Particle-1 uses 38% fewer tokens than DeepSeek V4.1 Flash, but its launch chart priced it at DeepSeek's rates. At Particle's own $0.25/$1 card, calling DeepSeek directly is 21% cheaper per task averaged over the week.
Particle AI's first model is a small fine-tune of DeepSeek V4.1 Flash that does the same coding work with far fewer tokens. The token savings hold up. The cost savings in the launch chart only hold if you pay DeepSeek's prices, and Particle doesn't charge DeepSeek's prices.
Image source: Particle AI
TL;DR
- Particle-1 costs $0.25 in, $0.03 cached and $1 out, half that from 20:30 to 04:00 New York time. That's the same card Particle charges for plain DeepSeek V4.1 Flash, and it's above DeepSeek's own peak price on cached input.
- Across Particle's four coding benchmarks it writes 38% fewer output tokens in total than the base model and scores 1.6 points lower on average. At Particle's own prices the mean task costs $0.142 at peak, not the $0.095 in the chart.
- Particle's peak covers the US working day, DeepSeek's covers the US evening. Run it overnight US time and Particle-1 wins. Run it during the US day and DeepSeek direct costs about 45% less per task.
What Particle charges
Particle-1 went live on October 28 alongside Particle's launch post. The rates come from the billing docs, which put it on the same card as Particle's hosted DeepSeek V4.1 Flash, with a note that Particle-1 has "independent price rules". Here it is next to DeepSeek's own API.
| Per 1M tokens | Input | Cached | Output | Peak hours (New York) |
|---|---|---|---|---|
| Particle-1, peak | $0.25 | $0.03 | $1.00 | 04:00-20:30 daily |
| Particle-1, off-peak | $0.125 | $0.015 | $0.50 | |
| DeepSeek V4.1 Flash, peak | $0.30 | $0.006 | $1.20 | 20:00-23:00, 01:00-05:00 weekdays |
| DeepSeek V4.1 Flash, off-peak | $0.15 | $0.003 | $0.60 | |
| GLM-5.3-Flash (Z.ai) | $0.15 | $0.03 | $0.50 | none |
Read November 2, 2026. DeepSeek's peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays, shown here in New York time now that US daylight saving has ended.
Input and output come in about 17% under DeepSeek's peak rates. Cached input costs five times as much, and that line matters more than it looks. An agent re-reads its whole context every turn, so most of what it pays for is cache hits.
A few more terms from the docs. Paying accounts can send requests on the flex tier, which waits behind standard traffic and takes another 50% off, on top of off-peak. There's no batch API and no cache write fee. Web search is $2.50 per 1,000 calls. According to the model page, the context is 1,048,576 tokens, output caps at 131,072, and reasoning defaults to high unless you set it to none.
One listing doesn't match. OpenRouter's Particle-1 page shows $0.05 in, $0.003 cached and $0.50 out, with Particle as the only provider and no discount flag. That is a fifth of the docs price on input. We don't know if it's a promotion or a mistake, so we leave it out of the math below and note it at the end.
The launch chart, repriced
Particle's headline numbers are cost per task on four agentic coding suites, run with the Terminus 2 agent at max reasoning. The footnote says both Particle-1 and DeepSeek were priced at $0.30 / $0.006 / $1.20, DeepSeek's peak card, with a fixed 95% cache hit rate. So the chart shows what Particle-1 would cost if DeepSeek sold it. It is a fair way to isolate the token savings, and the footnote is upfront about it. It is not what you pay.
Because Particle published its formula, output tokens and the resulting cost, the input tokens fall out of the arithmetic. On DeepSWE that is about 7.2 million input tokens per task for Particle-1 against 11.4 million for DeepSeek, so the fine-tune cuts input as well as output. We then repriced each task at both companies' real rates.
| Benchmark | Pass rate P-1 / DS | Chart P-1 / DS | Peak P-1 / DS | Off-peak P-1 / DS |
|---|---|---|---|---|
| DeepSWE 1.1 | 72.6% / 71.2% | $0.221 / $0.337 | $0.356 / $0.337 | $0.178 / $0.168 |
| Terminal-Bench 2.1 | 79.9% / 81.9% | $0.049 / $0.077 | $0.063 / $0.077 | $0.031 / $0.038 |
| ReactBench | 31.5% / 33.7% | $0.095 / $0.181 | $0.131 / $0.181 | $0.065 / $0.090 |
| ADE Bench | 64.5% / 68.0% | $0.016 / $0.034 | $0.019 / $0.034 | $0.009 / $0.017 |
| Mean of four | 62.1% / 63.7% | $0.095 / $0.157 | $0.142 / $0.157 | $0.071 / $0.078 |
Cost per task in USD, failures included. P-1 = Particle-1 on Particle's API, DS = DeepSeek V4.1 Flash on DeepSeek's API. Bold marks a benchmark where Particle-1 costs more. Input tokens are our back-calculation from Particle's published figures at its stated 95% cache-hit rate.
Compare peak with peak and the 39% saving in the chart becomes 9%. On DeepSWE, the longest and most input-heavy suite, Particle-1 at $0.356 costs more than DeepSeek at $0.337, because 7 million cached tokens at $0.03 outweigh everything the fine-tune saves on output.
Accuracy moves the other way, slightly. Particle-1 is ahead on DeepSWE and behind on the other three, by 1.9 to 3.5 points. Particle says pass@1 stays within the base model's confidence intervals, and GPQA Diamond barely moves (90.4% against 90.9%). No independent lab has run it yet; Artificial Analysis has no Particle-1 page. The data behind the launch chart also tags its token and accuracy fields "user-supplied-simulated". That may just be a label from the charting tool, but it's one more reason to wait for outside numbers before leaning on the pass rates.
There is one comparison where Particle-1 wins outright. If you already call DeepSeek V4.1 Flash through Particle, you pay $0.25 / $0.03 / $1 for it too. Switching that traffic to Particle-1 cuts the mean task from $0.236 to $0.142, 40% off for one changed model string.
Which one is cheaper, hour by hour
The two discount windows barely overlap. Particle is at full price for 16.5 hours a day, every day, centered on the US working day. DeepSeek is at full price for 7 hours on weekdays, and those hours fall in the US evening and small hours, which is when Particle is half price. Using the mean task above, here is which API is cheaper in each half hour.
Weekdays
Weekends
New York time (EST). DeepSeek's weekdays run on UTC, so its Monday 01:00 UTC peak lands on Sunday evening in New York and Friday evening is already Saturday in UTC, so off-peak; the chart ignores both edges. DeepSeek also drops its peak on Chinese public holidays.
On weekdays the line sits almost exactly at 05:00 and 20:00. From 05:00 to 20:00 New York time, DeepSeek is off-peak and Particle-1 isn't, so a task costs $0.078 on DeepSeek and $0.142 on Particle. At night it flips: Particle-1 drops to $0.071 while DeepSeek sits at its $0.157 peak for most of the window.
On weekends DeepSeek has no peak, so it wins from 04:00 to 20:30. Overnight, Particle-1's discount still puts it about 9% ahead, $0.071 against $0.078. Spread one task evenly over every hour of the week and the averages come out at $0.120 for Particle-1 and $0.095 for DeepSeek. That is where the 21% in the headline comes from.
An agent month at a lower cache rate
The 95% cache rate in Particle's chart punishes Particle's expensive cached input. Our usual workload is gentler: 1 billion input tokens a month with 85% cache hits, plus 50 million output. For Particle-1 we shrink those counts by the DeepSWE ratios backed out above, 37% less input and 28% less output, since DeepSWE is the suite closest to a real repo agent.
| When you run it | DeepSeek direct | Particle-1 |
|---|---|---|
| US daytime only | $55.05 | $75.57 |
| US weekday nights only | $110.10 | $37.78 |
| Evenly, all week | $66.52 | $63.76 |
DeepSeek: 1B input, 85% cached, 50M output. Particle-1: about 632M input, 85% cached, 35.8M output. "Nights" assumes DeepSeek is at peak for the whole window, which overstates it by a couple of hours. No flex tier.
At 85% cache hits the all-week totals land within $3 of each other, with Particle-1 slightly ahead. The real decision is scheduling. A team whose agents run while it works, in US hours, pays about $20.50 a month more on Particle for this workload, about 37%. Overnight batch work goes the other way by a lot.
Two things could move this. If OpenRouter's $0.05 / $0.50 listing is real, the same month costs $24.24 there at any hour, cheaper than anything in the table. And flex halves Particle's night price again, to $18.89, if your account qualifies and you can tolerate the queue.
Who should switch
Anyone already sending DeepSeek V4.1 Flash traffic through Particle should move it to Particle-1 today. It's the same price per token, with 40% less spent per task.
If you call DeepSeek directly, the answer depends on your clock. Nightly evals, backfills and CI agents that run after 20:00 New York time are cheaper on Particle-1. Interactive coding during the US day is cheaper on DeepSeek, and the fewer tokens don't make up the difference. Shorter outputs do mean faster turns, which may be worth more to you than the cents.
We've added Particle-1 to its own model page with the peak card, next to DeepSeek V4.1 Flash. For the base model's pricing history, our V4.1 Flash launch post covers the peak window in detail. To test your own token counts, use the calculator.
Sources
- Particle AI: Introducing Particle-1 - benchmarks, output tokens, cost-per-task method, training details, hero image
- Particle docs: usage and billing - peak and off-peak rates, New York time windows, flex discount, web search fee
- Particle docs: models and service tiers - context window, output cap, reasoning defaults, flex eligibility
- OpenRouter: Particle-1 - the $0.05 / $0.003 / $0.50 listing, read November 2
- DeepSeek and Z.ai pricing pages - comparison rates and DeepSeek's peak window