Finance and markets on x402: three engines, four jobs, one price

Financial analysis is sold on x402 three ways: a compute call, an assembled role, and a desk that runs three views on one brief. The model is the cheap part. What separates the three is what gets computed, who reconciles it, and what a month costs.

The job before the tool

A finance question arrives with an input attached. A company’s statements and a price. A curve snapshot and a release file. An OHLC series and a book. The work is to compute from that input a set of figures the reader can check, then write around them without quietly inventing the ones that are missing.

That is a narrower task than it sounds, and it is why the pricing on this network looks odd at first. The compute is sold by the operation, at one price, whatever the operation does. The prose is sold by the model call, and the model call is cheaper than most of the compute it explains. And the product a buyer actually wants — one view, computed from their own data, with the assumptions named — is assembled from both, which means the cost of the product is a small arithmetic sum rather than a subscription.

This pillar covers the four jobs on the hub that do that work for money: three roles, each with one view, and one desk that runs all three and reconciles them. It prices each of them at measured rates, states what the underlying compute and model calls cost, and says plainly what none of them will do. It contains no forecast, no recommendation to buy or sell anything, and no claim about the quality of the analysis.

Three engines at one price

Three services supply the compute. They are macro-analyst, financial-analyst and trade-analyst, and each is a deterministic toolkit: the caller names the operation, supplies the data where the operation needs it, and gets figures back with the formulas used. Table 1 lists them with the price and the state.

Table 1. The three analyst services, their paid operations, what they compute, the exact price of one basic operation on Base, and whether the service answers today. Source: catalog.json (updated 31 October 2026) and each service page, read 31 October 2026.

The exact figure is the Base USDC/EURC price for one call, and it is the same $0.006125 for all 45 basic operations across the three services. Nothing in the price distinguishes a discounted cash flow from a percentage change: the unit is the call, not the difficulty.

Two things follow from that flat price, and both matter to a buyer.

The first is that the expensive part of an analysis is not any single computation. A valuation that runs statements, multiples, a DCF and a cross-check is four calls and $0.02450 of engine time. A macro note that reads the curve, the breakevens, the inflation print, growth and FX is five calls and $0.030625. The bill is small because the operations are small: each is one deterministic calculation on data the caller already holds.

The second is that a service which is pending review is a price without a counter. All three analyst services are in that state: the pages publish the price and the operation list, and each one says in the same breath that nothing answers for it while it is pending review. The price is real — it is what the operation costs when it is bought over the wire — but no 402 handshake will quote it today. That distinction is worth holding on to, because every figure later in this piece that uses $0.006125 is a price for a capability that is not yet callable, and the assembled products that ship those same engines do not pay it at all. More on that below.

There is one operation on each service that breaks the flat price. The assembly calls — ma_ask, fa_ask, ta_ask — are priced at $0.256125 on Base, which is 41.8 times a basic operation. What they buy is different in kind: one call in which DeepSeek orchestrates the deterministic operations, supplies its own data, and returns a structured brief. It is the only product on the three services where the model and the compute arrive together. Whether that is worth 41.8 operations depends on whether you hold the data; if you do, the operations are cheaper and more checkable. If you do not, the assembly call is the only route that answers.

What each engine computes

The operation lists are long enough to be worth reading before the price, because the difference between these three services is not what they cost. It is what they can be asked.

Macro. The macro engine is arithmetic on series the caller supplies. It publishes eight compute operations, a validation call and two reference helpers. yield_curve returns the level, the 2s10s, 3m10s, 3m2s and 10s30s spreads, a shape label, and — where a prior curve is supplied — the per-tenor shift and steepener change. breakeven_inflation returns breakevens per tenor, the 5y5y forward from the 5-year and 10-year pairs, real yields and an expectations label. inflation_decomp splits a monthly series into the headline–core gap, MoM annualised, 3m, 6m and 12m annualised, the base effect and component contributions. growth returns year-on-year, period-on-period and quarter-on-quarter SAAR. fx returns the move against a prior spot, annualised carry from the rate differential, and a real-FX proxy. rates_math returns an implied forward, zero-coupon modified duration and DV01 on $100 face. real_commodity deflates a price by a CPI series and returns the real return and the distance off the all-time real peak. macro_stats covers z-scores, percentile ranks, drawdowns and CAGR.

Financial. The financial service is wider, because it holds both data and compute operations. The compute side carries statements (common-size, ratios, a five-way DuPont, free cash flow, the cash conversion cycle), multiples (EV assembly and valuation multiples with a market-cap reconciliation), dcf (a ten-year explicit FCFF path, the present-value-versus-terminal split, a WACC by growth sensitivity grid, and a reverse mode that returns the market-implied terminal figure), cross_verify (market cap, DuPont ROE, EBITDA bridge and EPS bridge, each against the contract’s hard-fail thresholds), risk (CAGR, volatility, Sharpe, Sortino, maximum drawdown, beta, alpha) and factor (correlation, covariance, a single-factor regression, z-scores, rolling statistics). Its data operations fetch a quote, a price history, a benchmark series, a risk-free rate, headlines, a screener resolution and exchange filings.

Trade. The trade service is the same shape at a shorter horizon. indicator computes from a supplied OHLCV envelope — moving averages, RSI, MACD, Bollinger bands, ATR, ADX, Supertrend. volume covers OBV, VWAP, MFI, accumulation/distribution, a volume moving average and CMF. orderbook takes a level-2 snapshot and returns spread, mid, depth imbalance and a weighted mid. signal reduces the readings to a composite. backtest runs a declarative strategy over a supplied series with an optional cost model, executes on the next bar’s open rather than the signal bar’s close, uses whole-share accounting, and marks to market daily. options_brief works an option chain into an ATM straddle expected move, an implied-volatility and delta ladder, a 16-delta strangle and a sample iron condor with its credit and maximum loss. verify compares two independent values within a stated tolerance.

Three properties of that list are worth more than any single operation.

The first is that the compute operations calculate over inputs and never fetch. For macro that is the whole service, and it is why a macro note’s evidence is the caller’s own files. Financial and trade do fetch, on their data operations, and those are the calls whose output needs a second observation before it is used; the roles treat filings as the primary document and an aggregator as a convenience.

The second is that the operations return the formulas they used. A discounted cash flow that arrives with its explicit path, its terminal split and its sensitivity grid can be re-derived by a reader who disagrees with it. That is a different product from a paragraph of prose containing a number.

The third is that the split between measurement and interpretation is visible in the interface. indicator measures; signal composes; the brief interprets. A buyer can take the first without the third, and a reviewer can check the first without accepting the third.

The model call: $0.00700 against $0.00875

The prose on top of those figures is bought from an inference endpoint, and on this network there are two worth pricing: DeepSeek and Qwen. Both answer POST /v1/chat/completions in the OpenAI request and response shape, neither asks for a key or an account, and both quote their price in the 402 response before anything is signed.

Measured on 31 October 2026, on Base, with one short prompt and max_tokens: 1000, the two do not quote the same amount. DeepSeek quotes $0.00700 for deepseek-v4-flash and a metered ceiling of $0.00955. Qwen quotes $0.00875 for qwen3.8-flash, with a ceiling at the same $0.00875.

The gap is 25 per cent, and it contradicted the sentence on Qwen’s own service page, which described qwen3.8-flash as billing "on the DeepSeek tier (same price as deepseek.x402.press)". It did not, and that page has since been corrected: the handshake is the price. One qualification belongs beside the figure, because it is easy to over-read. The comparison is against DeepSeek’s cheapest quote — an off-peak call with a short prompt. DeepSeek prices a long prompt, or a call inside its weekday peak windows, above $0.00955, which is more than Qwen’s flat $0.00875. The 25 per cent describes the floor of one service against the flat rate of the other, not a constant. The same 25 per cent gap appears on every chain the two answer for — $0.00250 against $0.00200 on Avalanche, $0.00625 against $0.00500 on Polygon, $0.00750 against $0.00600 on Arbitrum — which is what a tier-level markup looks like rather than a quirk of one quote.

The catalogue adds a second wrinkle worth knowing before you budget. The tier row for the DeepSeek operation in catalog.json carries $0.00875 on Base — the tier price rather than the model price, and 25 per cent above what the endpoint asks. The service page’s price table, which prices per model rather than per tier, carries $0.00700 for deepseek-v4-flash and $0.01325 for deepseek-v4-pro, which is what the handshake confirms. Budget from the handshake. Twelve drafting calls at $0.00700 is $0.084; the same twelve at the $0.00875 the tier row shows is $0.105, and that $0.021 difference is a 25 per cent error repeated on every month of the year.

The 25 per cent holds at one prompt length, and only at that one. Lengthen the prompt and the ordering moves. At roughly a thousand tokens DeepSeek still quoted $0.00700. At about 3,000 tokens it quoted $0.00955: 36 per cent above its own short-prompt amount, and 9.1 per cent above Qwen, which held $0.00875 at 1,000, 3,000, 10,000 and 30,000 tokens alike. The boundary sits somewhere between 1,000 and 3,000 tokens and we did not locate it.

That matters here more than it would elsewhere. Analysis prompts are long: a valuation that reads a full filing, a curve read that carries a decade of releases, a backtest write-up that reproduces the strategy specification. On that work DeepSeek is the dearer call at $0.00955 against $0.00875, and Qwen is the one to reach for. On a short prompt — a table lookup, a single-figure sanity check, a summary of three lines — DeepSeek is 20 per cent cheaper and Qwen is 25 per cent dearer. Neither endpoint is cheaper than the other; each is cheaper at one length.

The ladder above the entry tier is where the two part company further, and it is the subject of a comparison already published — DeepSeek, Qwen and HyperFusion compared. In short: Qwen’s max tier is 29.27 times its flash tier on Base, DeepSeek’s pro tier is a step of 1.89, and the metered ceiling runs to between 14 and 23 times the exact amount when max_tokens is left out. On the longest prompt tested, the ceilings moved with the prompt as well: $0.01730 on DeepSeek and $0.017040 on Qwen at about 30,000 tokens, against $0.00955 and $0.00875 at 1,000. Two fields, then, set the bill before the work does: max_tokens and the length of what you send.

Neither endpoint publishes a latency, throughput or accuracy figure, and this piece carries none. No benchmark exists on which to rank them, and any sentence claiming one model writes a better valuation note would be invention. The buying decision is on price, model access and the shape of the metering, and on nothing else that has been measured.

A month of one role

A role is the assembled product: one agent, one view, the parts it needs, and a ceiling on what it may spend. The three finance roles are built on the engines above, and each publishes its own monthly arithmetic. Table 2 puts the three side by side.

Table 2. An illustrative month for each finance role, twenty working days, at $0.006125 an engine operation and $0.00700 a model call, both on Base. Counts are planning assumptions, not observations. Source: the role pages, 31 October 2026.

Read across the rows and the pattern is consistent. A note costs between five and six cents in paid calls. A year of one role at this pace costs between $4.96 and $7.39. And the engine lines are between 52 and 64 per cent of the bill — the compute is the larger half, which is the opposite of what most people assume about analysis.

There is a caveat in the other direction, and it is the reason these figures are an upper bound rather than a forecast. The engines ship inside the workspace image and run in-process, so a role does not buy them over the wire. In-process, the engine lines cost nothing and the month is the model calls alone: $0.22400 for the financial role, $0.16800 for macro, $0.22400 for trade. The $0.006125 price is what the same operation costs when it is bought from the service page instead — the assembled agent’s ceiling, not its bill.

The metered ceiling is the third figure to hold in view. At max_tokens: 1000, a DeepSeek flash call may settle at $0.00955 rather than $0.00700, which is 36 per cent more, and at the monthly volumes above that is 13 to 17 per cent on the bill. That same $0.00955 is also DeepSeek’s price on a long prompt, so a month whose drafting reads a full filing lands on the ceiling column rather than the one beside it — or moves to Qwen at $0.00875, which does not step with prompt length. The difference between the two columns in Table 2 is the cost of not setting the field, and of not checking what your prompt weighs.

Three things sit outside every total in Table 2. The document parts — extract, report, data_table, office — have no published per-call price, so the reading, the writing and the rendering are not in the figure; a month of eight notes and twenty-four tables is understated by whatever those parts cost. The counts are assumptions rather than observations. And a month that adds a rates calculation, a factor model or an options read costs more, because each is another operation at the same unit price.

The opening balance is 0.50 USDC in the buyer’s own wallet, credited to the workspace as working capital rather than charged as a fee. At $0.00700 a call that is seventy-one calls: a month of the macro role, or close to one, depending on the pace. The three ceilings — calls per message, spend per message, spend per day — are the limit on what a single instruction can spend.

The desk: three views, one brief

The fourth job is not a fourth view. It is the reconciliation of the other three.

The finance desk takes one brief and works it three ways at once: the macro view computes the curve, breakevens, growth and FX; the financial view values the company from its statements; the trade view measures the price series into indicators, a signal and a backtest with costs. All three run against the same brief and the same as-of date, and each runs independently — no view reads another’s conclusion while it computes. That independence is the design, not a scheduling detail: a view that has seen its neighbour’s answer is not a second opinion but the same opinion restated.

What comes out is a synthesis and a disagreement table. Where the three agree, one line says so and names the inputs behind it. Where they diverge, the divergence is a row: the three values, the input and method behind each, the size of the gap, and whether the gap is one of inputs, method or timing. The note leads with the view the evidence supports, orders the disagreements by what they would change, and hands over to a named reader.

Table 3. One desk brief and one month of desk work, at $0.006125 an engine operation and $0.00700 a model call on Base. Counts are planning assumptions. Source: the finance desk page, 31 October 2026.

The arithmetic is short. Eighty-eight engine calls at $0.006125 is $0.53900; forty-eight model calls at $0.00700 is $0.33600; the two add to $0.87500, which is $0.10938 a brief and $10.50 over twelve months. At the $0.00955 ceiling the model lines rise to $0.45840 and the month to $0.99740, 14 per cent more. In-process, where the three engines are not bought over the wire, the month is $0.33600.

Put that beside the roles and the comparison is not the one it looks like. Buying the macro, financial and trade roles is $1.49800 a month for 232 paid calls, and what it buys is three artefacts that never speak to each other: a note about the economy, a valuation of the company, a brief on the price. A desk brief is 17 paid calls and $0.10938, and its deliverable is the reconciliation — the row where the macro read says the discount rate should fall and the fundamental read has already assumed it will not. The desk is cheaper per question because it asks the question once, and it is a different product because it is the only one of the four whose output includes the argument between its own parts.

The document parts sit outside this total in the same way they do for a role. The desk declares seven parts and no connection: it renders to PDF or DOCX inside the workspace and stops, and a buyer who needs the note to arrive by mail or message adds that at design time rather than assuming it is there.

What the desk is not

The distinction between a role and a desk is easy to blur in marketing and expensive to blur in practice, so it is worth stating three ways.

A role does one view. Ask the financial_analyst for a valuation and it values the company from its statements, runs the cross-checks, labels each figure and hands over the note. Ask the macro_analyst and it computes the curve and the inflation decomposition from your files. Ask the trade_analyst and it measures the series and gives you the level at which the view fails. Each is complete on its own question, and none of them contains the other two.

A desk reconciles three views and names the disagreements. It is not three notes stapled together, and it is not the average of three notes. The distinction is enforced: the desk will not produce a consensus sentence where the three diverge, because averaging is not synthesis — it is a fourth view nobody computed.

Neither is a bundle in the pricing sense. The desk’s price is a sum of calls, and the same eleven engine calls and six model calls bought separately would cost the same $0.10938. What the desk adds is the brief, the coordination and the table, not a discount.

The chain is a bigger lever than the model

Every figure so far is denominated in a chain, and on this network the chain moves the price more than the choice of model does.

DeepSeek’s measured quotes for one deepseek-v4-flash call, short prompt, max_tokens: 1000, on 31 October 2026, run as follows: $0.00200 on Avalanche, $0.00500 on Polygon, $0.00600 on Arbitrum, $0.00700 on Base, and $3.00 on Ethereum. Qwen’s run $0.00250, $0.00625, $0.00750, $0.00875 and $3.75 on the same chains. The analyst services carry the same shape at their own unit: $0.00175 on Avalanche, $0.004375 on Polygon, $0.00525 on Arbitrum, $0.006125 on Base and $2.625 on Ethereum.

Two consequences, both measured rather than argued.

The first is that settling in the cheapest chain on offer is worth more than switching models. Avalanche is 3.5 times cheaper than Base for the same DeepSeek call, and 3.5 times cheaper for the same Qwen call; the model gap is 1.25. A buyer choosing between the two endpoints on Base is optimising the smaller of the two numbers.

The second is the Ethereum multiplier. The same basic operation is $2.625 on Ethereum against $0.006125 on Base: 429 times the price. The same ratio holds for the measured inference calls — $3.00 against $0.00700, again 429 times — which is the signature of a settle cost passed through at a fixed markup rather than a service deciding to charge more. The DeepSeek page says as much: Ethereum’s figure carries the chain’s own settle cost, not the model’s.

None of that makes a chain unusable. It makes the chain a design decision with a bill attached: a treasury that already holds USDC on Base pays Base prices, and one that holds it on Avalanche pays 29 per cent of them. Note also that EURC is accepted on Base and not elsewhere, so a buyer holding euro-denominated stablecoin has one chain rather than five.

Where finance sits in the market

The network’s own index puts the numbers in context. Our pull of the x402 Bazaar on 30 October 2026 listed 105,478 resources, with a median price of $0.01 and a 99th percentile of $0.25.

Set against that distribution, the compute in this pillar is cheap and the assembled call is expensive. A basic analyst operation at $0.006125 is below the median — about three-fifths of it — and so is a DeepSeek call at $0.00700 or a Qwen call at $0.00875. The assembly calls at $0.256125 are above the 99th percentile: roughly forty times the price of a basic operation and twenty-six times the median listing, and the only product in this piece that arrives with the model’s reading already attached.

The comparison is indicative rather than like-for-like. The Bazaar is an index of listings, not of comparable products, and a median across weather, inference, image and analyst endpoints measures the shape of the market rather than the value of any one item in it. What it supports is one modest claim: at the operation level, deterministic finance compute is priced below the middle of this network, and the money in an analysis moves to the assembled product and the model calls around it.

What these jobs refuse

For a buyer of analysis, the refusals are the specification. Anyone can promise a valuation; the useful question is what the product will decline to produce, and the four jobs here are unusually explicit about it. Each refusal below is enforced by the workspace rather than offered as a preference.

No arithmetic in prose. Every multiple, discount rate, ratio and risk figure comes from the engine with its formulas echoed. A number written into a paragraph is not a calculation the reader can check, and where the engine cannot produce one the note says so instead of substituting.

No forecast dressed as a computed figure. The engines calculate over data they are handed. Where the next print matters, a view can be carried as opinion with a confidence level and the condition that would change it — and if that view cannot be kept apart from the arithmetic, it is left out.

No advice. No recommendation to trade, no position, no hedge, no allocation. Nothing in these products is investment, legal or tax advice, and the desks carry no execution part at all: no broker, no exchange key, no order. The decision and the liability stay with the person who has to defend them.

No central bank’s intentions reported as fact. What a committee said is quoted with the date, the speaker and the exact words. What policy is priced to do next is derived from breakevens, forwards and the curve, and labelled market-implied.

No gap filled quietly. A missing input comes back as null with a warning, is written into the coverage block as a stated gap, and the note says which conclusion it would have changed. A price or rate needs a second independent observation within 0.1 per cent, with a timestamp and a staleness test — 15 minutes intraday, one trading day on daily data — and a value that fails is reported as unconfirmed rather than averaged into something cleaner.

No strategy called profitable. A backtest is evidence over one period at stated costs. The brief reports the period, the assumptions and the risk that the rule was fitted to the data it was tested on, and a positive equity curve is a result to be examined rather than a property of the future.

No averaged consensus. The desk’s addition to that list. Where the three views diverge, the divergence is measured and left standing for the reader to decide.

What we did not measure

This piece holds no latency figure, no throughput figure, no uptime figure and no accuracy figure for any of the four jobs, because none is published and none was measured. It does not rank the two inference endpoints on the quality of their output, and it makes no claim about whether a computed valuation is a good valuation. Every figure here is a price, a multiple of a price, or an arithmetic total of prices.

What else a buyer should know is missing:

• The three analyst services are not callable. Their prices come from the published catalogue, not from a handshake, because nothing answers while they are pending review. The model prices are measured from the live handshake; the engine prices are published.
• The document parts are unpriced. extract, report, data_table and office carry no published per-call price, so every total in this piece excludes the reading, writing and rendering — the totals understate the work rather than overstate it.
• The monthly counts are planning estimates. Eight notes, ten setups, eight briefs: plausible months, not observed ones. The unit prices are measured; the volumes are assumptions made to turn them into a month.
• The catalogue and the handshake disagree in two places. The DeepSeek tier row shows $0.00875 on Base against a measured $0.00700 for deepseek-v4-flash, and $3.75 on Ethereum against a measured $3.00. The handshake is the price a client pays; this piece uses the handshake.
• The prompt-band boundary. DeepSeek’s exact amount steps from $0.00700 to $0.00955 somewhere between about 1,000 and 3,000 tokens. The boundary itself was not located, so a prompt near it cannot be priced from this piece. Qwen showed no step up to about 30,000 tokens.
• The Bazaar figures are a single pull. 105,478 resources, median $0.01, p99 $0.25, measured on 30 October 2026, not an audited census.
• The chain ratios are ratios of quotes. 429 times is $2.625 divided by $0.006125, and $3.00 divided by $0.00700, on the day of measurement. A chain’s settle cost can move without notice.

A decision rule

There is no best product here, and the differences that matter are constraints rather than quality. Four of them, each pointing at one of the four jobs.

If you need one figure set and you hold the data, buy the operation. One call, $0.006125 on Base, a formula back. It is the cheapest thing on the page and the easiest to check, and when the three analyst services are callable it needs no agent around it at all.

If you need an artefact — a note, a brief, a table — assemble the role. Five to six cents in paid calls buys the computation and the prose; the render is not in that price. Choose between the three roles by question, not by price: they are priced within a few dollars a year of each other. Then price the drafting by prompt length, because that is where a role’s bill moves: short prompts on DeepSeek at $0.00700, filings and long series on Qwen at $0.00875.

If one question needs three views, buy the desk. 17 paid calls and $0.10938 a brief, and the deliverable is the disagreement table. Do not buy three roles and reconcile them yourself unless the reconciliation is the work you want to do.

If the bill has to hold over a year, decide the chain first. Avalanche prices the same call at 29 per cent of Base, and Ethereum at 429 times it. Choose the chain your treasury already holds, then the model, then the volume — in that order of effect.

Two rules sit under all four. Take the amount from the 402 response and pay that rather than computing a charge from a table; the tables are for budgeting and the handshake is for paying. And set max_tokens: the difference between the exact amount and the ceiling is 36 per cent on DeepSeek at a 1,000-token allowance and up to 23 times the exact amount when the field is left out.

The clusters in this family

• How to value a private company
• DCF model structure
• Comparable company analysis
• Financial statement red flags
• Reading a yield curve
• Breakeven inflation
• Inflation decomposition
• FX exposure for importers
• ATR position sizing
• Backtesting without overfitting
• Support and resistance that holds
• Reading order flow

Method note

Prices for the three analyst services are from catalog.json at x402.press, updated 31 October 2026, read the same day; the services are pending review and return no 402 quote, so the engine price is published rather than measured. The DeepSeek and Qwen prices are measured from the live 402 handshake on 31 October 2026 at 08:26 UTC, Base, one short prompt, max_tokens: 1000, unpaid and unsigned, first round only; the same handshake supplied the per-chain figures. The prompt-length test sent about 4,000, 12,000, 40,000 and 120,000 characters — roughly 1,000, 3,000, 10,000 and 30,000 tokens — at max_tokens: 1000, and read the exact and ceiling amounts from each 402 response. Model identifiers are those the endpoints list. The role and desk figures are taken from their own pages and every sum has been recomputed: 40 × $0.006125 = $0.24500, 32 × $0.00700 = $0.22400, and so on through Table 2 and Table 3, with the monthly totals, the ceiling totals and the annual figures each derived from those products rather than copied. The Bazaar count, median and percentile are one pull on 30 October 2026. The ratios — 25 per cent, 20 per cent, 9.1 per cent, 29.27, 1.89, 36 per cent, 41.8, 429, 3.5 — are quotients of the figures above, not measurements, and the monthly counts are assumptions. Nothing here is published as a recommendation to buy, sell or hold any instrument.

The three roles are at financial_analyst, macro_analyst and trade_analyst; the desk at finance_desk. The engines are at macro-analyst, financial-analyst and trade-analyst; the models at DeepSeek and Qwen, reviewed at DeepSeek: per-call chat completions over x402 and Qwen: per-call chat completions over x402.

Filed under Agents