What happened
OpenRouter’s model page identifies Particle-1 as a model from Particle AI, post-trained on DeepSeek V4.1 Flash. It describes the intended work as software engineering, including terminal workflows, codebase changes, and data engineering. OpenRouter says Particle-1 was trained to complete those tasks with substantially fewer reasoning and tool-call tokens while keeping accuracy close to the base model. That is the stated design goal, rather than a quantified result on the page.
The model page gives a release date of October 28, 2026. It lists a context window of 1,048,576 tokens and says up to 131,072 completion tokens are supported. Particle-1 accepts text and images as input and returns text. OpenRouter lists function calling through tools and tool_choice, plus structured output through a JSON schema in response_format. Those interfaces matter for coding-agent harnesses that must call terminals, retrieve artifacts, or return machine-readable results.
OpenRouter lists a price of $0.05 per million input tokens and $0.50 per million output tokens, with cache reads at $0.003 per million tokens. Its page says one provider hosts the model—Particle AI—and that requests are forwarded directly, rather than being routed across alternative providers. The service describes its API as OpenAI-compatible and lists the model identifier particle-ai/particle-1. It supplies examples for streaming and non-streaming calls, as well as endpoints compatible with chat completions, responses, and messages formats.
The page also says Particle-1 supports reasoning controls, including a reasoning parameter and a reasoning-effort setting. Its operational dashboard displays throughput, latency, availability, tool-call error rate, structured-output error rate, and cache-hit rate over recent time windows. It notes that EU in-region routing is not currently available. These are useful implementation and serving details for an operator deciding whether a trial is technically possible; they are not measurements of coding-task success, review burden, or total cost of an accepted change.
What the archive adds
New Runtime’s earlier coverage on October 31 described Particle-1 as an API model post-trained from DeepSeek V4.1 Flash for software and data-engineering agents (earlier Field Note). A November 4 note recorded its listing in Vercel AI Gateway (earlier Field Note). This is a new fact on a known story: OpenRouter provides specific access, price, context, capability, and serving information.
Trend signal
Particle-1’s stated goal of reducing reasoning and tool-call tokens supports the direction New Runtime tracks in Agent economics moves to completed work: token price is not the economically meaningful unit if an agent needs extra turns, retries, or review before work is accepted. OpenRouter does not report cost per verified completed task, so this listing identifies a model built around the hypothesis without establishing that it delivers the outcome.
Why now
Particle-1 was released on October 28, 2026, and the earlier New Runtime notes document subsequent API-distribution listings. The OpenRouter listing matters at this stage because it makes the model’s interface, token prices, supported output controls, and routing constraint legible to teams already using an OpenAI-compatible gateway. That lowers the integration work required for a controlled comparison, even though it does not settle the performance claim.
Why it matters
For teams running coding agents, token efficiency matters only if it improves the cost, speed, or reliability of accepted work. Fewer reasoning and tool-call tokens may reduce part of an agent run’s bill, especially on long tool-using loops. But a cheaper run can still become expensive if it produces incorrect edits, needs retries, fails a structured-output contract, or shifts more verification and repair onto engineers.
Particle-1’s published API pricing and coding focus make it a plausible trial candidate for teams already using OpenRouter-compatible tooling. The useful comparison is not token usage in isolation, nor the nominal per-million-token rate. Hold the harness, repository tasks, tools, timeout, and acceptance criteria constant; then compare the total spend, elapsed time, tool failures, review time, and accepted-completion rate against the current production model or its DeepSeek V4.1 Flash base. Include unsuccessful runs in the result.
This distinction is particularly important for agent workflows. A model optimized to stop earlier can be valuable if it reaches the correct external outcome with fewer actions. It can be harmful if lower internal-token use merely moves work into another model call or a human repair loop. Teams therefore need a task-level ledger, not a model-level billing snapshot, before changing a routing default.
Limits and uncertainty
OpenRouter repeats Particle AI’s characterization of the model but does not disclose the training data or the post-training method beyond its relationship to DeepSeek V4.1 Flash. It does not publish a benchmark table, test set, comparison protocol, or quantitative result for the stated reduction in reasoning and tool-call tokens. Nor does the page quantify “accuracy close to its base model.”
The listed rates are API prices, not measured cost per completed task. Dashboard serving metrics are platform observations over bounded recent periods and do not demonstrate coding quality or reliability on a representative production workload. The model has one listed provider, which may constrain redundancy, and OpenRouter says EU in-region routing is unavailable. Independent tests of task completion, tool behavior, and real-world review effort remain necessary.
Skill to build now
Build a small evaluation ledger for coding-agent changes this week. Select 10 to 20 representative tasks with an objective acceptance check: a passing test suite, a reviewed patch, a validated migration, or a reproducible data result. Run the same harness and tools with Particle-1 and a baseline, record input, output, reasoning, and tool-call tokens, then add retries, wall-clock time, human review minutes, and pass or fail. Calculate cost per accepted task rather than announcing a winner from a single successful run.
Reader takeaway
Particle-1 is an API-accessible coding model with clear listed prices, a large context window, and a stated focus on reducing reasoning and tool-call tokens. That makes it straightforward to test, not proven to be cheaper in practice. For coding agents, decide with accepted task quality, total run cost, and verification effort—not token volume alone.
Open question
Can Particle-1 reduce total cost per accepted coding task against DeepSeek V4.1 Flash or established coding-agent defaults once retries, tool failures, and human review are counted?