Developer guides6 min read

DeepSeek V4 Pro after September 14: record the returned model

Direct answer

On September 22, 2026 the provider still lists deepseek-v4-pro as V4 Pro. Record the returned model version before calling a run a Pro-versus-Flash comparison.

Updated September 22, 2026: The September 10 note on this page, which said deepseek-v4-pro would route to V4.1 Flash on September 14 at 04:00 UTC, is withdrawn. The updates page now says the provider will keep V4 Pro API service after that date, with billing unchanged, and will give further notice if that changes. This site did not call the API.

What the current pages distinguish

Use deepseek-flash for V4.1 Flash. The pricing page still says the old names deepseek-v4-flash and deepseek-v4-flash-vision-exp are accepted, served by V4.1 Flash, and billed at the Flash price.

deepseek-v4-pro is listed separately. Its model version on that page is DeepSeek-V4-Pro-0813, not DeepSeek-V4.1-Flash. The table gives both names a 1M context length and a 384K maximum output, and it marks vision as supported for Flash and not supported for Pro. The documented concurrency limit is 2,500 for Flash and 500 for Pro. Those are page statements, not a load test.

The two names also have different list rates. The dated table is on the cost worksheet. Different prices are evidence that the provider is not billing them as one alias. They are not evidence that Pro produces more accepted changes in your repository.

Because this site did not send an authenticated request, do not treat the pricing-page version string as the version your account will return next week. Save the model metadata from the response with the date of the run.

Do not take the family name from GitHub

The MonkeyCode README lists DeepSeek among integrated model families. That sentence does not say whether a task used deepseek-v4-pro, deepseek-flash, or a hosted route whose display name differs from the API id. Write the returned identifier into the attempt log. A README family list is not a model-version check.

Decide what would justify a switch

Pick one defect that fails a test, one missing test and one documentation change. Define allowed files and acceptance checks before starting. Run Flash and Pro from separate copies of the same baseline, with the same instructions, tool access and timeout. Record the alias, the returned version and the date. Do not assume a dated release suffix in an old note is the current API id.

Retain every attempt

For each run save the input revision, model, runtime, prompt, final diff, check output, failure category and manual correction notes. Include unsuccessful runs when calculating cost per accepted change. An empty cost cell means unknown. Zero is valid only when measured as zero.

An upgrade is useful if it improves the result you need within your budget, including review effort. Three public exercises are an onboarding check, not a blind benchmark. Use previously unseen repository tasks before making a broader performance claim.

Start with the downloadable first task, record attempts in the CSV templates, and read the current API status before rerunning.

Frequently asked questions

Did deepseek-v4-pro switch to V4.1 Flash on September 14?

Not according to the provider pages checked on September 22, 2026. The updates page says V4 Pro API service continues after September 14, with billing unchanged. The pricing page lists deepseek-v4-pro as DeepSeek-V4-Pro-0813, separate from deepseek-flash.

Can two aliases still be compared as Pro and Flash?

Only if each run records the returned model version. This site did not call the API. A family name in the MonkeyCode README does not identify which alias a task used.

Do different list prices prove a quality difference?

No. The pricing page gives different dollar rates for the two names. That shows they are billed separately. It does not measure accepted changes in your repository.

Cookie settings

We use cookies only for analytics (GA4 + Matomo) to improve the docs. No ads, no tracking across sites.