Hook
Four upgrades, one number. Baidu's October 30 product briefing delivered a restructured search box, a visual entry point called AI Lens, a GUI visualization layer, an assistant with more than 170 named scenarios — and a single line item that actually matters: 4.9 RMB per thousand Search AI API calls.
Do the arithmetic. That is 0.0049 RMB per query. Roughly 0.0007 dollars. Cheaper than the coordination overhead on a single Ethereum log read. Cheaper by an order of magnitude than what I pay for the retrieve-rerank-generate chain inside my own monitoring stack.
The invoice does not state what the number covers. Retrieval only? Retrieval plus ranking plus a generated answer with citations? The briefing never says. And that omission, not the price, is the actual signal.
I have a professional reason to care, because I run the thing being priced. Last quarter I rebuilt a retrieval layer for an on-chain liquidity monitor — roughly 2.1 million monthly queries across fifteen chains, feeding a pipeline that flags migration events before they surface on aggregator dashboards. Fully loaded, my cost per query sat near 0.004 dollars. Baidu's headline sits roughly 5.7 times below that. Either the unit economics of this industry changed overnight, or the accounting is hiding in a footnote that nobody has published yet. I have watched enough protocol announcements to know which of those two is more common.
Context
The four product moves are engineering, not architecture. Search-box Agent-ification, a visual entry point, GUI presentation, assistant interaction — all of it lives in the application layer. No model version disclosed. No parameter count. No architecture change. No training or inference optimization. Mapped to any reasonable innovation ladder, these are composition-level and engineering-level changes. Not foundation-level.
Then there is the phrase "recursive self-improvement."
In the AI safety literature, that term carries a strict definition: a system that autonomously improves its own intelligence design, closing a capability flywheel that accelerates without human input. Deployed against a search product's feedback loop, it is a category error wearing a lab coat. What is being described is a conventional data flywheel — user behavior feeding ranking and answer-quality tuning — rebranded with vocabulary borrowed from AGI discourse. The borrowing is not neutral. It inflates what the product is while degrading what the term means.
One more thing about provenance. This briefing reached me in English, through a crypto newswire, republished from a Chinese AI product announcement. That is a domain mismatch with real consequences. A crypto desk does not have the reference points to flag a misused safety term, and an AI desk does not have the incentive to report that an API price tag was framed against a competitor's product that may not have a published price at all. Translation plus specialty mismatch is how a claim becomes a fact in eleven hours.
I have seen the same failure mode from the other side. In early 2022, I modeled the interdependencies of an algorithmic stablecoin while the public narrative was still describing it as uncollateralized innovation. My liquidity model showed a shortfall measured in billions of dollars. The published narrative lagged the on-chain reality by roughly eleven days. Eleven days is the entire gap between a warning and a postmortem. Anyone reading a translated press briefing should hold the same suspicion: the useful information is not in the sentence that was written, it is in the paragraph that was skipped.
The restated facts are simple. Four product upgrades. A developer API with an aggressive price. A consumer assistant with 170+ named scenarios. No model card, no benchmark, no evaluation set, no pricing documentation. Everything is a claim except the price, and the price is unmoored from its basis. Treat it accordingly.
Core
Start with the price, because the price is the only falsifiable claim on the table.
4.9 RMB per 1,000 calls. 0.0049 RMB per call. At an exchange rate near 7.0, approximately $0.0007.
Now build the cost stack that any real retrieval-augmented pipeline actually carries:
| Component | Typical unit cost | |---|---| | Retrieval (index lookup, top-k) | $0.0001 – $0.001 per query | | Reranking (cross-encoder over top 20) | $0.0002 – $0.002 | | Generation (~600–900 output tokens) | $0.0008 – $0.012 | | Citation mapping, post-processing | negligible to $0.001 | | Failed-query retry overhead | +10–25% blended |
Generation dominates. On any modern token sheet, 800 output tokens cost more than a thousand index lookups. So a per-call price that undercuts retrieval-only vendors by a wide margin can only be honest if generation is billed separately — and if generation is billed separately, the headline comparison is decorative.
That is the first structural read: retrieval is being sold as free shipping, and the cash register sits at the token meter. This is not a criticism. It is a competent two-part tariff, the same shape every mature infrastructure business has used to seed a downstream metered product. Sample cheap. Meter the loop. The danger is not the structure, it is that nobody is telling you which part of the tariff the number describes.
The second read concerns the benchmark. The briefing frames the price as below a competing search API. I went looking for that competitor's published search price sheet. Public material does not support a standalone, publicly priced search API product of the kind implied. The comparison may reference an unpriced internal offering, a partner integration, or a figure mangled in translation.
Every rug pull has a trail of paid gas. So does every misleading benchmark. When a vendor selects a comparison anchor that the counterparty never published, the anchor was chosen for mindshare, not accuracy. Anchoring to the most-searched AI brand of the quarter borrows heat. It does not establish category parity. The real category here is retrieval infrastructure — the specialized search endpoints, the cloud-bundled retrieval services, the paid crawl-and-answer APIs. That is the price curve being repriced. Not a chat competitor.
Now the part that actually matters to anyone running a data pipeline: lock-in does not happen at the price. It happens at the evaluation set.
I have migrated retrieval providers twice in three years. The migration itself took two days. The validation took five weeks. Here is what a migration genuinely costs a team like mine:
A golden query set. We hold 512 chain-specific questions with verified ground truth — things like which address first funded contract 0x… and on what date, or what was the collateral factor before the governance vote on proposal 41. Each has a defensible answer, checked by hand against a block explorer or a governance forum.
Recall measurement. Did the retriever surface the correct explorer page, the correct postmortem, the correct forum thread — or a plausible-looking summary of a different incident on a different chain?
Hallucination probing. This is where vendors quietly fail. I have watched a model confidently state that a token used 18 decimals when the contract returns 9, because 18 is the statistical prior. A wrong decimals value is not a cosmetic error. It is a factor-of-a-billion error in a supply calculation, and it will propagate silently into every downstream metric until someone reconciles against the contract. Six of the eleven providers I have tested scored above 15% on that probe.
Latency at p95. Irrelevant on a demo stage. Fatal in an alerting loop that must fire inside a nine-minute block window, because at minute ten the liquidity has already moved and the alert is a historical document.
Every prompt template, every chunking heuristic, every rerank threshold is tuned to one provider's quirks. The switching cost is not the API bill. It is the recalibration debt a team carries for a month after the switch.
Which is exactly why the price functions as a strategic instrument and not as a consumer benefit. A 0.0007-dollar query does not save a serious pipeline meaningful money — in mine, the retrieval slice is under a fifth of total inference spend. What such a price does is collapse the activation energy for the thousands of teams that have not yet built a pipeline at all. The target is not cost optimization. The target is greenfield default-setting, and defaults are worth more than discounts.
The third element is the GUI and visual layer.
Agents that perceive and operate a graphical interface were the loudest research theme of the past twelve months, driven by computer-use demonstrations from the major labs. Visual search maps onto multimodal understanding. Both directions align with the global frontier. Neither discloses a technical path: end-to-end vision-language control, or structured interface parsing plus rules? That difference is not semantic. One generalizes across applications. The other breaks the moment a button moves twelve pixels.
I hold a bias here and will state it plainly. In an alerting pipeline, I do not want an agent that reads a screen. I want an endpoint that returns a signed response against a schema. Visual agents are a solution to the problem of humans building interfaces for humans. Where a machine-readable API exists, the GUI agent is a downgrade — slower, less precise, and impossible to unit-test.
So the honest framing of that layer is: infrastructure for the long tail. Markets, dashboards, exchanges, and registries that will never expose an API. The long tail is real and it is enormous. It is also auditable only with difficulty, and that is the problem nobody at the briefing addressed.
If an agent operates an interface designed for humans, its action log is the only forensic record that exists. During my 2017 audit of a token migration contract in Estonia, the evidence chain was fourteen exchange withdrawal logs and a set of exact contract calls. All of it was reproducible. A GUI agent's click on a rendered button is not reproducible. It is an interpretation. The moment capital moves through interpreted clicks rather than signed transactions, the audit trail degrades from a ledger into a transcript. That is a future problem, not a reason to reject the product. It is a reason to demand an action-log standard before the volume arrives, while the format is still negotiable.
And back at the commercial layer, the disclosure gap remains wide. Two channels — an API for developers, an assistant for consumers, 170+ scenarios bridging them. No free tier published. No uptime commitment. No rate-limit ceiling. No enterprise pricing. No gross margin. No call volume. A pricing announcement with no volume data is half a sentence, and the missing half is the one that determines whether the price survives contact with a P&L.
Contrarian
Here is where I think the consensus reading is wrong.
The prevailing interpretation is that this pricing is a bid to become the default retrieval layer of the agent economy — the moment any agent needs to know something about the world, it calls one API, and the position becomes self-reinforcing, exactly as a default search engine did in 2008.
I do not think that door closes the way the bulls expect. Agents that touch on-chain state do not search. They query. They call RPC nodes, read contract storage directly, subscribe to indexers, pull from subgraphs, decode event logs. There is no vector retrieval anywhere in that loop. A liquidation bot does not search for the price of ETH. It reads an oracle feed and a health factor, in under a second, deterministically.
Search enters only at the long tail: news, documentation, forum sentiment, regulatory text, exploit postmortems. Necessary, certainly. Peripheral to the state machine. So the addressable prize is narrower than the narrative implies. Retrieval sits at the edge of the agent pipeline, not the center. The center is compute and indexed state. A vendor that wins retrieval has bought the periphery, and it has paid for it with a loss leader.
Meanwhile the damage from this price lands somewhere the briefing has no reason to mention. Specialist retrieval vendors have no model business against which to amortize a loss leader. Their unit economics are crawler infrastructure, index maintenance, and reranking — nothing else. A hyperscaler pricing retrieval below cost does not need to win those customers. It only needs to make the segment uninvestable. That is what price wars do at the infrastructure layer: not conquest, attrition. The casualty list is written in the vendors who quietly stop publishing a status page.
And the question the briefing skips entirely is self-cannibalization. Answer-first search reduces outbound clicks. Search advertising is billed on outbound clicks. Every incumbent search business on earth faces that contradiction and none has resolved it. Accelerating an answer-first product is a defensive self-revolution conducted while the revenue base being eroded is still funding the transition. That is a legitimate strategic choice. It is also a choice made under duress, and it should be read as one.
Which gives the honest characterization of the whole package: a product-and-commercial move executed under revenue pressure, wrapped in research vocabulary. Not a technical milestone. Not a capability breakthrough. Four application-layer integrations, one aggressive two-part tariff, and a narrative that borrowed a term from a field where the term means something else entirely.
I say this knowing what the alternative looks like. Every quarter a protocol team sends me a deck claiming to have solved something. Nearly all of them cite a metric nobody can independently reproduce. The ones that survive my skepticism share exactly two properties: an auditable number and a stated basis. Baidu supplied the number. The basis is still missing. That is not fraud. It is incomplete disclosure, and the market will fill the gap whether the company does or not — usually with the least charitable interpretation available.
Takeaway
When the official API documentation lands, and it will land within weeks, three items decide whether this is a genuine repricing or a marketing artifact. The billing basis: does 4.9 RMB per thousand calls include generated tokens, and if not, at what rate are those tokens metered? The operational envelope: free tier, rate ceiling, uptime commitment, overage ladder. The repricing clause: whether the agreement permits unilateral price changes, and on what notice. Read those three before reading any benchmark, because a price without a basis is a slogan with a decimal point.
The forward question is not whether retrieval becomes cheap. It will. Retrieval is heading toward zero the way block space commoditized and the way data availability commoditized after the last upgrade cycle. The question is what a fact costs in 2027, once the retrieval layer has been priced to nothing and the only surviving margin sits in the generation meter. If facts become the loss leader and synthesis becomes the toll booth, then the ability to know things quietly becomes a metered service — and the people who can afford to verify will be the people who can afford to ask. That is a market structure worth watching with the same cold eye we reserve for liquidity, because it fails the same way: slowly, then all at once, with everyone insisting the books were fine.
We followed the queries, not the promises. The queries are only now beginning to clear.
