The smallest number on DeepSeek’s pricing page applies only when the input is already in cache. A miss costs fifty times more before the model writes a token.
DeepSeek V4 Flash entered public beta through the official API on July 31. DeepSeek says the 0731 build keeps the Preview architecture and size, changes only post-training, adds native Responses API support, and is adapted for Codex-style agent work.
The release pairs a one-million-token context window with unusually low published prices: $0.0028 per million cached input tokens, $0.14 per million uncached input tokens, and $0.28 per million output tokens. The cache-hit price is the eye-catching number. It is not the safe default for a budget.
The cache miss is fifty times the cached-input rate
DeepSeek’s current pricing page makes the distinction explicit. Uncached input is $0.14 per million tokens, exactly 50× the $0.0028 cached rate. Output is billed separately at $0.28 per million tokens whether the prompt hits the cache or not.
That means the cheapest published rate belongs in a cache-performance model, not in a generic “cost per request” headline. If prompts change, cache keys churn, or the provider does not reuse the prefix you expect, the uncached rate applies.
A 10,000-input, 2,000-output request shows the gap
| Scenario | Input cost | Output cost | Total per request | 10,000 requests |
|---|---|---|---|---|
| Input fully uncached | $0.001400 | $0.000560 | $0.001960 | $19.60 |
| Input fully cached | $0.000028 | $0.000560 | $0.000588 | $5.88 |
In this simplified example, perfect input-cache reuse saves $13.72 across 10,000 requests. Output dominates the fully cached request, accounting for about 95% of its cost. The calculation excludes retries, tool calls, longer reasoning traces, image or file handling, and any future price change.
The benchmark table needs its harness attached
DeepSeek reports 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, and 70.3 on the verified Toolathlon split. It also reports 25.2 on Agent Last Exam and 25.1 on Automation Bench Public.
Those are company-reported results. The changelog says they used a minimal DeepSeek Harness with maximum effort, top-p 0.95, and temperature 1.0. DeepSeek says the harness will be released soon. Until it is public, outsiders cannot reproduce the exact agent wrapper from the official record.
The company also lists internal DSBench FullStack and Hard scores. Those results can inform a test plan, but an internal benchmark without a public task set and harness is not independently verifiable.
Only the Flash API changed
The official update says V4-Flash-0731 is the public-beta change. V4 Pro, the consumer app, and the web product were not updated with this release. Teams should avoid using a Flash API benchmark to describe every DeepSeek surface.
The API page lists one million tokens of context, up to 384,000 output tokens, and concurrency of 2,500. These are service specifications, not a recommendation to fill the window. Very long prompts increase latency, reduce cache predictability, and can amplify the cost of a miss.
Build a budget from cache behavior, not hope
- Log billed token classes: separate cached input, uncached input, and output in every cost report.
- Measure prefix reuse: record the real cache-hit rate for each workflow and prompt version.
- Model a cold path: budget the first request, changed system prompts, and cache eviction at the uncached rate.
- Cap agent loops: retries and tool calls can repeat both input and output charges.
- Recheck pricing: DeepSeek warns that a significant increase is expected and says the final plan will be announced separately.
Our review of GPT-5.6 pricing changes uses the same request-level discipline: translate token rates into a workload before comparing providers. Our guide to choosing an AI model explains why cost, latency, failure recovery, and task quality belong in one decision.
My verdict: test Flash now, price it at the miss
DeepSeek V4 Flash is worth testing for agent workflows because the official API combines a long context specification, Responses API support, and low current rates. The published benchmark table is useful for selecting tasks to reproduce, not for skipping evaluation.
For planning, use the uncached input price until your logs prove a stable hit rate. Treat savings from cache reuse as measured upside. That approach survives prompt changes, cold starts, and the price increase DeepSeek says is coming.
Read the primary documentation
- Read DeepSeek’s official API changelog.
- Check the current model specification and pricing page.
What is your measured cache-hit rate after a system-prompt change or a cold start?
Checked August 11, 2026. Release scope, API support, context and output limits, concurrency, benchmark results, harness settings, and prices come from DeepSeek’s official documentation. Benchmark figures are company-reported. Cost examples are Musthave.ai calculations and assume the published rates remain in effect.
