The tool layer now has a price list

2026-11-14 · Search, fetch, OCR and speech now come with a unit price — what that changes

On 13 November 2026 DeepSeek did something quietly more consequential than a model release. It took the capabilities its own chat app had been assembling behind the scenes — search, multi-step deep search, document parsing, URL reading, a safety check — and put them on the API with a unit price per call. Search and fetch at 5 CNY per 1,000 successful requests, OCR at 1.5 CNY per 1,000 pages, speech recognition at 0.39 CNY per hour during the current half-price promotion, synthesis at 1.5 CNY per 10,000 characters. Tokens are billed separately.

Read that as a sentence about agent architecture, not about DeepSeek. Capability is no longer something you assemble from vague promises. It has a price tag, a billing unit and a success condition.

What one yuan buys

About 200 page fetches.
5 CNY per 1,000 successful requests, and failures or empty results are free. An agent that reads 200 pages while researching costs one yuan.
About 660 pages of OCR.
1.5 CNY per 1,000 pages, whether the input is a scanned PDF, a Word file or a photo of a page. The pipeline work that used to justify a small project is now a rounding error.
Two and a half hours of audio transcribed.
0.39 CNY per hour at the promotional rate — 0.78 list. Whole courses, whole meetings, whole voice-note archives.

Stack those against what the work used to cost in human hours and the interesting question stops being "can the agent do this" and becomes "how many of these calls per run can I afford".

Three things a meter changes

Cost per run becomes computable. Model tokens were always one number; now the second number is the tools. An agent that makes 40 search calls and reads 60 pages costs 0.5 CNY in tools before the model has said anything. That is not a reason to avoid search — it is a reason to know which loop in your agent is expensive. Most agents waste tool calls on retries and on reading pages they already had.

Failures being free is an invitation, not a discount. If unsuccessful requests are not billed, the correct design is to retry, widen and fall back aggressively rather than handle every failure with a fragile one-shot. Cheap failure is a design constraint you should exploit deliberately.

The build-or-buy line moved. A crawler, a PDF pipeline and a transcription queue used to be the kind of thing an agent builder wrote themselves — a week each, then maintenance forever. At these prices the honest answer for small teams is usually "buy, and spend the week on the part that is actually yours".

The part the price list does not solve

A single key that reaches search, the open web, any uploaded document and any audio is a large blast radius. The metering is fine; the scoping is unchanged. What a tool can touch, where approval happens, and whether it can be pulled without breaking everything else — those questions do not get cheaper with a lower unit price. If anything, a cheap generic tool is the easiest one to connect and forget, and forgetting is where agents get hurt.

DeepSeek's own note is blunt about the obligation side: applications serving the public should enable the safety-check parameter, and the operator still carries content review, model registration and AI-labelling duties. Cheap capability does not transfer responsibility along with it.

One last thing worth stating plainly, since this is an MCP directory: these are REST endpoints, not MCP servers. Someone will wrap them, and when they do, each wrapper belongs in this list — the directory is organised by tool, not by vendor. Until then, the takeaway is the same one the crime logs taught, now with a receipt: the model is the cheap part. What your agent can do is the tools you gave it, and those now have a price you can read.

Sponsored ranking

No bids yet — first bid is $5 and takes row #1 · 0/10 rows filled
Read the rules before you bid
  1. This is an auction ranking, not a fixed block of slots. Highest cumulative bid is #1; rows go down to #10.
  2. First bid is $5. Outbid the leader by $1 and you take the top when the 24-hour protection window ends.
  3. An outbid bid is not lost — it stays on the board behind the leader, which is why this is a ranking.
  4. Every paid row is labelled Sponsored. The game leaderboard is separate and cannot be bought.
Full rules and how to bid →

More notes