NeverMetered vs cheap per-token APIs: when flat rate wins
Updated 10 November 2026
Pay-per-token hosts for open models can be remarkably cheap, and a flat rate doesn't always beat them. This page works out when it does. All the dollar figures come from the same rate cards and arithmetic as our cost calculator, so you can plug in your own numbers there.
The assumptions behind every number here
Coding agents resend the conversation on every step, so most of what they send is repeated input that a provider can serve from cache. Standard Compute measured one gateway's traffic from 13 September to 12 October 2026: a median active day of 8.5M tokens (94 requests) and a 90th-percentile day of 86.7M tokens (491 requests), with 87% of input served from cache. Their sample was OpenCode, Kilo and Roo traffic, not Claude Code or Codex, and it isn't universal. We use 87% cache hits, 4% of tokens as output (our own assumption, since agents read far more than they write) and 22 active days a month. Costs are list prices for the model and ignore cache-write charges, taxes and discounts.
What it costs per month, same model and others
| Provider and model | Price per M tokens (in / cached / out) | Typical day (8.5M) | Busy day (86.7M) |
|---|---|---|---|
| DeepInfra, Qwen3.8-27B (promotional price) | $0.15 / $0.038 / $1.875 | $23 a month | $240 a month |
| Novita, Qwen3.8-27B | $0.42 / $0.085 / $3 | $46 a month | $460 a month |
| OpenRouter via Parasail, Qwen3.8-27B (list prices; the monthly figures add its 5.5% fee) | $0.24 / $0.24 / $2.2 | $63 a month | $640 a month |
| DeepSeek V4.1 Flash, off-peak (a different model) | $0.15 / $0.003 / $0.6 | $8 a month | $86 a month |
| DeepSeek V4 Pro, off-peak (a different model) | $0.66 / $0.022 / $1.98 | $34 a month | $340 a month |
| NeverMetered Solo, Qwen3.8-27B (1 stream) | Flat | $24 a month | $24 a month |
| NeverMetered Pro, Qwen3.8-27B (3 streams) | Flat | $67 a month | $67 a month |
The NeverMetered rows use the live plan prices (see pricing). Note what the table says. On the same model at a typical day, DeepInfra's current price comes to about $23 a month, which is about the same as Solo's $24; Novita comes to about $46 (more than Solo) and OpenRouter's route about $63 (more than Solo). DeepSeek V4.1 Flash is about $8 a month, because its cache-hit price is a fraction of a cent per million tokens and agents cache heavily. It's a different model, so that's a comparison of price, not of equivalent quality.
Where the break-even sits
| Per-token option | Cheaper than Solo ($24) up to | Cheaper than Pro ($67) up to |
|---|---|---|
| DeepInfra, Qwen3.8-27B | 8.7M tokens a day | 24M tokens a day |
| Novita, Qwen3.8-27B | 4.5M tokens a day | 13M tokens a day |
| OpenRouter, Qwen3.8-27B | 3.2M tokens a day | 9.1M tokens a day |
| DeepSeek V4.1 Flash | 24M tokens a day | 67M tokens a day |
| DeepSeek V4 Pro | 6.1M tokens a day | 17M tokens a day |
So if your agent processes fewer tokens a day than those figures, the per-token option is cheaper; above them, the flat plan is. For the same model at DeepInfra's price, the crossover against Solo is around 8.7M tokens a day, which is roughly a typical measured day. A busy day of 86.7M tokens would cost about $240 a month at DeepInfra and $460 at Novita, against $67 for Pro. The break-even moves with your cache-hit rate and output share, which is why the calculator lets you change them.
What the per-token price doesn't capture
- DeepInfra's price is currently promotional. The page labels it a 25% discount with no end date we could find, so treat it as likely to change.
- Together and Fireworks aren't in the table because neither publishes a serverless per-token price for Qwen3.8-27B (Together lists dedicated endpoints, contact for pricing).
- DeepSeek's off-peak rates are half its peak rates. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays, which overlaps working hours in some time zones; peak costs double.
- Bills vary. A runaway agent loop on a per-token plan is a larger bill, not a stalled agent. A flat plan caps the cost and queues instead.
- Rate limits and concurrency differ by provider and can throttle parallel agents; on NeverMetered the parallel-stream count is the limit by design.
- Not every host is the same quality. The same model can be served at different precision. NeverMetered serves a 4/5-bit quantized build, and its quality evals are in progress.
Who should pick which
| If you are... | Pick |
|---|---|
| Running light or occasional agent work | A per-token host. You pay for little, and DeepSeek or DeepInfra can cost a few dollars a month |
| Running a typical day, one agent | Close call on the same model. Choose on bill predictability and setup, not price |
| Running heavy days or several agents at once | A flat plan: cost stops rising, and you're buying parallel capacity |
| Wanting a bill you can budget to the dollar | A flat plan |
| Needing a larger or stronger model | A frontier API or DeepSeek V4 Pro: NeverMetered only serves Qwen3.8-27B |
Is a per-token API cheaper than NeverMetered?
Why is DeepSeek so cheap?
Does OpenRouter cost more than going direct?
What if my usage is spiky?
Where can I try my own numbers?
- DeepInfra: Qwen3.8-27B pricing (checked 10 November 2026)
- Novita: Qwen3.8-27B model page (checked 10 November 2026)
- OpenRouter: Qwen3.8-27B providers and prices (checked 10 November 2026)
- OpenRouter: pricing and fees (checked 10 November 2026)
- DeepSeek API pricing (checked 10 November 2026)
- Standard Compute: AI coding agent cost per developer (checked 10 November 2026)
- NeverMetered pricing and plan limits (checked 10 November 2026)