DeepSeek put a meter on the free part, and swapped the model under a fixed ID

DeepSeek shipped two things on 17 October 2026 and framed both as good news. The changelog entry for DeepSeek-V4.1-Flash-1017 calls it a “Performance–cost Pareto improvement” and reassures developers that “the API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model.” The second entry announces that DeepSeek is “opening up” its in-house search engine as an API.

Read as line items rather than as release notes, those are the same sentence twice: a cost you were not tracking is about to appear, and the thing you are buying changed without the identifier changing. Neither is a price rise. Both land on your invoice.

The free part now has a meter

DeepSeek’s Search and Fetch APIs are now standalone, billed per request at $0.75 per 1,000 successful requests — charged separately for each endpoint. Until now they have been free, and crucially they have not been free unused.

DeepSeek’s pricing page is explicit about who has already been calling them: server-side tool calls from DeepSeek Harness, Claude Code, or Codex already hit these APIs and are “currently free of charge.” Billing begins 31 October 2026 at 16:00 UTC — the same moment the pricing page expresses as “00:00 Beijing time on November 1, 2026,” which is the identical instant rather than a discrepancy.

That means the affected population is not “people who decided to buy a search API.” It is everyone running an agent harness against a DeepSeek key with web tools switched on, who in thirteen days starts paying for an endpoint they never chose and may not know they are calling.

The arithmetic the announcement does not do

Per-call pricing invites you to wave it away. $0.00075 is a rounding error. Compare it to the tokens, though, and the ranking inverts.

Take one agent task that searches the web twenty times and fetches twenty pages — unremarkable for research or debugging work:

Component	Volume	Off-peak cost
Search calls	20	$0.015
Fetch calls	20	$0.015
Input tokens (cache miss)	100,000	$0.015
Output tokens	10,000	$0.006
Total		$0.051

The tool calls are $0.030 of a $0.051 task — more than the model. Rates from DeepSeek’s pricing page, checked 18 October 2026: deepseek-flash bills $0.15 per million input tokens on a cache miss and $0.60 per million output tokens off-peak, with peak windows doubling both.

Two things follow. First, DeepSeek’s headline advantage is a token advantage, and this change moves a meaningful share of agent spend outside the thing DeepSeek is cheapest at. Second, the usual optimisations stop helping: prompt caching cuts input tokens to $0.003 per million, which makes a well-cached agent’s bill more dominated by search calls, not less. Off-peak scheduling does not help either — the $0.75 rate carries no peak/off-peak split, so the one lever DeepSeek buyers rely on does not apply to this line at all.

If you run agents on DeepSeek, the useful thing to do before 31 October is not a price comparison. It is an inventory: pull your harness logs, count actual search and fetch calls per task, and multiply. That number has been invisible because it was free.

The model changed underneath a fixed name

The second half is a lifecycle problem this publication keeps writing about, now with a third vendor.

deepseek-flash is unchanged as an API model ID. The model version behind it moved from V4.1-Flash to V4.1-Flash-1017, and it is not a small move. On DeepSeek’s own published figures, comparing the 10 September launch entry with the 17 October one:

Benchmark	V4.1 Flash (10 Sep)	V4.1-Flash-1017	Change
HLE (no tools)	36.8	42.6	+5.8
Terminal-Bench 4.0	31.2	42.3	+11.1
DeepSWE v1.1	74.2	77.0	+2.8
ProgramBench	20.3	29.1	+8.8

DeepSeek also states that runtime and output length were reduced. For a model billed per output token, shorter outputs are an unannounced price cut — real, welcome, and equally unpinnable.

That is the issue. A double-digit jump on Terminal-Bench is exactly the kind of change that invalidates an evaluation suite, a prompt library tuned to the old model’s verbosity, or a latency budget. And there is no dated model string to pin: nothing you can send to the API asks for the September build. The pattern is familiar — OpenAI shipped an image-encoding fix with no dated snapshot to reproduce the pre-fix behaviour, and Google rerouted deprecated Gemini Flash models silently.

DeepSeek has one defence the other two do not: the weights are published. The open-weights repository is named DeepSeek-V4.1-Flash-1017, so a team that needs a frozen build can self-host a named release rather than negotiate with a vendor’s roadmap. That is the concrete, under-appreciated value of open weights, and it is the same property that makes DeepSeek the sensible choice for anyone whose compliance posture cannot absorb a silent model swap.

Also worth noting in the same breath, because it is the one place DeepSeek’s lifecycle hygiene is good: the retired deepseek-v4-flash and deepseek-v4-flash-vision-exp names still resolve, are served by V4.1-Flash-1017 and bill at Flash rates. Nothing breaks — but anyone still sending those strings is silently on a different model than the name says, two generations removed from what they originally integrated.

Tokens are no longer the whole API

It is worth putting the $0.75 next to the only other metered harness surface a buyer can price today. OpenAI’s Agents API, in public beta since 10 September 2026, charges nothing for the harness itself and bills OpenAI-hosted sandboxes at $0.09–$5.76 per hour of wall-clock time, with a non-configurable one-hour idle deletion. Different unit, same structural move: the orchestration layer around the model becomes a separate, metered product.

The two are priced on opposite axes, which is the useful part. OpenAI bills time — an agent that sits idle mid-task still accrues. DeepSeek bills actions — an agent can run for an hour for free and then cost three cents the moment it searches. Neither charge appears in a per-million-token comparison, which is still how almost every “cheapest model” table on the internet is built, including the ones that put DeepSeek at the top.

For anyone budgeting agent work in 2027, the practical consequence is that a model’s token price is now one of three or four lines, and it is increasingly not the biggest. Our Codex review works through the same problem on the OpenAI side, where the subscription, the API key and the hosted sandbox are three different prices for one product.

What to do before 31 October

Count your search and fetch calls. Per task, per day. If you run Claude Code or Codex against a DeepSeek key with server-side tools enabled, you are in scope whether or not you configured anything.
Re-baseline evals. The model changed on 17 October. Scores from before that date describe a build you can no longer request.
Upgrade the harness. DeepSeek asks for DeepSeek Harness 0.2.1-rc.1 or later for the new build.
If you pin builds for compliance, mirror the weights. The API cannot express a version; the Hugging Face release can.
Re-check the tool-call line against alternatives. DeepSeek’s token prices remain among the lowest published, and that has not changed — but a search-heavy agent is now buying two products from DeepSeek, and only one of them is cheap for structural reasons.

The broader read: as agent harnesses standardise on server-side tools, the metered surface of an “API” stops being tokens alone. DeepSeek is the first to put a number on its search endpoint and tell you which third-party harnesses were already using it. It will not be the last, and the free tier that quietly becomes a line item is going to be a recurring 2027 story. Our DeepSeek review carries the current rate tables, and the API cost calculator covers the token side.

Frequently asked questions

Did DeepSeek raise its prices on 17 October 2026?

Not on tokens. DeepSeek's own changelog says the calling method and pricing are unchanged, and the pricing page confirms deepseek-flash still bills $0.15 per million input tokens on a cache miss, $0.003 on a cache hit and $0.60 per million output tokens off-peak, with peak rates exactly double. What changed is that two endpoints which had been free — the Search API and the Fetch API — get a price on 31 October 2026 at $0.75 per 1,000 successful requests each. If your usage is pure text generation, nothing about your bill changes. If your agent searches the web, it does.

When exactly does Search and Fetch billing start?

One instant, stated two ways in DeepSeek's own documentation. The release post says billing starts 31 October 2026 at 16:00 UTC; the pricing page says per-request billing starts at 00:00 Beijing time on 1 November 2026. Beijing is UTC+8, so those are the same moment and not a contradiction. Both figures are $0.75 per 1,000 successful requests, charged separately for Search and for Fetch, and only successful requests count.

Does this affect me if I use Claude Code or Codex rather than DeepSeek's own harness?

Yes, if those harnesses are pointed at DeepSeek and using server-side tools. DeepSeek's pricing page names DeepSeek Harness, Claude Code and Codex specifically as clients whose server-side tool calls already hit the Search and Fetch endpoints, and says those calls are 'currently free of charge' with per-request billing starting at the deadline above. The harness is not the thing being billed — your DeepSeek API key is. So a Claude Code or Codex setup running against DeepSeek acquires a new line item without any change on your side.

How much will Search and Fetch actually add to a task?

Enough to exceed the model cost on search-heavy work. At $0.75 per 1,000 calls each endpoint costs $0.00075 per request, so an agent issuing 20 searches and 20 fetches in one task pays about 3 cents in tool calls. The same task at 100,000 input tokens on a cache miss and 10,000 output tokens costs about 2.1 cents in tokens off-peak. The tool calls are the larger half of the bill. Cheap per call does not mean negligible — it means the cost moved from a line you were watching to one you were not.

Can I pin the previous V4.1 Flash build?

Not through the API. DeepSeek changed the model version behind deepseek-flash from V4.1-Flash to V4.1-Flash-1017 without introducing a dated model ID, so there is no string you can send to request the build you benchmarked in September. The one real escape hatch is that the weights are published — the open-weights repository is named DeepSeek-V4.1-Flash-1017 — so a team that genuinely needs a frozen build can self-host a specific release. That option does not exist with OpenAI or Google when they do the same thing.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.