Cheap tokens are the point

This service resells nothing and marks up nothing, so its existence hangs on a single number: what a token costs. That number has quietly become one of the widest price spreads in software. As of September 2026, a million output tokens from deepseek-flash (DeepSeek-V4.1-Flash) cost $0.60 off-peak and $1.20 at peak. The same million from a frontier flagship cost $25 to $50 — $180 from the pro tiers, at the list prices read on 2026-08-06. The models are not equivalent, but on everyday work they are far closer in capability than they are in price.

The rate cards

DeepSeek's card today, USD per million tokens, read from its pricing page on 2026-09-18. Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday; every other hour is off-peak.

The cross-vendor snapshot below was read on 2026-08-06, against each vendor's own pricing page. DeepSeek sold one flat card then, and its two rows are that card, kept as history: V4 Flash was retired on 2026-09-10 and the flat card ended on 2026-08-16.

*Introductory pricing through 2026-08-31; $3 / $15 after. Sources: DeepSeek, Anthropic, OpenAI. Prices change monthly; when this table and a vendor's page disagree, the vendor's page is right.

Read today's flash card against that column: flash to Fable 5 is 83× on output off-peak, 42× at peak. The cached-input column is starker — $1.00 against $0.003 is 333× — and DeepSeek's cache is automatic and free to write, where Anthropic bills cache writes at 1.25–2× the input rate.

The same task

Rate cards mislead without a workload, so take an ordinary agent exchange — the shape Anthropic itself uses as a worked example: 50k input tokens of which 40k are cache reads, 15k output tokens. Counting the same nominal tokens at deepseek-flash's card today, off-peak, and at the other vendors' list prices of 2026-08-06:

Not twenty percent cheaper. Seventeen to a few hundred times cheaper off-peak, half that at peak, depending on the model and how much of the prompt caches. *Sonnet 5 at its introductory price; the table on the V4 flat card, which billed $0.0057 for this exchange, read 31× to 735×.

That table is still arithmetic on a rate card, which is the weakest kind of cost claim — it assumes every model spends the same tokens on the same work, and they do not. The section below is the stronger evidence: what independent benchmarks actually paid.

What the benchmarks actually paid

The figures in this section and the next measured V4 Flash and V4 Pro, read on 2026-08-07. V4 Flash was retired on 2026-09-10; deepseek-flash is V4.1 Flash, a different model at a different price, and none of these rows describe it. They stay as dated V4 measurements until independent ones exist for V4.1.

Two leaderboards publish a dollar figure next to the score, measured from their own runs rather than multiplied out of a price list. Artificial Analysis reports cost per task — total spend to run its Intelligence Index divided by the number of tasks — alongside the index score itself:

The same source publishes what the whole index cost to run per model, which is the least ambiguous number on this page: $72.03 for deepseek-v4-flash against $3,836.05 for Claude Opus 5 and $5,455.22 for Claude Fable 5 — 53× and 76× on identical work.

On coding specifically, SWE-rebench runs one fixed harness over 111 real GitHub issues and publishes cost per problem next to the resolve rate:

And one independent developer ran the same build prompt three times back to back in early August 2026 and published the receipts: Claude Opus 5 $0.3183, Kimi K3 $0.2485, deepseek-v4-flash $0.0049. All three produced working output; DeepSeek failed silently on image handling where Claude improvised around bugs that stopped the others.

Where the cheap-token argument breaks

Two numbers in those tables cut against the story this page is telling, and leaving them out would make it propaganda. On Artificial Analysis, GPT-5.6 Luna scores the same 52 as flash for $0.05 against flash's $0.03 — a 1.7× gap, not a hundredfold one. On SWE-rebench, Luna beats deepseek-v4-pro on both axes at once: more issues resolved, less money per issue. So the honest shape of the claim is narrower than "DeepSeek is the cheapest way to get work done":

• Against the flagship tier, the gap is real and enormous. One to two orders of magnitude per task, on published same-harness runs, confirmed by two independent measurers.
• Against the cheap frontier tier, the gap mostly evaporates. The cheap tiers of the big labs now land within a factor of two of DeepSeek's per-task cost, sometimes with better scores.
• The top of the quality range was not for sale at this price. V4 Flash's 52 sat ten-plus index points below Opus 5 and Fable 5, and Artificial Analysis pairs that $0.03 with a 37% score on Humanity's Last Exam and an 84% hallucination rate. Paying 76× more buys something; whether it buys enough for your task is your measurement to make, not ours.

Every figure above was read on 2026-08-07 and they drift for two reasons at once: vendors reprice (DeepSeek's own cost-to-run fell from $1,071 to $176 for v4-pro after a price cut), and benchmarks reweight their indices. Treat the ratios as this month's, not as constants.

What this does not claim

Three smaller caveats on top of the big one. Tokenizers differ, so identical text is not identical token counts across vendors — Anthropic notes its current tokenizer emits roughly 30% more tokens for the same input than its previous one. DeepSeek bills peak hours at 2× the off-peak rates since 2026-08-16 (01:00–04:00 and 06:00–10:00 UTC, Monday to Friday), which halves the gap during working hours in Beijing. And a harness costs money too: one comparison held V4 Flash constant across four agent frameworks and watched cost per successful task swing 2.7× — $0.073 on the cheapest, $0.195 on Claude Code — so the tool you drive the model with can matter as much as the model.

What survives all of that is still worth saying plainly: for the broad middle of real work — summarize, translate, refactor, answer, glue — the going rate differs by one to two orders of magnitude depending on whose API you call, and the cheapest credible option is no longer a toy.

Why it matters

Chat is measured in thousands of tokens; agents are measured in millions. The moment a model works unattended — reading files, retrying, checking its own output — token consumption stops tracking human attention and starts tracking machine patience. An overnight agent run that emits ten million output tokens costs $6 at flash's off-peak price and $500 at Fable prices. One of those is "leave it running"; the other is a line item that gets a meeting. At frontier prices, autonomy is a luxury good. At flash prices, it is a background process.

Whatever AGI turns out to be, it will be made of tokens, and nobody runs civilization-scale inference at $50 per million. Every 10× drop in token price makes a class of applications viable that was silly the day before — the same way compute-per-dollar curves, not any single breakthrough, decided what software got built. Cheap tokens are not the budget option. They are the substrate.

This page is also the explanation of the gateway you are reading it on. At flash prices, a dollar buys roughly nineteen hundred ordinary conversational turns off-peak, half that at peak; at frontier list prices, the same dollar buys about fifty. A free tier funded by donated keys and pocket money is arithmetic that only works at the bottom of that table — which is why it runs on deepseek-flash, and why there is no paid tier to upsell you to.