DeepSeek API Pricing: V4.1 Flash and V4 Pro

DeepSeek API pricing still looks deceptively cheap if you land on an old V4 Flash rate. I checked the live rate card on November 3, 2026, then calculated the DeepSeek API cost of the same coding workload against Claude and GPT. I build octomind, so the numbers below come from vendor pages and a published benchmark, not a product pitch.

The short answer: use deepseek-flash for the low-cost current model. At peak list rates, a 20,000-input/2,000-output turn costs $0.0084 without cache hits, or $0.00252 when all input is cached. deepseek-v4-pro costs more; it beats GPT-6.1 Terra on this workload once about 66.3% of input tokens hit cache during peak hours. Rates below were checked November 3, 2026.

What is DeepSeek V4 API pricing in November 2026?

The official API lists two models: deepseek-flash and deepseek-v4-pro. Both have a 1-million-token context window and a maximum output of 384,000 tokens; Flash supports vision, Pro does not. The cache-hit, cache-miss, and output prices are separate, and each has peak and off-peak rates. DeepSeek Models & Pricing (2026)

DeepSeek’s current alias is deepseek-flash. The older deepseek-v4-flash name is still accepted, but DeepSeek says it routes to V4.1 Flash at the Flash price. That is different from the older model row in Octomind’s catalog snapshot, which was already marked retired on November 1 and has different historical rates; use the vendor card for a current API quote. DeepSeek Models & Pricing (2026)

The current rate card lists V4-Pro-0813 and separate Pro rates, while DeepSeek’s September release announcement (2026) said deepseek-v4-pro requests would route to V4.1 Flash at Flash rates until V4.1 Pro launched. I use the current rate card below, but confirm the model and billed rate in your own response and invoice before setting a Pro budget.

When do cache-hit and off-peak prices apply?

Cache-hit and cache-miss input tokens are priced separately. Off-peak rates are half of peak rates. DeepSeek defines peak as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Weekends and those holidays are off-peak all day. DeepSeek Models & Pricing (2026)

For a quick estimate, multiply each token bucket by its own rate:

cost = cache_hit_input / 1,000,000 × cache_hit_rate
     + cache_miss_input / 1,000,000 × cache_miss_rate
     + output_tokens / 1,000,000 × output_rate

Do not apply the cache-hit price to the whole prompt just because an agent repeats its system prompt. DeepSeek charges Search and Fetch separately at $0.75 per 1,000 successful requests; those tool charges are outside the token examples below. DeepSeek Models & Pricing (2026)

What does a coding turn cost against Claude and GPT?

With a fixed 20,000 input tokens and 2,000 output tokens, V4.1 Flash is cheaper than the selected Claude and GPT models even with no cache hits. V4 Pro is also below Claude Sonnet 5.5 at these list rates, but against GPT-6.1 Terra its peak-hour result depends on cache share. The table uses each provider’s standard API token prices checked November 3, 2026: Anthropic’s Sonnet 5.5 pricing (2026) and OpenAI’s GPT-6.1 Terra pricing (2026).

For Claude and GPT, the cached-input columns use their published cache-read rates ($0.10 and $0.05 per million respectively); cache writes are excluded. This is a token-cost comparison, not a claim that the models produce equivalent fixes. It also assumes every turn has the stated 20K/2K usage, even though a real agent’s context and output vary from turn to turn.

We have one older first-party datapoint, not a current-model benchmark: on 50 real pull-request tasks, V4 Flash solved 45/50 in octomind and 43/50 in OpenCode, at about $0.032 and $0.031 per case. That August run used the retired V4 Flash model, so it does not establish V4.1 Flash quality or predict the cost of this fixed-size example.

See the full DeepSeek V4 Flash benchmark and its limits. For the larger question of when to route turns across models, see DeepSeek model routing; this post owns current DeepSeek API pricing.

At what cache-hit share does DeepSeek V4 Pro pricing beat GPT-6.1 Terra?

At peak rates, V4 Pro becomes cheaper than GPT-6.1 Terra when roughly 66% of the 20,000 input tokens are cache hits. At off-peak rates, V4 Pro is already cheaper at zero hits. V4.1 Flash is cheaper than GPT and Claude even at zero hits in either schedule.

For peak hours, the fixed-turn equations are:

V4 Pro:        $0.03432 - ($0.02552 × cache-hit share)
GPT-6.1 Terra: $0.03000 - ($0.01900 × cache-hit share)

Set them equal: $0.00432 ÷ $0.00652 = 0.6626, or about 66.3%. This assumes cached input costs the listed hit rate and ignores cache-write fees because the DeepSeek rate card lists no separate cache-write charge. If your agent starts with a short-lived context or changes its prompt prefix each turn, measure actual cache hits before choosing Pro for price reasons.

If you only need the cheapest current DeepSeek model, choose Flash: at peak it costs $0.0084 for this turn with no cache hits, compared with $0.03 for GPT-6.1 Terra and $0.06 for Claude Sonnet 5.5. Pro is the higher-priced route when you specifically want that endpoint; the rate comparison alone says nothing about which model will solve your task better. The catalog’s comparison of V4 Flash and V4 Pro uses older model rows, so treat it as historical context. The LLM API pricing comparison covers other model families.

Does DeepSeek cost the same through OpenRouter, Together, and Fireworks?

No. Together and Fireworks list V4.1 Flash at $0.30 per million uncached input, $0.006 cached input, and $1.20 output, matching DeepSeek’s peak rates but not its off-peak schedule.

OpenRouter exposes a provider list with a much wider range; the row with the lowest input price on November 3 was Open Inference at $0.03325 input, $0.00475 cache read, and $0.475 output per million. Other rows were lower on output or cache read, so the cheapest provider depends on your token mix. Those rates are provider-specific and can change, so a gateway route is not one fixed DeepSeek price. Together’s pricing page (2026), Fireworks’ pricing page (2026), and OpenRouter’s V4.1 Flash provider table (2026).

For the example’s 100 turns, multiply the zero-cache column by 100: $0.84 direct at peak, $0.42 direct off-peak, $0.84 through Together or Fireworks Standard, $0.588 on Fireworks Spot, and about $0.1615 using the Open Inference rates shown on OpenRouter. At 100% cache hits the provider rates change the result, so inspect the row’s cache-read price too.

OpenRouter’s standard routing can choose among providers; pin or exclude providers if you need predictable pricing or behavior. The live page listed 32 providers when checked.

Which DeepSeek endpoint should you choose?

Choose deepseek-flash when the priority is lower token spend and vision input matters. Schedule flexible work off-peak if you control timing: the direct API’s whole rate card is half price then. Compare V4 Pro with GPT only after measuring cache-hit share; under this example’s peak rates, the crossover is about 66%.

For a third-party host, compare the exact provider row, cache-read rate, and service tier before selecting it. A cheap-looking OpenRouter row is not a promise that automatic routing will always use that provider, while Together and Fireworks list the same standard V4.1 Flash rates as DeepSeek’s peak schedule. The host can still be the right pick for its routing, tooling, or service setup; this table only compares token charges.

Treat this DeepSeek pricing as checked rates rather than permanent prices. DeepSeek says prices can change, and third-party provider rates or promotions can move independently. Recheck the official rate card before setting budgets or shipping a production model ID.