Particle announced Particle-1 on October 28, 2026, as a coding-focused model post-trained on DeepSeek V4.1 Flash. The company’s main claim is lower token use while preserving task performance. An October 31 evaluation from Plotly provides an additional check on how that efficiency pitch transfers to a different workload.
The practical answer is narrower than a universal efficiency breakthrough: hosted access exists, while workload-specific savings still need verification. This article reviews published evidence, not firsthand Nerova testing.
What actually became available on October 28
Announcement and API availability are not the same as downloadable weights. OpenRouter dates Particle-1's release to October 28 and currently lists Particle AI as an inference provider. That supports treating it as an accessible hosted model, not merely an announcement of future access.
Particle's documentation supplies an OpenAI-compatible client example using the model identifier particle-1 and its hosted API endpoint. Developers can therefore evaluate it inside an existing agent harness rather than wait for a dedicated application. That compatibility should be treated as a starting point, not proof that every harness feature behaves identically.
The reviewed release materials do not establish a separate preview-to-general-availability timetable. Nor should the open-weight base model be confused with an open release of Particle's derivative: Plotly explicitly lists Particle-1's weights as unavailable. As of November 1, these sources establish hosted access, not a downloadable Particle-1 checkpoint.
What the token-efficiency numbers actually measure
Particle reports 28.4%–52.2% fewer output tokens and 34.4%–51.6% lower average task costs versus its base model across Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1. It says first-attempt pass rates fall within the base model's confidence intervals. These are vendor-reported results, not proof of identical performance.
The setup matters: Terminus 2, maximum reasoning, a 512,000-token budget and 16 rollouts per task. Cost calculations assume a fixed 95% cache-hit rate; averages include failed attempts. Those assumptions are part of the claim, not incidental implementation details.
Our interpretation: compare complete accepted changes, not just generated text. A shorter run that leaves a regression for a developer to repair is not automatically an operational saving.
Plotly's October 31 test found a different trade-off
Plotly's firsthand analytics evaluation provides a useful counterweight. Particle-1 passed 40 of 43 questions, compared with 36 for the base model, but generated approximately 6,900 output tokens per question versus 6,200. It also made more model calls. Here, stronger answers did not mean less generated text.
That test also found practical savings: Particle-1 had a median completion time of 16 seconds per question versus 21 seconds for the DeepSeek baseline, and an estimated cost of $0.35 per run versus $0.46. The baseline was evaluated in September, and the reported costs reflect historical prices. More output tokens therefore did not prevent lower elapsed time and cost in this particular comparison.
Plotly cautions that each result represents one run; repeat evaluations of other models moved by as many as three points. This does not invalidate Particle's coding results: analytics is a different workload. It does rule out assuming that coding-oriented token savings automatically transfer to every business agent.
The adoption question is accepted work per budget
Our recommendation for teams evaluating this release is a controlled coding pilot, not an immediate wholesale replacement:
- Hold the workflow constant. Compare the same repository tasks, tools, tests and reasoning settings against the existing model.
- Separate three measurements. Record output tokens, total billed usage and elapsed completion time independently.
- Count recovery work. Include failed attempts, retries and human repair time before declaring a saving.
- Define acceptance externally. Require tests and reviewer approval rather than trusting the agent's completion message.
That evaluation would answer the business question the announcement cannot settle: whether a more economical coding trajectory produces the same deployable result in your environment. Until then, treat Particle-1 as a candidate for measured adoption, not a guaranteed percentage reduction in the engineering budget.