AI modelsDeepSeek

DeepSeek V4

Decision matrix

DeepSeek V4 model decision matrix
Overview

DeepSeek V4 comes in two API variants, a fast Flash model and a more capable Pro model, both with a 1M-token context window, priced from well under a cent per million cached input tokens up to a few dollars per million output tokens, with lower rates during off-peak UTC hours. Pick it when raw price matters and you can schedule heavy jobs during the cheaper window.

Other good fits
  • Batch or non-urgent workloads that can be scheduled during DeepSeek's off-peak discount hours.
Pros
Read pros

DeepSeek V4 is available through the API in two variants: DeepSeek-V4.1-Flash (faster, supports vision, has both thinking and non-thinking modes) and DeepSeek-V4-Pro (more capable, text-only). Both share a 1M-token context window with up to 384K tokens of output, and both are priced low enough that cost-sensitive, high-volume applications can benefit meaningfully, especially with prompt caching and off-peak scheduling.

Pick it when…

  • Per-token cost is a real constraint and DeepSeek's rates, particularly with cache hits, are meaningfully cheaper than your current provider.
  • You want vision input on a cheap, fast model (Flash) rather than the more expensive alternative of a vision-capable flagship elsewhere.
  • Your workload can tolerate being scheduled during off-peak hours (01:00-04:00 and 06:00-10:00 UTC, weekdays) for a further discount.
Cons
Read cons

The time-of-day pricing split means the price you actually pay depends on when your application happens to call the API, not just which model you choose; for an application that must run during peak hours every time, factor in the higher rate. DeepSeek-V4-Pro doesn't support vision input, so a task needing both Pro-level text capability and image understanding isn't served by a single variant here.

Pick something else when…

  • You need vision on the more capable variant. Neither DeepSeek variant currently offers that combination; compare against a provider whose flagship model supports both.
  • You need consistent pricing regardless of time of day. A flat-rate provider avoids the off-peak/peak complexity here.
  • You want to compare DeepSeek against several other cheap models from one integration. OpenRouter makes that easy.

Avoid if

  • You need vision (image) input and are using the Pro variant; only the Flash model currently supports vision.
Context

Both Flash and Pro variants offer a 1M-token context window with up to 384K tokens of output, support JSON output and tool calls, and work with the Anthropic API format as well as DeepSeek's own. Flash supports both thinking (default) and non-thinking modes and vision input; Pro does not support vision.

Requirements

Not documented yet.

Licence & cost
Read licence & cost

Prices checked 23 September 2026 (pricing (opens in a new tab)), per million tokens (off-peak to peak range):

ModelInput (cache hit)Input (cache miss)Output
DeepSeek-V4.1-Flash$0.003–$0.006$0.15–$0.30$0.60–$1.20
DeepSeek-V4-Pro$0.022–$0.044$0.66–$1.32$1.98–$3.96

Off-peak pricing (01:00-04:00 and 06:00-10:00 UTC, weekdays) is roughly half the peak rate on every line. Cache hits are dramatically cheaper than cache misses, so an application that repeats similar prompts benefits disproportionately from DeepSeek's caching. There's no subscription; billing is pay-as-you-go per token.

First task

Documentation-based suggestion — not run here.

Send the same made-up drafting or extraction task to DeepSeek-V4.1-Flash and DeepSeek-V4-Pro, comparing quality against their price difference, then repeat the same request a second time to see the effect of prompt caching on cost for a repeated or similar prompt.

Done when you have a real cost-and-quality comparison across both variants, and you understand how much prompt caching actually saved on the repeated request.

Links & evidence
Documentation-based · not hands-on · Sep 23, 2026

Rests on DeepSeek's API pricing documentation, read on 2026-09-23. No API call was made for this review.

  • DeepSeek API pricing · Sep 23, 2026 — V4.1-Flash and V4-Pro per-million-token pricing, cache hit/miss rates, off-peak/peak windows, 1M context window.
  • OpenRouter pricing · Sep 23, 2026 — Comparison alternative for testing several low-cost models from one integration.
  • GLM API pricing · Sep 23, 2026 — Comparison alternative for a similarly low-cost model family.
Further review notes & alternatives

Alternatives and trade-offs

  • GLM wins as another low-cost alternative worth comparing on price and capability for your specific task.
  • Claude Haiku wins when consistent, time-of-day-independent pricing matters more than DeepSeek's off-peak discount.
  • OpenRouter wins when you want to compare DeepSeek against several other providers from one integration.

Learn more

Catalogue facts

Process long text prompts with a 1M-token context and thinking mode. Choose it when hosted access or published weights fit the project constraints. It is text-only, with full self-hosting requiring cluster-scale resources.

DeepSeek · paid · self-host

  • frontier model
  • large language model
  • high context window
  • open weights
  • low cost api
  • reasoning mode
  • write-code
  • open-source
  • draft-docs
  • talk-think
  • data-analytics

Browse all tools