What would your coding agent cost per token?
Enter how much your agent uses and see the monthly bill at Anthropic, OpenAI, DeepSeek and hosts of the same model NeverMetered serves, next to a flat plan. It shows where paying per token is cheaper too, because that depends on how much you use.
Pick a starting point, then adjust. Nothing here is saved or sent anywhere.
Median active day measured by Standard Compute (94 requests).
Input plus output. Your tool's usage screen or your provider's dashboard shows this.
Our assumption: agents read far more than they write. Raise it for chatty, write-heavy work.
The cheapest per-token host of the same model (DeepInfra) costs about $23 a month for this workload. NeverMetered Solo is $24 a month for 1 agent at once and NeverMetered Pro is $67 a month for 3 agents at once, with no token meter. On these settings it's close, so the deciding factors are predictability and how many agents you run at once. A flat price doesn't move with a long agent session.
One stream generating your output alone would need about 2.2 hours a day at NeverMetered's measured 42 tokens a second. Reading prompts takes extra time on top.
The same model, paid per token
Qwen3.8-27B, the model NeverMetered serves, from hosts that bill by the token. This is the like-for-like comparison.
- Qwen3.8-27BLowest in this groupDeepInfra · $0.15/M in · $0.038/M cached · $1.875/M outDeepInfra labels this a promotional 25% discount; it may end.Price source$23a month
- $46a month
- Qwen3.8-27BOpenRouter (Parasail route) · $0.24/M in · $0.24/M cached · $2.2/M out · +5.5% feeNo cached-input price is listed for this route, so cached tokens are charged as normal input. Includes OpenRouter's 5.5% credit-purchase fee.Price source$63a month
Other low-cost per-token APIs
Different models. Often very cheap when most of your input is cached.
- DeepSeek V4.1 Flash (off-peak)Lowest in this groupDeepSeek · $0.15/M in · $0.003/M cached · $0.6/M outOff-peak price. Peak hours (01:00-04:00 and 06:00-10:00 UTC on weekdays) cost double.Price source$8a month
- DeepSeek V4 Pro (off-peak)DeepSeek · $0.66/M in · $0.022/M cached · $1.98/M outOff-peak price. Peak hours cost double.Price source$34a month
Frontier models, paid per token
Different and generally stronger models, so this is a price check, not a like-for-like. Many people keep one of these for their hardest work.
- Claude Haiku 5.5Lowest in this groupAnthropic · $0.1/M in · $0.01/M cached · $0.5/M outRate for prompts up to 100K tokens; over 100K it is $0.50 in / $2.50 out.Price source$8a month
- $69a month
- $140a month
- $140a month
- $170a month
- $270a month
NeverMetered: one price, whatever the tokens
No token quota, 5-hour window or weekly cap. Plans differ mainly in how many requests run at once. Each request still has the context, output and time limits shown, and the plan's queue is finite.
- Solo1 stream (requests at once) · 2 can queuePer request: 256K context · up to 248,320 output tokens · up to 15 minutes$24a month
- Pro3 streams (requests at once) · 2 can queuePer request: 256K context · up to 248,320 output tokens · up to 15 minutes$67a month
It serves one open-weight model, Qwen3.8-27B, and is slower than the frontier APIs above. See current plans or what the model is good at.
Per-token list prices read from each provider's own page on 2026-11-10. They change often, and taxes, volume discounts and cache-write charges are not included. Subscriptions such as Claude Pro or Max and ChatGPT Plus or Pro aren't shown because they publish no token quota, so any conversion to dollars per token would be invented.
Questions about these numbers
Where do the usage numbers come from?
The typical and busy presets are measured: Standard Compute, a rival gateway, reported a median active day of 8.5 million tokens and a 90th-percentile day of 86.7 million across real coding-agent traffic (13 September to 12 October 2026), with 87% of input served from cache. That sample used OpenCode, Kilo and Roo, not Claude Code or Codex, so your own numbers may differ. The light and heavy presets are round examples.
Why aren't Claude Pro, Claude Max or ChatGPT Plus in the list?
Subscriptions don't publish a token quota, so there is no honest way to turn $20 or $100 a month into dollars per token. They're flat prices with usage limits, which is a different thing from both per-token APIs and a plan with no limits. Our guides compare them directly.
Is a cheaper per-token host always the better deal?
Not always, and not never. At light usage, a per-token host of the same model can cost less than a flat plan, and some low-cost APIs are very cheap when most input is cached. A flat price wins as usage grows, and when you want a bill that doesn't move with a long agent session or several agents running at once.
What does cached input mean here?
Coding agents resend the conversation with every step, so most of what they send was already seen on the previous step. Providers bill that repeated part at a much lower cached rate. The calculator lets you set how much of your input is cached; the default is the measured 87%.
Does NeverMetered count tokens?
No. Plans are priced by how many requests can run at once, and there is no token quota, 5-hour window or weekly cap. The Usage page shows token counts for your own visibility, and they are never billed.
More detail: what running a coding agent costs per day, how NeverMetered compares and how parallel streams work. Rates checked 2026-11-10.
Let your agents run.
No token meter and no 5-hour or weekly limits, for one flat price. Upgrade when you want more agents working at once.