Bottom line first
- Since 2026-08-16 16:00 UTC, DeepSeek dropped flat pricing: deepseek-v4-flash bills by UTC peak/off-peak; off-peak is exactly half peak.
- Off-peak: cache-miss input ~$0.22/M, output ~$0.66/M; peak doubles to $0.44 / $1.32. Cache-hit input off-peak: $0.007/M.
- Pre–Aug 16 flat rates were $0.14 in / $0.28 out / $0.0028 cache-hit—dead for new budgets.
- The watershed is output-token share × which UTC window, not Flash vs Pro. Thinking is on by default at the same rates.
- 1M context and OpenAI-compatible
https://api.deepseek.comease migration; peak scheduling and cache hit rate are the levers you control.
Model id is not the watershed. UTC peak/off-peak tier and output-token share are.
0. Bottom line
As of 2026-08-22, the official DeepSeek pricing page (effective 2026-08-16 16:00 UTC) splits deepseek-v4-flash (checkpoint DeepSeek-V4-Flash-0731, available since 2026-07-31) into peak/off-peak × cache-hit/miss × input/output. Base URL https://api.deepseek.com, OpenAI-compatible; 1M-token context; thinking on by default at the same rate card as non-thinking.
Working numbers (USD / 1M tokens; verify live):
- Off-peak deepseek-v4-flash: cache-hit in $0.007, cache-miss in $0.22, out $0.66.
- Peak deepseek-v4-flash: cache-hit $0.014, cache-miss in $0.44, out $1.32 (2× off-peak).
- Off-peak deepseek-v4-pro: cache-hit $0.022, cache-miss in $0.66, out $1.98; peak doubles again.
- Legacy flat (before Aug 16): in $0.14, out $0.28, cache-hit $0.0028—obsolete.
Peak UTC windows: 01:00–04:00 and 06:00–10:00; everything else is off-peak at half price. Moving batch jobs to off-peak saves 50%, but if output is 70% of tokens, scheduling alone saves ~35% because off-peak output is still 3× cache-miss input. Compare Gemini track-based pricing: Gemini API 2026 pricing; Agent stack: 2026 AI Agent stack.
1. Why “what does DeepSeek cost?” has no single answer
Teams budgeted pre-Aug 16 $0.14 / $0.28, launched on Aug 17, hit UTC 07:00 peak, and output at $1.32/M blew the PoC. Others quoted off-peak $0.22 input while agent loops emit ≥3× output vs input, with thinking billed the same. No cache means an 80k system prompt at peak cache-miss $0.44/M costs $0.035 per call input-only—10k calls/day ≈ $350; the same prefix cache-hit off-peak is $0.00056/call.
August 2026 adds a time axis: fixed UTC peak windows, off-peak exactly half—not Gemini Batch, not Claude seats. Search results frozen before Aug 16 miss peak/off-peak, cache-hit rows, and checkpoint 0731. Claude subscription comparison: Claude Code 2026 pricing.
- Peak is whole-table ×2, not a small surcharge; one agent session can span both windows.
- Cache-hit vs miss gap ≈31× off-peak ($0.007 vs $0.22)—skipping cache is a voluntary repeat-input tax.
- 1M context is a capability cap, not free quota; filling 1M cache-miss input at peak is $440 input alone.
- Thinking defaults on at the same price; turning it off does not discount.
- MCP retries and lid-close timeouts rebill output—often at peak. Host choice: MCP on Cloud Mac vs VPS vs local.
2. Four billing dimensions
2.1 Peak vs off-peak (UTC)
Peak: 01:00–04:00 UTC and 06:00–10:00 UTC. Off-peak = half price. Cron batch jobs to off-peak; US/EU morning traffic hits the 06:00–10:00 peak band.
2.2 Cache-hit vs cache-miss input
Stable prefixes (system prompt, tool JSON Schema) hit cache at Flash off-peak $0.007/M. Changes force cache-miss full input price.
2.3 Input vs output tokens
Off-peak output $0.66 is 3× cache-miss input $0.22; peak output $1.32 is 94× cache-hit input $0.014. Agent hops and long CoT inflate output—output share is the main dial; peak/off-peak is a multiplier.
2.4 Flash vs Pro and execution host
deepseek-v4-pro off-peak cache-miss in $0.66, out $1.98 (~3× Flash). Upgrade only when Flash quality fails. API bills exclude Macs, MCP, and peak retries after failed iOS builds.
3. Comparison with one header set
| Layer / option | Entry | Execution | Context | Cost | Permission boundary |
|---|---|---|---|---|---|
| Off-peak V4-Flash | https://api.deepseek.com + key | 1M ctx; thinking same price | cache hit/miss tiers | in $0.007–0.22, out $0.66/M | OpenAI compatible; keys on backend |
| Peak V4-Flash | Same; UTC timestamp in peak | Same; queues may grow | Same tiers | in $0.014–0.44, out $1.32/M | Same key; time-sliced billing |
| Legacy flat (dead) | Pre–8/16 invoices | No peak/off-peak | No cache tiers in old table | in $0.14, out $0.28 | Do not budget |
| Off-peak V4-Pro | model=deepseek-v4-pro | Stronger reasoning | cache tiers | in $0.022–0.66, out $1.98/M | Hard-case router only |
| Cache strategy | Stable prefix | Cut miss input | system + Schema fixed | hit can cut input 97%+ | Prefix change = remiss |
| Execution host | Laptop / VPS / Cloud Mac | MCP, tests, signing | Repos, certs | Peak retries waste out tokens | Always-on host cuts waste |
Asymmetric takeaway: moving peak → off-peak saves 50%; cutting output share from 70% to 40% can save as much or more. Do both. Pro does not fix a forked schema; off-peak does not fix an uncached 80k system prompt.
3.1 Official cheat sheet (USD / 1M, Aug 2026)
| Model | Window | cache-hit in | cache-miss in | output | Notes |
|---|---|---|---|---|---|
| deepseek-v4-flash | Off-peak | $0.007 | $0.22 | $0.66 | Checkpoint 0731; thinking same rate |
| deepseek-v4-flash | Peak | $0.014 | $0.44 | $1.32 | UTC 01–04, 06–10 |
| deepseek-v4-pro | Off-peak | $0.022 | $0.66 | $1.98 | Peak ×2 |
| deepseek-v4-pro | Peak | $0.044 | $1.32 | $3.96 | Router only |
| (legacy to 8/16) | Flat | $0.0028 | $0.14 | $0.28 | Obsolete |
4. Scenario matrix
| Layer / option | Entry | Execution | Context | Cost | Permission boundary |
|---|---|---|---|---|---|
| Learn API / demo | https://api.deepseek.com + Flash | Short turns, no multi-hop | Single shot <8k | Anchor off-peak; <$5/mo | Test keys; backend proxy |
| Daily product, mixed TZ | Flash + cache system | Cap tool steps | Short ctx + cache hits | Budget out 3× in; mix peak/off | Monitor UTC distribution |
| Coding agent / MCP | Flash default; Pro 5% route | Thinking same price; fuse fails | Repo summaries | Output + retries dominate | Always-on MCP on Cloud Mac |
| Night batch / evals | Flash off-peak cron | 1M window in chunks | Bulk JSONL | Off-peak out $0.66 saves 50% | Queue policy written down |
| Long-doc RAG | Flash + high cache hit | Static blocks cached | Short dynamic query | Miss input controlled | Version prefixes |
| iOS / Xcode colocated | Cloud Mac Host + MCP | DeepSeek infers only | Certs on Mac | Day lease + avoid peak retries | Keychain never leaves box |
5. Recommended stacks
A | Personal PoC deepseek-v4-flash + off-peak trials → short prompts, no agent hops → use official table; not legacy $0.14/$0.28 B | Indie product (default) Flash + system/Schema cache → batch/evals cron in UTC off-peak → agent: max steps, peak queue or throttle → Pro only after repeated Flash failures C | Multi-model stack DeepSeek Flash for cost-efficient inference → route with Gemini / Claude by scenario (see stack article) → one JSON Schema, not three tool defs D | Apple delivery DeepSeek = inference only → MCP / xcodebuild / signing on Cloud Mac → lid-close retries at peak rebill $1.32/M output
6. Pitfalls
- Pitfall 1: Quoting pre–Aug 16 $0.14/$0.28. Peak/off-peak replaced flat pricing.
- Pitfall 2: Off-peak $0.22 input only. Output $0.66 off-peak, $1.32 peak—agents live on output.
- Pitfall 3: Expecting thinking-off discount. Same rate card; thinking on by default.
- Pitfall 4: No cache; 50k system every miss. Hit off-peak $0.007/M is ~31× cheaper.
- Pitfall 5: Ignoring UTC peaks—US morning is 06:00–10:00 UTC peak.
- Pitfall 6: Unbounded retries. MCP timeout at peak doubles output tokens.
- Pitfall 7: Defaulting to Pro. Flash off-peak out $0.66 is enough for many tasks; Pro out $1.98 is 3×.
7. Seven steps
- Open the official pricing page; confirm 2026-08-16 16:00 UTC effective date and peak windows; record six-cell Flash prices.
- Run the real production prompt once off-peak and once peak; log input/output and cache hit/miss—never estimate from “hello.”
- Stabilize system prompt and tool JSON Schema for cache; target >80% input tokens on hit.
- Cap tool steps and daily USD; on peak cost alerts, queue to off-peak or shorten replies.
- Cron batch, evals, and backfill to UTC off-peak; accept peak premium or throttle live entry.
- Move MCP and builds off lid-closing laptops. See MCP and Cloud Mac.
- Internal rate card: off-peak/peak columns, Flash default, Pro router threshold; diff invoices by UTC hour monthly.
8. FAQ
What does DeepSeek V4-Flash API cost now?
Depends on UTC window and cache. Off-peak cache-miss in $0.22/M, out $0.66/M; peak doubles. Cache-hit in off-peak $0.007/M. Pre–Aug 16 flat rates are dead.
How are peak and off-peak defined?
UTC 01:00–04:00 and 06:00–10:00 are peak (2× off-peak). Billing uses the API request UTC timestamp, not the user’s local timezone.
Are thinking and non-thinking priced the same?
Yes. V4-Flash defaults thinking on; same rate card—budget output including thinking tokens.
DeepSeek vs Gemini—which is cheaper?
No single sticker comparison. DeepSeek sells peak/off-peak × cache tiers; Gemini sells tracks × models. High cache hit + off-peak batch favors DeepSeek input; peak-heavy output agents need $1.32/M math. See Gemini API 2026 pricing.
Which model id?
deepseek-v4-flash (checkpoint DeepSeek-V4-Flash-0731). Pro: deepseek-v4-pro. Base https://api.deepseek.com; point OpenAI SDK at that URL.
Why a cloud Mac?
DeepSeek sells inference. If MCP, Xcode, or Keychain fail on a laptop, peak retries rebill output at $1.32/M. Always-on Cloud Mac cuts wasted output tokens, not the unit price.
9. Summary
Post–Aug 2026 DeepSeek V4-Flash API pricing is UTC peak/off-peak × cache hit × input/output mix. Off-peak half price saves 50% on the table, but 70% output share means scheduling alone saves ~35%; cut output, raise cache, kill peak retries together.
Order of work: lock six-cell prices on the official table → measure prod prompt peak and off-peak → cache system/Schema → anchor batch to off-peak → agent caps and peak fuses → keep the execution host up. Retire $0.14/$0.28; write output share and UTC windows into every finance model revision.
Spend tokens on inference, not on peak retries
Token cost is already printed per peak/off-peak and cache tier. What blows the budget is the agent timing out on a laptop, MCP dropping, and Xcode signing failing—then output runs again at peak $1.32/M. Dedicated M4 Cloud Mac keeps Host, repo, and MCP on one always-on path. Day-lease acceptance then monthly lock-in makes every million output tokens hit a real step.