DeepSeek V4.1 Flash Pricing: $0.003 Cache Hits & Peak Rates

DeepSeek sets V4.1 Flash at $0.003 per 1M cached input tokens, $0.15 per 1M uncached input tokens, and $0.60 per 1M output tokens during off-peak hours. Peak hours are exactly double. The rates took effect at 04:00 UTC on September 10, 2026^[3].

The API model name is deepseek-flash. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but they now route to V4.1 Flash and are billed at these Flash rates^[1].

DeepSeek continues the peak/off-peak scheme it introduced on August 16, 2026. Peak windows are Monday through Friday from 01:00-04:00 UTC and 06:00-10:00 UTC. All other hours — including all weekend — are off-peak and cost half as much^[3][5].

For schedulable agent jobs, the time a workload runs is a cost lever every bit as real as token count. A batch job moved out of the weekday morning windows pays 50% less for identical work. The windows are defined in UTC, so teams in the US or Asia should map them to local time carefully^[3][5].

• Peak: Mon-Fri 01:00-04:00 UTC and 06:00-10:00 UTC
• Off-peak: all other hours, plus all weekend
• Off-peak output: $0.60/1M — half the $1.20 peak rate
• Shift flexible batch and eval workloads off-peak to halve the bill^[3]

Cached input is where the economics of agents get decided. At $0.003 per 1M off-peak, a cache hit costs 2% of what an uncached input token costs ($0.15). DeepSeek itself says cache-hit charges often account for a large share of agent costs^[5].

An agent repeatedly working against the same repository, tool definitions, system instructions, or conversation history rereads far more cached context than fresh context. DeepSeek cut V4.1 Flash's global KV cache to 890 bytes per token — about a quarter of V4-Flash's HBM footprint — which is what makes the low cache-hit price sustainable^[2][4].

The practical implication is that comparing models on uncached input price alone can obscure the fastest-growing component of agent economics. The metric that matters is cost per completed task, not headline input price^[5].

• $0.003/1M off-peak vs $0.15/1M uncached — a 50x difference
• Reuse stable prompt prefixes to maximize cache-hit ratio
• Measure cache-hit ratio and cost per successful task, not just token totals^[5]

Consider an agent that retains a 500,000-token reusable prefix and hits that cache across 100 requests — 50 million cached input tokens total. Ignoring cache-write, new uncached context, and output, those cache reads cost about $0.15 on V4.1 Flash off-peak, according to VentureBeat's calculation^[5].

The same 50M cached tokens would cost about $15 on Kimi K3, $20 on GPT-5.6 Sol, and $25 on Claude Opus 5 at their published cache-read rates. That is a 100x-to-160x gap on the cached portion alone — which is why input-heavy, repetitive, cacheable workloads are where V4.1 Flash's advantage is largest^[5].

V4.1 Flash is dramatically cheaper than the old V4 Flash pricing after the August peak/off-peak repricing, and far cheaper than V4 Pro. On off-peak output, V4.1 Flash is $0.60/1M versus V4 Pro's $1.98/1M and V4 Flash's old $0.66/1M peak-rate structure^[3].

The September 10 change effectively reverses part of the August price rise. Bloomberg Intelligence analysts put the cut at as much as 32%. Meanwhile, every V4-Pro request routes to V4.1 Flash from September 14 and is billed at Flash rates — so V4-Pro customers get a cheaper, nominally more capable model without changing their code^[3][5].

At $0.15 input / $0.60 output off-peak, or $0.75 blended per 1M tokens, V4.1 Flash sits near the absolute low end of the global paid API market. Only Meta's Contributor-tier Muse Spark and Xiaomi's MiMo-V2.5 Flash are cheaper in the comparison table VentureBeat assembled^[5].

Even at peak, V4.1 Flash ties with MiniMax-M3 and LongCat's promotional pricing and sits far below the mid-tier cluster around $4.50-$8.00. It is dramatically cheaper than Claude Fable/Mythos 5.1 and GPT-5.6 Sol Fast mode^[5].

Three levers dominate V4.1 Flash cost: cache-hit ratio, reasoning effort, and scheduling. Reasoning effort is now a continuous 1-100 integer, and DeepSeek's own tests show the gains are front-loaded^[5].

DeepSeek says going from effort 25 to 100 raises DeepSWE v1.1 from 66.0 to 74.2, but consumes roughly 2.5x as many output tokens. The final step to 100 makes agent trajectories 1.6-1.8x longer for comparatively small gains — so the leaderboard configuration is often not the economical one^[5].

• Maximize cache hits: keep stable prefixes, reuse system prompts and tool definitions.
• Pick effort 60-80 instead of 100: most of the accuracy at under half the output tokens.
• Schedule batch and eval jobs off-peak to halve the rate.
• Use non-thinking mode for simple tasks to avoid spending budget on the trace.
• Track cost per completed task, not cost per token^[5]

Billing is straightforward: expense = number of tokens x price. Fees are deducted from your topped-up balance or granted balance, with the granted balance used first when both are available^[3].

New accounts historically start with a free token grant. DeepSeek reserves the right to adjust prices and recommends topping up based on actual usage while checking the pricing page regularly^[3].