representative tasks, keep prompts and tool permissions constant, and record completion rate, input and output tokens, review time, and failures for Particle-1 and the current baseline. Use the listed model ID and calculate cost from the published token prices; keep any cache treatment explicit. This turns the “fewer tokens” positioning into a workload-specific result rather than an assumption.
Reader takeaway
Particle-1’s Vercel AI Gateway listing makes the model easier to try through a documented API route and publishes useful limits and prices. It does not show that the model is more capable or cheaper on your coding work. Treat this as an integration update: test the model on representative tasks before changing routing or production defaults.