DeepSeek API Pricing (November 2026)

DeepSeek exposes an OpenAI-compatible base URL at https://api.deepseek.com. Create a key in the DeepSeek platform, keep it on the server, and send it as a Bearer token. This request uses the current Flash model name rather than the retired deepseek-v4-flash or deepseek-chat aliases:

curl https://api.deepseek.com/chat/completions \
  -H 'Content-Type: application/json' \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{"model":"deepseek-flash","messages":[{"role":"user","content":"Reply with OK"}]}'

A successful response includes token usage. Do not place the API key in browser code, logs, or a public repository. The LLM API pricing calculator can then turn measured input, output, and cache-hit tokens into a monthly estimate.

V4 Pro launched on April 24, 2026 at $0.435 per million input tokens and $0.87 per million output tokens, one flat rate at every hour. From 16:00 UTC on August 16, DeepSeek moved it to a peak schedule of $1.32 input and $3.96 output, with off-peak hours at half. That is about 3x the old input rate and 4.6x the old output rate at peak. Even off-peak, at $0.66 and $1.98, it costs more than it did before the change. The September 10 launch of V4.1 Flash, at $0.30/$1.20 peak, then gave DeepSeek a cheaper default again.

DeepSeek's September 10, 2026 pricing table lists peak and off-peak rates for cache-hit input, cache-miss input, and output on deepseek-flash and deepseek-v4-pro. 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. The published rates remain the billing baseline: the V4.1 Flash and V4 Pro rows on this page store the peak rate so an estimate never understates the bill.

Its pricing page, re-read October 29, 2026, lists deepseek-v4-pro at $1.32 cache-miss input, $0.044 cache-hit input, and $3.96 output at peak, and $0.66, $0.022, and $1.98 off-peak. The table, structured price data, and calculator use those peak rates. The calculator does not apply the off-peak discount. Halve the result for work that runs entirely off-peak.

DeepSeek now sells web Search and Fetch as standalone APIs at $0.75 per 1,000 successful requests each. When you use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit those APIs. They are free until 00:00 Beijing time on November 1, 2026 (16:00 UTC on October 31), when per-request billing starts.

That changes the bill for agent work in a way the token table doesn't show. A coding agent that runs 2,000 searches a day adds $1.50 a day at that rate, before any tokens. Check your agent's tool-call logs for search and fetch requests before November 1, and set a spend limit if you can't cap them. The charge comes out of the same prepaid balance as token usage. These rates are from DeepSeek's pricing page, read October 29, 2026.

DeepSeek prices input tokens in two lanes: cache miss (the standard rate) and cache hit, billed when the service can reuse a prompt prefix. A V4.1 Flash cache hit costs $0.006 per million tokens at peak and $0.003 off-peak. V4 Pro's cache-hit input is $0.044 at peak and $0.022 off-peak. Those are explicit provider rates, not a percentage inferred from the standard input price. DeepSeek says V4.1 Flash's smaller KV cache is what let it cut prices, and cache-hit charges are often the largest share of an agent bill.

Agents, RAG pipelines, and chat applications can benefit when they resend the same system prompt or document context. But the table does not promise a hit rate: a workload with mostly new prefixes pays the cache-miss rate. Enter your measured cached share in the estimator above. DeepSeek lists no separate Batch API tier.

The response exposes prompt_cache_hit_tokens and prompt_cache_miss_tokens in its usage object. Log those two fields and calculate the real hit share before budgeting. DeepSeek says unused cache entries are usually cleared after hours to days, so a one-time warm-up does not establish a permanent discount.

DeepSeek currently lists account-level concurrency limits of 500 for V4 Pro and 2,500 for V4.1 Flash. A request occupies one slot until its response completes. Requests beyond the account limit receive HTTP 429, regardless of how many API keys the account uses.

Production clients should cap concurrency, retry 429 responses with jittered backoff, and keep a fallback path for availability-sensitive work. DeepSeek also documents empty-line or SSE keep-alive traffic while a request waits; custom HTTP parsers must ignore those keep-alives without treating them as the JSON result.

V4.1 Flash ($0.30/$1.20 peak) is now the default for new work. DeepSeek's launch table puts it ahead of V4 Pro on every agentic coding, cyber, and tool-use row it reports, it reads images natively, and it keeps the 1M-token context window. V4 Pro ($1.32/$3.96 peak) is the previous-generation flagship with separate reasoning-effort rows in the current Chinese model rankings. DeepSeek cancelled the September 14 reroute and still serves it, with no end date published, so keep it only for integrations that depend on it and plan the move to V4.1 Flash.

Older V3.2 and R1 rows remain in the registry for historical comparison. DeepSeek retired the legacy deepseek-chat and deepseek-reasoner aliases after July 24, 2026 at 15:59 UTC, and retired deepseek-v4-flash on September 10, 2026. New requests must use deepseek-flash or deepseek-v4-pro. DeepSeek publishes the V4.1 Flash and V4 Pro 0813 weights under the MIT license, so both models can be self-hosted.

The current prices and the legacy rows answer different questions. Use the V4.1 Flash row for a new API integration; use the older rows only to reconstruct a historical bill or migration. DeepSeek changes the model behind an alias, so a familiar model ID does not prove that the underlying version stayed fixed.

The price table above links every row to DeepSeek's current or archived first-party evidence and gives the effective version date. BenchLM does not treat historical alias pricing as a currently callable SKU.

OpenAI's GPT-6 lineup changed this comparison. At its $0.30/$1.20 peak rate, V4.1 Flash costs about 3x as much as GPT-6 Luna ($0.10/$0.50) on input and about 2.4x as much as Luna on output. Off-peak, at half the peak rate, it still costs about 50% more than Luna on input and about 20% more than Luna on output. Against GPT-6.1 Terra ($1/$5), V4.1 Flash is still cheap: it costs about 3x less than Terra on input and about 4x less than Terra on output.

V4 Pro at $1.32/$3.96 peak costs about 21% less than GPT-6.1 Terra on output, about 60% less than GPT-6.1 Sol ($2/$10) and about 5x less than Claude Opus 5.5 ($20 output). Price alone doesn't settle it: the Western flagships still score higher on BenchLM's overall leaderboard, so compare quality per dollar on your own workload.

Compare the full tables: OpenAI API pricing and Claude API pricing, or use the current product-level DeepSeek vs Claude comparison. The Opus 5 vs V4 Pro model page keeps the benchmark-by-benchmark evidence separate.