DeepSeek Flash-1017: Cheap Tokens Are Only Part of the Bill
![]() |
| AI-generated illustration for Routes, Bytes & Budgets. The task log is a concept interface, not a DeepSeek screenshot or benchmark. |
Cheap AI tokens catch my attention. But the number I actually care about is how much it costs to finish something useful: a checked research summary, a working script or a piece of admin that no longer needs my attention.
That is why DeepSeek’s October 17 update interests me. DeepSeek-V4.1-Flash-1017 promises better results with less waiting and shorter output. Those improvements could matter more than a tiny price on a comparison chart. They could also disappear from the budget if the task needs several attempts, a pile of web searches and an afternoon of corrections.
My starting point is simple: compare the bill for a completed task, then decide whether switching tools makes sense.
What has actually changed?
DeepSeek says the new Flash model improves benchmark performance, task completion speed and output efficiency. It is available through its website, app and API. The company’s release notes also recommend updating DeepSeek Harness to version 0.2.1-rc.1 or later.
I find the emphasis on shorter, more effective work encouraging. A model that reaches a usable answer in fewer steps has an obvious practical advantage. But these are DeepSeek’s results and claims, not measurements from my own workflow. A coding benchmark also cannot tell me how well it will handle the particular documents, writing or travel research I give it.
There is another distinction worth keeping clear: the prices below are API usage prices. They are not a monthly subscription comparison, and a third-party application may add its own charges.
The token price depends on when and how you use it
On October 19, DeepSeek’s official pricing lists the following rates for deepseek-flash, which currently serves Flash-1017. All figures are US dollars per one million tokens.
| Usage | Off-peak | Peak |
|---|---|---|
| Input served from cache | $0.003 | $0.006 |
| Input not served from cache | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. In Beijing and Shanghai, that is 09:00–12:00 and 14:00–18:00. Other hours, weekends and Chinese public holidays are off-peak.
That schedule matters because it overlaps with an ordinary working day in China. Moving a non-urgent batch to the evening can reduce its token bill, but delaying urgent work to save fractions of a cent makes little sense.
Nor do I assume every input token gets the lowest rate. DeepSeek’s caching documentation explains that reuse depends on matching persisted input prefixes. Cache hits are best-effort, not guaranteed. I want the usage report to show the actual split.
Searches and page retrieval have their own bill
The Search and Fetch documentation prices each service at $0.75 per 1,000 successful requests. That is $0.00075 per successful call. Search finds pages; Fetch retrieves a page’s contents.
The charge is per successful request, not per search result. Failed requests, timeouts and empty search results are not billed under the documented rules. Retrieved text can still contribute to model input usage when the assistant reads it.
There is a temporary exception: specified server-side Search/Fetch calls made through DeepSeek Harness, Claude Code or Codex are currently free for DeepSeek API users. Billing begins on October 31, 2026, at 16:00 UTC, which is November 1 at 00:00 in Beijing and Shanghai. This is not a blanket free period for every use of the standalone APIs.
For a tool comparison, I therefore want both today’s bill and the bill after that waiver ends. A short promotion should not decide a longer-term setup.
A small example makes the difference clearer
Imagine one task uses 50,000 uncached input tokens, 10,000 output tokens and 20 successful Search/Fetch calls combined. This is an illustrative budget, not a benchmark or a claim about normal usage. Applying the standard tool charges gives:
| Cost component | Off-peak task | Peak task |
|---|---|---|
| 50,000 uncached input tokens | $0.0075 | $0.0150 |
| 10,000 output tokens | $0.0060 | $0.0120 |
| 20 successful tool calls | $0.0150 | $0.0150 |
| Total | $0.0285 | $0.0420 |
Both totals are small. The useful detail is that the tools cost more than the off-peak tokens in this example. A token-only comparison misses that completely.
If the whole attempt has to be repeated three times with the same usage each time, those totals become $0.0855 and $0.1260. Real retries can be cheaper through caching or more expensive because they carry longer histories. Either way, I count every attempt, including the unsuccessful ones.
My switching checklist starts with the result
For a fair comparison, I want a small set of representative tasks and a clear definition of success before running them. For research, that might mean an answer with working references and no unsupported claims. For code, it might mean a script that completes the actual job and passes a useful check.
- Count accepted results. Divide total spending across the trial by the number of tasks that genuinely passed review.
- Record the whole route. Include retries, tool calls, fallback models and any application fees.
- Record my time separately. Five minutes checking a result is different from thirty minutes repairing it, even when both API bills look tiny.
- Keep the conditions comparable. Use similar tools, instructions, time limits and spending caps.
- Check what moves with the switch. Documents, credentials and connected services deserve as much attention as price.
This also connects with my interest in smaller local AI on existing devices. Local processing changes the privacy and cost trade-offs, but it brings its own setup and hardware constraints. There is no single price label that settles every case.
Cheap enough to explore, useful enough to keep?
Flash-1017 looks worth considering because the potential savings are substantial at the API level. My enthusiasm depends on whether those savings survive the entire job.
I do not need the cheapest answer if I have to ask for it five times. I want the least expensive dependable way to get the work finished, with a bill I can understand and limits I can control.

Comments
Post a Comment