Editorial correction: This page no longer repeats the July price-cut percentages, launch prices or comparative cost figures previously taken from secondary reporting. The table below is a primary list rate captured on September 22, 2026. It is not a cost per accepted task, and the provider says prices can change.
Capture the list rate before the run
On September 22, 2026, the DeepSeek pricing page listed two current API names, deepseek-flash and deepseek-v4-pro, with versions DeepSeek-V4.1-Flash and DeepSeek-V4-Pro-0813. Both are listed with a 1M context length and a 384K maximum output. Prices are dollars per 1M tokens. Off-peak is half of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Weekends and Chinese public holidays are off-peak.
| Rate per 1M tokens | deepseek-flash |
deepseek-v4-pro |
|---|---|---|
| Cache-hit input, off-peak | $0.003 | $0.022 |
| Cache-hit input, peak | $0.006 | $0.044 |
| Cache-miss input, off-peak | $0.15 | $0.66 |
| Cache-miss input, peak | $0.3 | $1.32 |
| Output, off-peak | $0.6 | $1.98 |
| Output, peak | $1.2 | $3.96 |
The same page says retired Flash aliases are still accepted, served by V4.1 Flash, and billed at the Flash price. It also lists a concurrency limit of 2,500 for Flash and 500 for Pro. That is a documented cap, not a throughput this site measured. Refresh the quote when the alias, version or price page changes. An estimate is not a receipt.
Keep the numerator and denominator honest
For a defined window, add actual model charges, tool fees and compute charges across all attempts, including retries and rejected diffs. Divide by the number of changes that pass your acceptance checks. If no change is accepted, report “no accepted result” rather than a zero cost per task. A cache-hit rate applies only to tokens the invoice marks as cache hits. Track human review minutes separately, or disclose the hourly rate when converting them into money.
What the GitHub README prices, and what it does not
The MonkeyCode README, read on September 22, 2026, does not publish a token price or a cost per accepted change. It does publish minimum machine sizes for self-hosting: the console is listed as at least 2 CPU, 4 GB memory and 40 GB disk; the development-environment host is listed as at least 8 CPU, 16 GB memory and 100 GB disk. Those are capacity floors from the project README. They are not a cloud bill, and they do not show that a given model fits on that host.
Use them as the compute side of the worksheet: record the machine you actually ran, the hours it was up, and the rate you paid. Do not substitute the README minimum for that bill. The repository is AGPL-3.0. The license does not set the model price.
Hosted MonkeyCode access and a direct DeepSeek API call are different endpoints. The maintainer confirmed hosted DeepSeek V4.1 Flash access on September 10. That confirmation is not an unlimited-use measurement, and it does not make the official API free. Write the endpoint type on every row.
Fill the columns that already exist
The public attempt template has these fields: task_id, run_id, repository_commit, monkeycode_version, model_id, endpoint_type, run_started_at_utc, accepted, retries, elapsed_seconds, review_minutes, input_tokens, output_tokens, model_cost, currency, compute_cost, failure_category, notes.
Write the returned model version into model_id, not only the alias you requested. Leave model_cost or compute_cost empty when the charge is unknown. Zero is valid only when the measured charge is zero.
Do not borrow a public exercise as a price
The first-task pack checks a workflow. Its published reference solutions make it a poor blind price comparison. Repeat the same columns on unseen repository tasks before using the figure in a purchasing decision. Three public exercises are an onboarding check, not a ranking.
Frequently asked questions
Does the MonkeyCode repository publish a price per accepted task?
No. The README checked on September 22, 2026 publishes deployment minimums and an install path. A per-task figure has to come from your invoices and the attempt log.
Are the DeepSeek list rates the cost of an accepted change?
No. The September 22 pricing page lists dollars per 1M tokens, with separate cache-hit, cache-miss, output and peak rates. Retries, rejected diffs, compute and review time sit outside that list rate.
Can a hosted free model be recorded as zero API cost?
Only for the hosted path you actually used, and only if that path charged zero. A free hosted entry does not make a direct DeepSeek API call free. Record endpoint_type so the two are not mixed.