No-BS Take · Pricing
DeepSeek V3.2 API Pricing: 2.7x Cloud Markup and the 20% Cache-Hit Flip
Pick the cheapest host on a DeepSeek V3.2 sticker sheet and you can still pick the wrong provider by 32% the moment your prompts start reusing context.
DeepInfra advertises $0.26 input and $0.38 output per million tokens, undercutting DeepSeek's own API at $0.28 / $0.40. Look one column over at prompt caching and enterprise cloud rates and two flips appear:
- The
20%cache-hit flip: DeepSeek cuts cached input tokens by90%down to$0.028per million, whereas DeepInfra ($0.13), Novita AI ($0.1345), SiliconFlow ($0.135), and FriendliAI ($0.25) only cut cached reads by50%. Once more than19.7%of your input tokens hit prompt cache, DeepSeek's first-party API overtakes every reseller on cost. - The
4.4xcloud output markup: Fireworks AI ($0.56 / $1.68) and Microsoft Azure AI Foundry ($0.58 / $1.68) charge more than twice as much for input and more than four times as much for output ($1.68vs$0.38), with no cached discount listed in our tracked feed. On a10Minput plus2Moutput month, a cold workload jumps2.7x($3.36on DeepInfra to$9.16on Azure), and an80%cache-hit agent loop jumps5.8x($1.58on DeepSeek direct vs$9.16on Azure).
Picture two gas stations across the street from each other: the discounter puts a price two cents lower on its highway sign ($0.26 vs $0.28), while the station next door gives 90% off repeat refills ($0.028 cached) and the airport cloud pumps charge nearly triple for the exact same tank with zero loyalty discount.
Why cold vs cached bills flip the leaderboard
Cost metric (10M input + 2M output / mo) | Cold (0% cache hit) | Coding or RAG loop (80% cache hit) |
|---|---|---|
DeepInfra ($0.26 in / $0.13 cache / $0.38 out) | $3.36 | $2.32 |
DeepSeek ($0.28 in / $0.028 cache / $0.40 out) | $3.60 | $1.58 |
Novita AI ($0.269 in / $0.1345 cache / $0.40 out) | $3.49 | $2.41 |
SiliconFlow ($0.27 in / $0.135 cache / $0.42 out) | $3.54 | $2.46 |
FriendliAI ($0.50 in / $0.25 cache / $1.50 out) | $8.00 | $6.00 |
Fireworks AI ($0.56 in / $1.68 out) | $8.96 (2.7x) | $8.96 (5.7x) |
Microsoft Azure AI Foundry ($0.58 in / $1.68 out) | $9.16 (2.7x) | $9.16 (5.8x) |
Two edge cases matter before switching endpoints:
- SambaNova's
11.6xconfiguration outlier: SambaNova is tracked at$3.00input and$4.50output ($39.00for the same month) with a32,768-token context window rather than128Kto164Kon the other hosts. Verify whether you specifically need that hardware configuration on the SambaNova pricing page. - Do not assume newer V4 models cost more either: DeepSeek V4.1 Flash lists at
$0.14 / $0.42on DeepInfra ($2.16cold at this volume) and$0.30 / $1.20($5.40) across larger hosts like Together AI, though DeepSeek's own off-peak rules complicate direct comparisons (see our DeepSeek V4.1 Flash off-peak breakdown). Test both on your actual prompt template on the DeepSeek V3.2 model page.
Which host to pick for your workload
- Long system prompts, RAG, or multi-turn coding (
>20%cache hits): Start with DeepSeek direct. The$0.028cached rate beats every reseller in our8-host snapshot cleanly once context repeats. - One-shot batch classification (
~0%cache hits): DeepInfra ($3.36), Novita AI ($3.49), and SiliconFlow ($3.54) sit within cents of each other. Pick by latency, uptime, and data residency rather than the third decimal. - Locked inside Azure enterprise procurement: Paying
$9.16instead of$1.58to$3.60buys consolidated invoicing and existing cloud compliance. Just price that$5.56to$7.58per10M/2Mpremium intentionally before scaling background loops there.
Caveats and tracked pricing sources
We compare standard tracked list prices, not batch tiers, committed-spend discounts, latency, rate limits, quantization, or cache Time-to-Live windows (a host with a short cache TTL can erase the $0.028 headline advantage if your loop pauses between turns). Always benchmark your own prompts before cutting over production traffic. How our daily pipeline ingests and filters LiteLLM and models.dev feeds is documented on the methodology page.
Tracked host links checked live 200 OK on November 14, 2026 (feed synced through November 13): DeepInfra, Novita AI, SiliconFlow, DeepSeek, FriendliAI, Fireworks AI, Microsoft Azure AI Foundry, and SambaNova.
The numbers
| Cheap 160K tier, DeepSeek V3.2 (per 1M tokens: input / cache read / output)Tracked list price | DeepInfra $0.26 / $0.13 / $0.38; Novita $0.269 / $0.1345 / $0.40; SiliconFlow $0.27 / $0.135 / $0.42; DeepSeek $0.28 / $0.028 / $0.40 | DeepInfra pricingChecked Nov 13, 2026 by AISelect daily price tracking (LiteLLM, models.dev), medium confidence |
| Cloud and premium hosts, DeepSeek V3.2 (per 1M tokens)Tracked list price | FriendliAI $0.50 / $0.25 cache / $1.50; Fireworks $0.56 / $1.68; Azure $0.58 / $1.68; SambaNova (32K) $3.00 / $4.50 | Fireworks AI pricingChecked Nov 13, 2026 by AISelect daily price tracking (LiteLLM, models.dev), medium confidence |
| DeepSeek V4.1 Flash comparison rates (per 1M tokens)Tracked list price | DeepInfra $0.14 / $0.42; Together AI $0.30 / $1.20 | Together AI pricingChecked Nov 13, 2026 by AISelect daily price tracking (LiteLLM, models.dev), medium confidence |
| Monthly cost for 10M input and 2M output tokens (cold vs 80% cached)Our calculation | Cold: DeepInfra $3.36, DeepSeek $3.60, Fireworks $8.96, Azure $9.16. 80% cached: DeepSeek $1.58, DeepInfra $2.32, Azure $9.16 | Our calculation from tracked list pricesChecked Nov 14, 2026 by AISelect editorial research, high confidence |
Frozen at publication. Current prices are shown below.
Today's API prices
Updated daily from our price tracking; the numbers above are frozen at publication.
| Model | Cheapest provider | Input / 1M | Output / 1M | Providers |
|---|---|---|---|---|
| DeepSeek-V3.2 | DeepInfra | $0.26 | $0.38 | 8 |
| DeepSeek-V4.1-Flash | DeepInfra | $0.14 | $0.42 | 11 |
Opinion, with receipts
More No-BS Takes
This is an opinion piece. The numbers are sourced; the conclusions are the author's. No provider paid for or reviewed it.