Coding agents have a billing problem that nobody warns you about. They do not send one prompt. They send your repository, over and over, every turn, and each turn the meter starts again at the top of the file tree. The cost of an hour's work is not the cost of the answers you got - it is the cost of the context you resent to get them.
We wanted a real number for that, so we instrumented five live beta sessions and let developers work on their own projects with their own tooling. No synthetic load test, no scripted prompts. Cline, agent harnesses, and hand-rolled setups, pointed at a node and left alone.
The first turn pays for all the others
In a single one-hour run, developers pushed 827 million input tokens through the node. Only 19 million of those were computed fresh. The remaining 97.8% were recognised as repeat context and served from a warm cache at a fraction of the cost of prefill.
That ratio is the whole economic story of agent coding, and it is why a shared pool behaves so differently from a reserved lane. On a shared pool your context competes for cache residency with everyone else's; a stranger pasting 800k tokens can evict yours, and your next turn pays full prefill again. On a reserved lane nothing else is scheduled against you, so the cache stays warm for the length of the session.
| Session | Length | Total tokens | From cache | Failures |
|---|---|---|---|---|
| Run 1 | 1 hr | 830M | 97.8% | 0 |
| Run 2 | 1 hr | 721M | 97.0-97.8% | 0 |
| Run 3 | 1 hr | 950M | 97.0-97.8% | 0 |
| Marathon | 3 hr | 1.93B | 96.9% (1.87B) | 0 |
What one hour looks like at full tilt
The heaviest single hour we recorded: 227 million input tokens and 439,000 output tokens, across 1,017 requests. That is roughly 63,000 input tokens per second, sustained, for sixty minutes - by one developer, on one lane, with no rate limit interrupting them.
That developer is not typical. Across the three one-hour runs the node served 830M, 721M and 950M tokens among 18, 15 and 16 active developers respectively - so a typical developer moved tens of millions of tokens in an hour. Still a number that would make a per-token dashboard uncomfortable reading.
The part that answers the question
To price it honestly we took every request and costed it at DeepSeek's own published API rates for the same model, including their cached-input discount. That is the most conservative comparison available - it gives the per-token side its best case.
| Cost for that hour | |
|---|---|
| Reserved lane, flat | $0.20 - $0.40 |
| Median developer, off-peak equivalent | $0.80 - $1.20 |
| Average developer, off-peak equivalent | $1.50 |
| Average developer, peak equivalent | $3.00 |
| Heaviest hour, peak equivalent | $2.63 - $5.30 |
The average developer drew four to seven times the lane price in off-peak API-equivalent value. The heaviest hours cleared 13-26x at peak rates. Across the three free Friday hours, the cohort collectively drew $42 of off-peak DeepSeek-equivalent usage, or $83 priced at peak.
What the testers said
Nineteen written reviews, averaging 4.8 out of 5. Every respondent said they would use a reserved lane.
"I've worked a session up to around 940k context without any hallucination. The agent behaved well - the time didn't waste."
"Zero dropped requests in Cline, and cache speeds were great during long coding sessions."
Methodology, and what this does not prove
- All figures are live telemetry measured at the serving node - request counts, token totals, cache-hit rates. Nothing is extrapolated. Sessions ran 18-24 August 2026.
- "Value drawn" prices every request at DeepSeek's published API rates for the same model, including their cached-input discount, using their own off-peak and peak windows.
- Cache-hit rates are prefix-cache hits measured at the engine: 97.0-97.8% across the one-hour runs, 96.9% in the three-hour marathon.
- The model is full-weight DeepSeek V4 Flash 0731 at 1M context. Not a distill, not a quantised cut-down.
- Developer counts overlap across sessions. Your workload will differ from all of these - that is what the next beta hour is for.