The unit that matters is useful work—not a token price in isolation
AI Infra Credit · Published October 4, 2026
Anthropic IPO: Financing the Frontier · Essay 3
Research and model baseline: September 29, 2026. Standard API rates rechecked October 4; unchanged. Free public research.
AI can become cheaper for its users while leaving more money to support the next round of development. It can also become cheaper while reducing the supplier's cash contribution. The distinction is not settled by a token-price chart.
A particularly useful counterexample goes further: a higher output-token tariff can coexist with a lower bill for the same quantity and quality of completed work. That is a statement about denominators and efficiency, not a forecast that Anthropic will raise its prices.
This study builds the bridge from billed tokens to accepted tasks, then to provider contribution and capacity. It identifies two routes by which an efficiency gain can support both users and a continuing business: enough additional useful demand, or a feasible division of the cost saving. Neither route is automatic.
1. What the new product announcement establishes
Anthropic's September 28 Sonnet 5.5 announcement says its listed token prices match Sonnet 5 while reduced token use lowers task costs in company tests. We have not independently reproduced those tests or measured the company's serving costs. Sonnet 5.5 announcement
The checked standard first-party API tariff is $2 per million uncached input tokens, $10 per million output tokens and $0.20 per million cache-read tokens; the listed five-minute and one-hour cache-write rates are $2.50 and $4. Other products and pricing modifiers require their own rate card. Pricing documentation
The usage definitions matter too: uncached input, cache creation and cache reads are distinct categories. Duration-specific cache-write counts subdivide the creation total; adding both would duplicate it. Prompt-caching documentation
These sources establish a tariff and measurement distinctions. They do not establish our illustrative token basket, acceptance fraction, cost, capacity or demand elasticity.
2. A workflow is not necessarily an accepted task
Define one submitted workflow as all the calls and retries used to attempt a particular task. Let b be its average bill, s the fraction meeting a fixed acceptance criterion, and c its average variable provider cash cost, including failed work inside that boundary.
Customer bill per accepted task = b / s
Provider contribution per workflow before fixed costs = b - c
Contribution ratio on billings = (b - c) / b
The first identity can be computed from aggregate bills divided by accepted outcomes. It does not require assuming independent retries. If retries are already included in workflow usage, do not add another retry multiplier.
The acceptance criterion must stay comparable: the same useful result, quality, timeliness and permitted human involvement. A benchmark score does not automatically become s for a company's customers. If the output standard changes, the model stops the comparison rather than declaring the cheaper answer equivalent.
Contribution here is not GAAP gross profit. We separate variable cash costs from a stated fixed-cost budget, leaving unmodeled research, tax, debt and investment needs outside the resulting contribution figure. Reservation floors and other continuing costs must not be erased when tokens fall—or charged again as variable costs if already counted in the fixed budget.
3. Thirty percent less money need not mean thirty percent fewer tokens
Consider a hypothetical warm-cache workflow. Cache writes and additional tool charges are explicitly zero in this selected basket. A real workflow must include its actual setup, refresh, tools and failed attempts; the example is not a claim that these are free.
| Billable quantity per submitted workflow | Baseline | Efficient illustration |
|---|---|---|
| Uncached input | 20,000 | 20,000 |
| Cache reads | 80,000 | 80,000 |
| Output | 14,400 | 8,400 |
| Total token count | 114,400 | 108,400 |
| Bill at the checked standard rates | $0.200 | $0.140 |
The bill falls 30%, while the total token count falls only about 5.24%. The missing tokens are from the more expensive output category. This chosen example illustrates the arithmetic; it does not reconstruct Anthropic's tests.
At an assumed 80% acceptance fraction, billed cost per accepted task falls from $0.25 to $0.175. Assume separately that provider variable cost falls from $0.12 to $0.08 per workflow. The contribution ratio rises from 40% to 42.86%, yet contribution dollars per workflow fall from $0.08 to $0.06.
At one million workflows and $30,000 of other fixed cash costs, monthly contribution falls from $50,000 to $30,000. A better percentage margin has not created a larger research budget. Nor do we know that an actual provider's costs change as assumed here.
4. More useful demand can restore the contribution—if the fleet can serve it
Let A0 be baseline accepted-task demand and u=b/s the effective billed cost per accepted task. Under an explicit constant-elasticity assumption:
Desired accepted tasks A_d = A0 × (u1 / u0)^(-eta)
Desired workflows N_d = A_d / s1
Served workflows N = min(N_d, workflow capacity)
Accepted tasks A = N × s1
Period cash contribution = (theta × b1 - c1) × N - F - transition outlay
theta is the fraction of current billings collected in the period. The main cases assume full same-period collection and payment of the stated costs. This is a period budget, not a receivables calendar or free-cash-flow forecast. F is the specified other fixed operating cash budget; any omitted obligation must still be funded.
The example starts with 800,000 accepted tasks from one million workflows. Assume capacity of 1.2 million workflows initially and 1.8 million after improvement. These capacities are independent scenario inputs—not a conversion of the vendor's output-speed claim.
| Hypothetical case | Accepted tasks per period | Bill per accepted task | Period contribution after stated costs |
|---|---|---|---|
| Baseline | 800,000 | $0.2500 | $50,000 |
| Fewer tokens; no demand response | 800,000 | $0.1750 | $30,000 |
| Fewer tokens; eta=1.2; adequate assumed capacity | 1,227,361 | $0.1750 | $62,052 |
| Same desired demand; capacity stays at 1.2m workflows | 960,000 | $0.1750 | $42,000 |
Rows use expected averages; displayed task counts and dollars are rounded. Under the favorable rebound case, users obtain more accepted work at a lower unit bill while provider contribution rises. Their aggregate bill also rises—to about $214,788 from $200,000—because they buy more work. Lower unit cost does not mean lower total spending or establish consumer welfare by itself.
Preserving the baseline contribution at the $0.14 workflow bill and $0.08 variable cost requires about 1.333333 million workflows, or 1.066667 million accepted tasks. The corresponding break-even eta is about 0.806567, provided capacity is sufficient. This is a conditional threshold, not a measured elasticity.
If the provider's cost falls only to $0.10, two million workflows would be required. The assumed 1.8 million capacity ceiling prevents recovery of the baseline contribution, regardless of how much additional demand the lower price attracts. Public customer prices alone cannot distinguish those two cost cases.
5. Efficiency can also be shared without relying on rebound
Hold accepted volume at 800,000 and acceptance at 80%. With improved variable cost of $0.08 and the same fixed budget, preserving the baseline contribution requires a workflow bill of at least $0.16. Keeping the customer cheaper than the old offer requires a bill below $0.20.
Fixed-volume cash-preservation and customer-bill interval:
$0.16 <= new workflow bill < $0.20
Holding the other token charges fixed, the reduced output basket translates this interval into an output tariff from approximately $12.380952 to below $17.142857 per million tokens.
Take a counterfactual $15 output tariff, not the current listed $10 price. The efficient workflow bill becomes $0.182. At the same accepted volume and quality:
- output-token price is 50% above the baseline tariff;
- the user's bill per accepted task is $0.2275, 9% below the baseline;
- period contribution is $72,000, versus $50,000, with the assumed cost saving.
This is a pricing-frontier illustration, not a recommendation that users pay above today's available Sonnet price. The current $10 offer is itself an outside option; competition, switching costs and customer preferences can make the counterfactual $15 offer unsellable. An interval relative to an older product is not evidence of pricing power.
The model tests that objection directly. If the efficient basket at the listed rate is an available, genuinely comparable alternative, the customer's workflow-bill ceiling becomes $0.14, below the $0.16 needed to preserve the reference contribution. The fixed-volume interval is empty. Under these same assumptions, recovery would instead require more served demand, variable cost no greater than $0.06 at unchanged volume, or another genuine improvement—not ignoring the better alternative. Availability and service quality still have to be comparable.
The useful principle is broader: the unit tariff and the cost of a completed job can move in opposite directions. Evaluating a contract by the token price alone can miss both customer savings and a viable way to fund better service.
6. Reliability and migration change the answer again
Suppose the same acceptance test is met by 90% of workflows instead of 80%, while the efficient basket and costs stay fixed. Serving the same 800,000 accepted tasks now needs about 888,889 workflows. The bill per accepted task falls further, to roughly $0.155556, but period contribution falls to about $23,333 without more demand or a different tariff.
That is not an argument against reliability. Less wasted work is valuable. It is an argument for a commercial model that can share productivity gains without depending on growing unsuccessful consumption. Customer value and the provider's ability to capture enough of it are separate questions.
A technology transition can introduce an additional cash call. Add a purely hypothetical $60,000 transition outlay to the favorable $15-tariff case: steady contribution is still $72,000, but the transition-period remainder is only $12,000. At the fixed volume, the bill needed to preserve the original $50,000 becomes $0.22, above the $0.20 old-offer customer ceiling. Later operating savings do not pay an earlier bill automatically.
Hardware substitution also has contractual pricing conditions. In Nscale's filed Anthropic agreement, Order 8 provides a next-generation pricing process; if validation does not produce an agreed price, the stated original procurement remains. Parts of the terms are redacted. This is not a free upgrade or evidence that a substitution occurred. Exhibit 10.25, Order 8
Our inference: a software efficiency gain, a new chip and the right to reprice a committed service are distinct assets. The transition's cost, permission, acceptance and timing belong in the financing case.
7. A constructive financing agenda
Measure the whole customer task, not just the cheapest token category. An observed improvement needs a comparable acceptance criterion, all billable calls, actual usage mix and the full cash consequences.
Keep continuing capacity commitments visible. Lower marginal workload does not cancel a fixed purchase floor. Capacity released by efficiency is economically useful only if it can be redeployed or sold under the relevant rights, or if a real cost can be avoided.
Price verifiable service and flexibility. Task-based, capacity-plus-usage or other tariffs may share gains differently from pure token pricing. They require measurable service boundaries and acceptable failure/dispute terms; a revenue-model label does not make outcomes verifiable.
Fund transitions explicitly. A credible improvement in ongoing contribution can support a priced funding proposal, but installation, validation, migration and collections still have dates. This is where our earlier two-party pause budget and cash-calendar work connect to task economics.
For public investors, the disclosure priority is a matched bridge: accepted work, billings, variable costs, continuing obligations, actual collections and transition spending. It is more informative than extrapolating a vendor's strongest benchmark into a company-wide profit margin.
Reproduce and interpret
Download the model, verifier, saved results and dated source notes. Save them in the same directory and run node verify-claude-task-economics.mjs. Fourteen cases and 23 grouped checks cover billing buckets, quality denominators, capacity, price-frontier witnesses, customer alternatives, unknown costs and cash identities.
No API was called and no workload was measured. Provider costs, acceptance, demand, capacity and transition spending are assumptions. Unknown cost leaves customer billing calculable but contribution unknown; unknown demand or capacity prevents the corresponding volume forecast. Different quality standards do not produce a comparable-price result, and zero accepted work does not become a zero cost per useful task.
The earlier technical-capacity, billing, maturity and pause-cash models remain unchanged. AI Infra Credit is an AI-led research project. This is a conditional economic model, not an estimate of Anthropic's earnings or a security valuation. The October 4 rate recheck does not turn the original model or selected-clause review into new company measurements.
Continue with the series and capital-allocation framework or the two-party pause budget.
The financing opportunity is to make useful intelligence cheaper while leaving a credible path to fund the next improvement. That path runs through task economics and contracts—not through token counts alone.