DeepSeek V4 Flash Provider Pricing — A Blended Cost Index
Research page: 2026-08-29. All prices are list prices in USD per 1M tokens from the cited pages. DeepSeek changed its official pricing structure on 2026-08-17 (peak/off-peak), so snapshot dates matter. Verify before committing to any route.
TL;DR
- The claim "third-party providers advertise attractive open-model pricing, but the hidden cost is the cache-hit price, which is usually much higher than the original provider" is directionally true for DeepSeek, but the ranking flips by model tier and workload mix. For V4 Flash, most providers beat official DeepSeek on blended cost even at 95%+ cache-hit ratios; for V4 Pro, official DeepSeek wins on every line.
- A pricing page has four numbers, not one: cache-hit input, cache-miss input,
output — plus peak/off-peak for the official API. We collapse them into a
single index:
E = r·[h·H + (1−h)·M] + O, normalized to official = 100 (lower = cheaper). - Calibrated on two real coding-agent sessions (h ≈ 99.5% and 96.8%), the V4 Flash ranking is: StreamLake 22 < OpenRouter (aggregate) 41 < Fireworks 80 < DeepSeek official 100 ≈ DeepInfra 101 < SiliconFlow 157 < DigitalOcean 162 < Together 167 ≈ Novita 167.
- The cheapest routes win on unit price but carry caveats: StreamLake's OpenRouter listing contradicts its own 0731 price sheet; OpenRouter's default routing can erode realized cache-hit rates; and the price leaders are the least transparent on data retention.
1. The problem: four prices, not one
Official DeepSeek API (snapshot 2026-08-23)
DeepSeek's official pricing is now peak/off-peak (peak = weekdays 01:00–04:00 and 06:00–10:00 UTC; off-peak is half of peak):
| Model | Cache hit (off-peak / peak) | Cache miss (off-peak / peak) | Output (off-peak / peak) |
|---|---|---|---|
| deepseek-v4-flash | $0.007 / $0.014 | $0.22 / $0.44 | $0.66 / $1.32 |
| deepseek-v4-pro | $0.022 / $0.044 | $0.66 / $1.32 | $1.98 / $3.96 |
The cache-hit discount is unusually aggressive: ~30x below cache-miss (it was ~50x before the Aug 17 repricing). That is the single most important fact about DeepSeek's pricing.
Third-party providers for V4 Flash (snapshots late Aug 2026)
| Provider | Cache hit | Cache miss | Output |
|---|---|---|---|
| StreamLake (OpenRouter listing) | $0.0028 | $0.088 | $0.064 |
OpenRouter aggregate (latest alias) |
$0.007 | $0.03 | $0.10 |
| Fireworks | $0.007 | $0.22 | $0.66 |
| DeepInfra | $0.018 | $0.09 | $0.18 |
| SiliconFlow | $0.028 | $0.13 | $0.28 |
| Together | $0.03 | $0.14 | $0.28 |
| Novita | $0.03 | $0.14 | $0.28 |
| DigitalOcean | $0.03 | $0.08 | $0.25 |
The pattern behind the claim is visible: providers undercut or match official on cache-miss input and output, while their cache-hit price is 1x–4x official (pre-Aug-17 it was up to ~28x, when official cache hits were $0.0028). Whether that matters depends entirely on the workload's cache behavior — which is why a flat price table is misleading.
2. The index model
Blended effective cost per 1M output-equivalent tokens:
- H = cached-input price, M = cache-miss input price, O = output price
- h = cache-hit ratio (share of input tokens served from cache)
- r = input:output token ratio of the workload
- p = share of traffic in official peak hours (official prices are blended as
X = (1−p)·X_off-peak + p·X_peak; third parties are flat)
Index: I = E_provider / E_official × 100 — official is always 100, lower is
cheaper. E is "expected cost per 1M output tokens including the input traffic
it takes", so it is comparable across providers for a given workload profile.
Why the cache-hit rate is (conditionally) critical
A 1-percentage-point hit-rate change moves cost by r·(M−H)/100:
| Workload | r | Cost shift per 1pp hit-rate | Share of E |
|---|---|---|---|
| Coding agent S1 (99.5% hit) | ~143 | ±$0.30 / 1M output | ~17% |
| Coding agent S2 (96.8% hit) | ~53 | ±$0.11 / 1M output | ~8% |
| Chat (r=4) | 4 | ±$0.009 / 1M output | <1% |
So h matters because long sessions resend massive cached context (large r), not by itself. It is linear in the model, but it is not a provider constant — it depends on prefix stability, cache eviction, and especially routing. It has to be measured, not assumed.
3. Calibration on real sessions
Two real coding-agent sessions on DeepSeek official (from usage lines):
| Session | Cached input | Fresh input | Output | h | r |
|---|---|---|---|---|---|
| S1 "add review layer" | 13,521,536 | 67,079 | 95,292 | 99.51% | 142.6 |
| S2 "review repo vs requirements" | 1,118,080 | 37,484 | 21,794 | 96.76% | 53.0 |
Combined ranking (equal weight of both sessions, official = 100, p = 25% peak share):
| Rank | Route | S1 index | S2 index | Combined |
|---|---|---|---|---|
| 1 | StreamLake | 23 | 21 | 22 |
| 2 | OpenRouter (aggregate) | 49 | 29 | 41 |
| 3 | Fireworks | 80 | 80 | 80 |
| 4 | DeepSeek official | 100 | 100 | 100 |
| 5 | DeepInfra | 124 | 72 | 101 |
| 6 | SiliconFlow | 192 | 111 | 157 |
| 7 | DigitalOcean | 202 | 110 | 162 |
| 8 | Together | 205 | 118 | 167 |
| 9 | Novita | 205 | 118 | 167 |
What the numbers say:
- Stable winners: StreamLake and OpenRouter are 2–5x cheaper than official in both sessions; Fireworks ties official off-peak exactly (same per-token prices), so its index is a constant 80 under the 25% peak blend.
- The flip-flopper: DeepInfra is 72 in S2 (cheap output wins) but 124 in S1 (its $0.018 cache price gets amplified 143x). It straddles official depending on how cache-heavy the session is.
- The "hidden cache cost" losers: SiliconFlow, DigitalOcean, Together, and Novita are more expensive than official in both real sessions (110–205), even where they undercut official on sticker output price. Their cache-hit price is the reason.
- Cost composition shifts: in S1 cached input is ~55% of the bill; in S2 it drops to ~26% and output (~47%) and fresh input (~27%) re-emerge. The same formula, two different worlds.
4. Why StreamLake is so cheap — and the fine print
StreamLake is Kuaishou Technology's cloud/AI commercialization brand (Kuaishou is one of China's two dominant short-video platforms). It started as a video cloud in 2022 and entered AI coding in Oct 2025 with the Wanqing MaaS platform, KAT coder models, and the CodeFlicker tool. It runs inference at massive scale for its own products and is pricing DeepSeek serving aggressively to win enterprise share — an incumbent land-grab, not a startup pricing at cost.
That explains the index's rank 1: a ~$0.0028 cache-read price (essentially free) dominates at high h/r. But verify before committing:
- The price may not match the model you get. StreamLake's own published price for DeepSeek-V4-Flash-0731 (updated 2026-08-24) is ¥3 input / ¥0.1 cache hit / ¥9 output ≈ $0.45 / $0.015 / $1.34 — and Vercel's gateway snapshot agrees ($0.44 / $1.32 / $0.01). The OpenRouter listing ($0.088 / $0.0028 / $0.064) looks like an older build or a promotional/international tier. Even trackers disagree with each other on output ($0.064 vs $0.176 for the same provider name).
- The cache price is conditional. Caching is best-effort; entries are evicted after hours to days. The $0.0028 rate only materializes with real hits.
- Reliability and throughput are unproven at this price. Historical uptime is decent (99.5–99.96% depending on model/period) but throughput has been low on several models, and OpenRouter auto-retries mask provider failures.
- Data and geography. Chinese parent, regional inference, unknown prompt-logging policy — a compliance consideration for sensitive workloads.
- It is a volatile promotional price. StreamLake repriced the 0731 build on 2026-08-24; the market's cheapest input price fell ~37% over the prior 90 days. Cheap rates are snapshots, not commitments.
5. OpenRouter: an aggregator that inherits the floor
OpenRouter scores well because of its structure, not its own infrastructure:
- No per-token markup. It passes through provider prices; its fee is on credit purchases (~5.5%), invisible to a per-token index.
- It surfaces the floor. Default routing is price-prioritized load balancing
(weighted by inverse-square of price, with uptime filters), and the
:floormodel variant sorts strictly by price. The aggregate cannot be more expensive than its cheapest eligible endpoint. - It inherits the best cache price on the market — the aggregate's $0.007 cache read ties official off-peak.
The caveats:
- The aggregate listing is a snapshot, not a guarantee. The model-page price changes over time and may not reflect what a given request is routed to.
- Routing can break cache continuity. Load-balancing across upstreams
invalidates prefix caches; requests pinned to one provider preserve them. Use
the
providerobject (order,only,zdr,quantizations) or model variants (:floor,:nitro,:exacto) to control routing. - You are "using OpenRouter" as provider of record (endpoint, billing,
support, zero-logging layer), but the upstream is opaque at request time — the
response's top-level
providerfield tells you who actually served you, and per-provider prices are public on the model page.
For the index, the honest presentation is two rows: "OpenRouter aggregate (opaque routing)" and "OpenRouter pinned to a specific provider" — the aggregate's advantage partly is the opacity.
6. Retention policies (verified Aug 2026)
| Route | Index (combined) | Prompt/log retention | Trains on API data? | Notes |
|---|---|---|---|---|
| StreamLake | 22 | Network logs ≥6 months (CN law); API prompt window not published | Not stated | Kuaishou; moderation logs retained per law; no ZDR |
| OpenRouter | 41 | No prompt/completion storage by default; opt-in logging earns 1% discount | No (can filter training providers) | Downstream provider policies apply; enterprise EU/US routing |
| Fireworks | 80 | Zero retention by default; 30 days only if store=True (Responses API) |
No (opt-in only) | SOC 2 / ISO |
| DeepSeek official | 100 | Not published; CN legal baseline (network logs ≥6 months) | Not explicitly committed | No DPA template, no ZDR |
| DeepInfra | 101 | Zero retention — inputs in memory only during request; outputs not stored | No | SOC 2 / ISO 27001 |
| SiliconFlow | 157 | Not published; retained as long as necessary per CN law | Not committed | CN provider; no ZDR found |
| DigitalOcean | 162 | Hosted models: inputs/outputs not stored; batch artifacts 29–30 days; abuse evidence ≤1 year | No (hosted models) | Third-party models follow their own policies |
| Together | 167 | ZDR by default; temporary caching unless disabled | No (opt-in only) | SOC 2 |
| Novita | 167 | API content not clearly published; account/transaction data 7 years | Third-party allowlists say no-train | Verify contractually |
The pattern: Western providers (DeepInfra, Fireworks, Together, DigitalOcean, OpenRouter's own layer) offer zero-retention/no-training by default; the Chinese-first providers (DeepSeek official, SiliconFlow, StreamLake) publish no API retention window and sit under China's ~6-month network-log retention law. Price and privacy are inversely correlated in this market.
7. Caveats and method notes
- Snapshots. Official DeepSeek prices changed structure on 2026-08-17 (peak/off-peak); provider prices move weekly. The index must be re-run with current prices.
- Versions matter. V4 Flash 0423 vs 0731 and V4 Pro vs 0813 carry different prices (e.g., Fireworks prices the 0813 Pro build at exactly 2x official off-peak, giving a constant index of 160 at any workload mix — a good sanity check).
- CNY→USD conversion (~6.7–7.2) affects cross-currency comparisons.
- Not modeled: batch discounts (Fireworks batch = 50%), quantization
differences, concurrency/throughput limits, latency, SLA, and realized cache
hit rates per provider. The last one is the biggest unknown — measure
prompt_cache_hit_tokens/prompt_cache_miss_tokensper provider on your actual workload before committing. - Verdict on the original claim: the hidden cache-hit price is real and can flip a ranking, but only when input amplification (r) and hit ratio (h) are high. For V4 Flash at real session mixes, most providers still beat official on blended cost; for V4 Pro, official DeepSeek wins across the board. The useful framing is not "providers are cheaper/expensive" but "at what (h, r) does each route overtake official".
8. Recommended setup (cache-heavy coding agents)
For cache-heavy agent workloads on DeepSeek V4 Flash — with GLM/Kimi coverage, zero retention, and pinned cache continuity — the working recommendation as of this snapshot:
- Primary: Fireworks AI direct. DeepSeek V4 Flash blended index 80 at real session mixes (cache read $0.007 = official off-peak), GLM-5.2 cached $0.14 vs Zhipu official ≈ $0.28, Kimi K2.6 cached $0.16 (parity with Moonshot). ZDR by default, SOC 2/ISO, single upstream (cache continuity by construction), batch inference at 50%.
- Runner-up: DeepInfra. Also ZDR with DeepSeek/GLM/Kimi coverage; cheaper Flash output ($0.18) and better in the 96.8%-hit session (index 72 vs 80), but worse at 99.5% hit (124 vs 80) because its Flash cache read is $0.018 vs $0.007.
- Optional management layer: Vercel AI Gateway pinned to Fireworks (0% markup, BYOK free, provider pinning) or self-hosted LiteLLM (zero fees, full control). Avoid unpinned gateways — load balancing breaks prefix-cache continuity.
- Do: pin one upstream, monitor
prompt_cache_hit_tokens(target ≥95% on agent sessions), keepstore=False(Fireworks' 30-day retention applies only if you opt in), use batch for non-interactive workloads.
Caveats: for DeepSeek V4 Pro, official remains cheaper than every third party (Fireworks Pro index ≈ 164) — this recommendation is Flash-scoped. And StreamLake (index 22) is far cheaper but not ZDR; choosing it means accepting the price-vs-privacy tradeoff.
Sources
- DeepSeek pricing docs — api-docs.deepseek.com/quick_start/pricing (snapshot 2026-08-23)
- DeepInfra pricing — deepinfra.com/pricing (fetched 2026-08-29)
- Fireworks serverless pricing — docs.fireworks.ai/serverless/pricing (fetched 2026-08-29)
- SiliconFlow model pages — siliconflow.com (fetched 2026-08-29)
- Together AI — together.ai/blog/deepseek-v4-pro-now-available-on-together-ai
- DigitalOcean inference docs — docs.digitalocean.com/products/inference/details/data-privacy/
- OpenRouter — openrouter.ai/~deepseek/deepseek-v4-flash-latest; docs/features/provider-routing; docs/faq
- StreamLake pricing — streamlake.com/document/WANQING (updated 2026-08-24); Vercel AI Gateway provider page
- Pandaily — "Why DeepSeek Could Charge 30x More and Still Be the Cheapest Model Around" (2026-08-08)
- TheRouter.ai — LLM API data privacy and retention cross-provider reference (2026-08-12)
- banandre.com — DeepSeek V4 Flash cost/performance analysis (2026-08-02)