DeepSeek V4 Flash Provider Pricing — A Blended Cost Index for Cache Hits, Cache Misses, and Output

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, among the kept routes Fireworks beats official at every hit mix (80) and DeepInfra straddles it (72–124) — the rest lose 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 sync ranking (StreamLake excluded — region; OpenRouter aggregate excluded — unusable endpoints) is: Fireworks 80 < DeepSeek official 100 ≈ DeepInfra 101 < SiliconFlow 157 < DigitalOcean 162 < Together 167 ≈ Novita 167. Batch mode flips the top: Fireworks 40 < DeepInfra ~81 < official 100.
• Companion: GLM-5.3-Flash net-of-fee index — the high-cache leader is Vercel’s pinned cheapest upstream (~43.5 even after its ~12% fee), ahead of Fireworks batch (~48.5) and Z.ai promo (50); low cache hit flips to batch and promo (~50) with Vercel-pinned dropping to ~70.5. Fees cost ~5 index points on OpenRouter routes and ~4.6 on the Vercel-pinned route (see §9).
• Trust: both recommended providers are NVIDIA-backed, SOC 2 / ISO 27001 certified per company disclosures, and founded in 2022. Fireworks is the enterprise heavyweight ($17.5B valuation, $1B+ ARR, Sequoia/Benchmark/ Lightspeed/Index, Microsoft partnership, customers Uber/Shopify/GitLab); DeepInfra is the lean specialist (~$133M total, 500 Global/Felicis plus NVIDIA/Samsung, zero retention, 99.982% uptime SLA) (see §10).
• The cheapest advertised routes are excluded, not recommended: StreamLake is CN-hosted with unknown prompt-logging (region constraint), and OpenRouter’s aggregate low price routes to endpoints with no throughput/retention guarantees — pin one provider or pay the 5.5% fee on a real host. Among the kept routes the levers are batch discounts (Fireworks 50%, DeepInfra 20%) and the fee rails (Fireworks/DeepInfra net fee 0; OpenRouter 5.5%; Vercel ~12%).

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):

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)

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:

$$ E(h, r, p) = r \cdot \big[ h \cdot H + (1-h) \cdot M \big] + O $$

• 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.

Sync vs batch. Rankings below are quoted in both modes. Sync (interactive) = serverless list prices. Batch = documented async discounts only: Fireworks Batch API bills 50% off serverless on input and output (so its index halves), DeepInfra Batch API bills 20% off with a ≤24-hour window. Batch is async-only — an interactive agent cannot use it, and providers without a documented V4 Flash batch price keep their sync index.

Why the cache-hit rate is (conditionally) critical

$$ \frac{\partial E}{\partial h} = -r \cdot (M - H) $$

A 1-percentage-point hit-rate change moves cost by r·(M−H)/100:

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):

Combined ranking (equal weight of both sessions, official = 100, p = 25% peak share). StreamLake is excluded (region constraint) and the OpenRouter aggregate is excluded (its advertised low price routes to endpoints with no throughput or retention guarantees — see §4/§5); only pinned, usable routes are ranked:

What the numbers say:

• The only host under official in both sessions: Fireworks ties official off-peak exactly (same per-token prices), so its sync index is a constant 80 under the 25% peak blend — and its batch index drops to 40.
• 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; its Batch API (20% off) pulls it to ~81 combined.
• 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. DigitalOcean is the extreme case: its $0.03 cache read is 3.4x official blended (4.3x off-peak), so cached input alone is ~93% of its S1 bill (vs ~55% for official) — its cheap $0.08 cache-miss and $0.25 output can’t compensate at 99.5% hit.
• 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.

Batch note: Fireworks batch discounts the cache-read line too (50% off all input/output), so its 40 holds at any hit rate; DeepInfra batch is 20% off all lines. Batch values for providers without a documented V4 Flash batch price are shown at sync (no discount).

4. Excluded routes (and why)

Two routes that score well on raw price are deliberately left out of the rankings:

• StreamLake — region constraint. Kuaishou’s Wanqing MaaS platform prices DeepSeek serving at an incumbent land-grab (its OpenRouter listing reads $0.0028 cache / $0.088 miss / $0.064 out, roughly the index’s best raw price), but it is a Chinese parent with regional inference, an unknown prompt-logging policy, no ZDR, and no DPA template — and its own price sheet contradicts the listing (¥3 / ¥0.1 / ¥9 ≈ $0.45 / $0.015 / $1.34), which looks like an older build or promotional tier. Not usable for this workload’s compliance posture, so it is excluded rather than ranked.
• OpenRouter aggregate — unusable endpoints. The aggregate’s advertised low price (index ~41) is a snapshot of its cheapest eligible endpoints, but those endpoints carry no throughput, retention, or reliability guarantees, and auto-retries mask upstream failures. The low price is not real for production use; only pinned routes (a specific host + the 5.5% fee) count here (see §5).

Removing them flattens the market: the surviving spread is Fireworks 80 → official 100 ≈ DeepInfra 101 → SiliconFlow 157 → DigitalOcean 162 → Together/ Novita 167, with batch mode adding Fireworks 40 and DeepInfra ~81.

5. OpenRouter: an aggregator that inherits the floor

OpenRouter scores well because of its structure, not its own infrastructure:

• No per-token markup, but a credit-purchase fee. It passes through provider prices; its fee is on credit purchases — 5.5% (with a $0.80 minimum) for card top-ups, 5% for crypto — so it’s invisible to a per-token index but real at the till. That’s roughly half the bite of Vercel’s AI Gateway credit top-ups, which add payment processing plus tax (about 12% combined, see §7).
• It surfaces the floor. Default routing is price-prioritized load balancing (weighted by inverse-square of price, with uptime filters), and the :floor model 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 — and not a usable price. The model-page price changes over time, may not reflect what a given request is routed to, and its floor routes to endpoints with no throughput, retention, or reliability guarantees. This is why the aggregate is excluded from the rankings (§4); the only usable OpenRouter price is a pinned host.
• Routing can break cache continuity. Load-balancing across upstreams invalidates prefix caches; requests pinned to one provider preserve them. Use the provider object (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 provider field tells you who actually served you, and per-provider prices are public on the model page.

For the index, the aggregate row is excluded (§4); the honest presentation is “OpenRouter pinned to a specific provider” at that provider’s price × 1.055 (e.g., Fireworks 80 → ~84, DeepInfra 101 → ~107).

6. Retention policies (verified Aug 2026)

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 — and StreamLake, excluded for region) 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: 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_tokens per provider on your actual workload before committing. (Batch discounts are modeled: Fireworks 50%, DeepInfra 20% — see §3.)
• Not modeled: payment fees and tax. The index compares list token prices; the money actually leaving your account includes a billing layer on top. OpenRouter charges 5.5% (min $0.80) on card credit purchases. Vercel AI Gateway charges no token markup or platform fee, but credit top-ups carry payment processing plus tax — reported around 12% combined in mid-2026 (e.g., 2.9% + $0.30 processing plus ~9.3% sales tax; exact rate depends on card and tax jurisdiction, and invoice billing avoids the processing fee). Fireworks and DeepInfra were checked for this refresh: Fireworks uses prepaid credits with monthly postpaid billing and DeepInfra bills cards monthly — neither documents a payment-processing surcharge, so their net fee is 0 (the one soft cost is Fireworks’ prepaid credit float). DigitalOcean is the outlier on the cheap side of the till: top-ups via PayPal (the only prepay option) carry no documented surcharge — you pay the list price and the balance just lags a few hours while PayPal processes — which keeps its effective cost at the token price — though at index 162 the token price is still the problem. At real monthly spend this widens the effective gap between routes billed through Vercel and the same provider billed directly, and it makes OpenRouter (~5.5%) cheaper at the till than Vercel (~12%) even where token prices tie.
• 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, Fireworks beats official (80) and DeepInfra straddles it (72–124) once StreamLake and the OpenRouter aggregate are excluded; 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”.

Net-of-fee ranking (what actually leaves your account)

Applying the billing rails from the caveat above (OpenRouter ×1.055, Vercel ×1.12, Fireworks/DeepInfra ×1.0 — no documented surcharge; sync mode unless noted):

Fees don’t reshuffle the leaderboard — the gaps (40 → 80 → 100) dwarf the 5.5–12% spread — but they matter in two places. First, the management-layer math: Fireworks is ~90 via Vercel AI Gateway vs ~84 via OpenRouter pinned vs 80 direct, so Vercel’s ~12% bite costs roughly 10 index points over direct billing — the gateway only earns that fee if you use its routing/observability features (or switch to invoice billing, which drops the processing fee). Second, among the kept routes only Fireworks and DeepInfra direct have a documented net fee of 0; every OpenRouter-pinned route pays 5.5% on top of the host’s price.

DigitalOcean is the fee-free outlier, but that doesn’t save it for cache-heavy agents: its cache read ($0.03) is 4.3x official off-peak, and it sits at 202 / 110 on the two real sessions. It beats official only below ~92–95% hit rate at agent r values (crossover is workload-dependent: ~92% at r≈143, ~95% at r≈53); for low-r chat it is actually cheaper at any hit rate because output ($0.25) beats official blended ($0.825). For cache-heavy coding agents it’s the wrong profile.

Worked example: why Fireworks via Vercel (89.6) still beats official (100)

Fireworks’ three prices ($0.007 / $0.22 / $0.66) are exactly official off-peak. Official blended at p = 25% peak share multiplies every line by 1.25 (0.75×1 + 0.25×2), and since E is linear in prices, official E = 1.25 × Fireworks E for any (h, r). The raw index is therefore exactly 100/1.25 = 80 regardless of cache behavior; Vercel’s ~12% fee makes it 80 × 1.12 = 89.6.

The entire edge is peak-hour avoidance: official charges 2× during peak and the model assumes 25% of traffic lands there, so official costs 1.25× off-peak while Vercel-gated Fireworks costs 1.12× off-peak. Break-even: the fee would need to reach ~25% (or your peak-hour share drop below ~12%) before Fireworks via Vercel loses to official. If your workload is 100% off-peak (p = 0), direct Fireworks ties official and Vercel-gated Fireworks becomes 112 — worse.

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 (sync). 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), net fee 0. Batch mode drops it to 40 — the cheapest route on the board.
• 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. Net fee 0; Batch API (20% off, ≤24h) pulls it to ~81.
• Optional management layer: Vercel AI Gateway pinned to Fireworks (0% markup, BYOK free, provider pinning) — but factor in its ~12% credit-purchase fee plus tax (vs OpenRouter’s 5.5%; see §7) — 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), keep store=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. StreamLake is excluded (region constraint) and the OpenRouter aggregate is excluded (unusable endpoints) — see §4; if those constraints don’t apply to you, they are the cheapest raw routes on the market.

9. Companion index: GLM-5.3-Flash (net of fees, high vs low cache hit)

Full GLM-5.3-Flash research (intelligence column, promo rules, Vercel gateway note): the glm-5-3-flash-cost-index.md companion. Same blended formula, same fee rails as this page (direct ×1.0, OpenRouter 5.5%, Vercel ~12%). Unlike DeepSeek, GLM-5.3-Flash has no hidden cache markup — every major host matches or undercuts Z.ai’s $0.03 cache read — so the board is flatter and the real levers are promo, batch, and the fee rails.

Scenario definitions:

• High cache hit — S1 calibration (h = 99.51%, r = 142.6)
• Low cache hit — cache-light session (h = 50%, r = 53)

Prices (USD/1M, snapshot 2026-08-28): Z.ai list / promo-to-Sep-9 $0.03 / $0.015 hit, $0.15 / $0.075 miss, $0.50 / $0.25 out; Fireworks serverless $0.029 / $0.15 / $0.50 and batch 50% off; Makora $0.024 / $0.14 / $0.47; parity hosts $0.03 / $0.15 / $0.50; DigitalOcean hosts GLM-5.3-Flash directly at the same parity price (list, not promo), billed through its prepaid balance with no PayPal top-up surcharge; Vercel’s cheapest GLM upstream $0.01 / $0.10 / $0.40 (dynamic routing — pin it).

Takeaways:

• Fees don’t decide this board either. OpenRouter’s 5.5% costs ~5 index points at parity (100 → 105.5) and ~4.5 on Makora; Vercel’s ~12% adds ~4.6 points on the pinned upstream but it still holds the high-cache #1 because its $0.01 cache read beats even the promo.
• Cache sensitivity shows up in the ordering, not the raw prices. Vercel-pinned drops from #1 (43.5) to #4 (70.5) at low h because its edge is the cache read; Makora (86.2 → 96.4) and Fireworks serverless (97.1 → 99.5) drift toward parity for the same reason. Batch and promo are hit-rate-agnostic and own the low-cache board.
• DigitalOcean is the fee-free parity route. It hosts GLM-5.3-Flash at the same list prices as Z.ai ($0.03 / $0.15 / $0.50) but bills through its prepaid balance — PayPal top-ups carry no surcharge — so it nets 100 instead of the ~105.5 the same token prices cost through OpenRouter. That’s the clearest case of a billing rail deciding between tied providers; it also keeps DO’s ZDR and 88 t/s on OpenRouter.
• Cross-model: at high h, DeepSeek V4 Flash official (46.5 on the GLM=100 basis) still beats every GLM route except Vercel-pinned (net ~43.5). At low h the roles flip — DeepSeek jumps to ~158 on the GLM=100 basis, so GLM list wins outright, consistent with the companion’s h ≈ 87–90% crossover for agent-shaped workloads.
• Post-promo (after Sep 9): the leaderboard becomes Vercel-pinned (~43.5), Fireworks batch (~48.5), DeepInfra batch (~80), Makora (~86.2), Fireworks serverless (~97.1) — the promo row disappears and Z.ai list/parity (100) drop to the middle.

10. Trust and backing: Fireworks AI vs DeepInfra

The two providers this page actually recommends (Fireworks primary, DeepInfra runner-up) are both young — each founded in 2022 — both backed by NVIDIA, and both SOC 2 / ISO 27001 certified per company disclosures. The real difference is scale, not pedigree: Fireworks is a four-year-old enterprise heavyweight; DeepInfra is a lean specialist inference cloud.

Verdict. Both clear the production bar for the §8 workload, and both are credible partly because of who stands behind them: NVIDIA, Sequoia/Benchmark (Fireworks) and 500 Global/Felicis plus Samsung/Supermicro (DeepInfra), with enterprise customers (Uber, Shopify, GitLab) or operator pedigree (imo’s 200M users). The tradeoff is scale: Fireworks has the bigger balance sheet and more independent references; DeepInfra is smaller, so watch its security/abuse operations and capacity as it grows — its 99.982% SLA and zero-retention stance are the strongest signals it publishes. Neither is risk-free: both are 2022 startups in a market that reprices weekly, and certification claims (especially Fireworks’ HIPAA/PCI scope) should be re-verified on their trust pages before regulated workloads.

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
• Vercel AI Gateway pricing — vercel.com/docs/ai-gateway/pricing (no token markup; payment-processing fees on credit top-ups)
• OpenRouter FAQ — openrouter.ai/docs/faq (5.5% credit-purchase fee, $0.80 minimum)
• GLM-5.3-Flash pricing — companion page glm-5-3-flash-cost-index.md (snapshot 2026-08-28): Z.ai, OpenRouter endpoints, Fireworks, Vercel AI Gateway; 50% promo to 2026-09-09 16:00 UTC
• Fireworks AI — fireworks.ai/blog/series-c (Oct 2025); Jul 2026 $1.5B @ $17.5B coverage (ISV World, Reuters); Operator Collective founder profile; zhidx (2026-05)
• DeepInfra — deepinfra.com/blog/deepinfra-series-b (May 2026); about page + homepage (SOC 2/ISO, uptime SLA); Felicis “Deep Dive on DeepInfra”; 500 Global profile; Data Center Dynamics coverage
• Fireworks billing/credits — docs.fireworks.ai (prepaid credits, postpaid monthly, no documented surcharge); Batch API docs (50% off serverless, async)
• DeepInfra billing/batch — docs.deepinfra.com billing portal (monthly card billing, no documented surcharge); deepinfra.com/blog/batch-api + docs (20% off real-time, ≤24h)
• 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)