Per-Token Pricing Is Dead. The Outcome Tax Is How AI Companies Actually Charge in 2026.

In Q4 2025, Cursor's enterprise pricing page added a new tier alongside the existing per-user model: an outcome-priced enterprise plan that charges based on pull requests shipped rather than seats provisioned. Around the same time, Intercom's Fin pricing made the per-resolution model the default for new enterprise customers. By Q1 2026, Sierra had publicly disclosed that its primary commercial structure was per-successful-agent-interaction rather than per-query. Across the AI SaaS landscape, the most successful enterprise products were shifting pricing in the same direction.

The aggregate result is that per-token pricing — the dominant commercial model of 2023-2025 AI commercialization — is being replaced as the primary pricing wrapper for enterprise AI products. Its replacement is what the industry has started calling the "outcome tax": a pricing model that charges customers based on measurable business outcomes the AI delivers, with the customer paying a defined share of the value created rather than a usage-based meter on inputs.

The transition is not a marketing repositioning. It is a structural reset of AI SaaS economics, driven by the failure of per-token pricing on three specific dimensions, and it changes which AI companies survive the next 18 months.

Why Per-Token Pricing Failed

Per-token pricing seemed obviously correct in 2023. AI inference cost was high, variable across providers, and a function of input and output volume. Charging customers based on tokens passed through to the model created a transparent unit economic relationship between customer behavior and vendor cost. The early AI SaaS companies adopted some version of per-token pricing because the alternatives — fixed-price subscriptions decoupled from variable inference cost — created either margin compression risk or pricing power capture by customers who used aggressively.

Three years of operating per-token pricing at scale revealed three structural failures.

Failure 1: Customer-side unpredictability. Enterprise customers cannot budget against AI features that produce variable monthly invoices. Procurement teams responsible for software spend control increasingly reject purchases of AI products that cannot offer fixed-cost or capped-cost models. Signal's earlier analysis of the AI pricing crisis documented multiple cases in 2025 where enterprise AI deals collapsed at procurement stage specifically because per-token billing created budget exposure the customer could not control. The pattern accelerated through 2025 as more enterprises hit unexpected month-over-month invoice volatility.

Failure 2: Vendor-side margin compression. Per-token resale models generate gross margins of 30-50% on the underlying inference cost, far below the 70-85% gross margins of traditional SaaS. As Signal documented in the API economy repricing analysis, AI startups operating pure resale models reported average gross margins of 45% in 2025, against an industry benchmark of 75% for software companies. The lower margin makes it nearly impossible to build venture-scale AI businesses on a pure per-token resale model: customer acquisition cost remains at SaaS levels but contribution margin is half what SaaS produces, so payback periods extend and CAC efficiency degrades.

Failure 3: Customer trust erosion. The most damaging failure has been the steady drip of public incidents where customers received surprise invoices ranging from 5x to 50x expected costs. A B2B SaaS product that integrated an AI agent feature into its core workflow saw average customer invoices triple month-over-month when the AI was used aggressively during an end-of-quarter sales push. A developer tools company integrated an AI-assisted code review feature billed per-token; one developer's heavy use produced a $42,000 month-over-month invoice increase for a single account. These incidents become public, get amplified on social media and in enterprise procurement networks, and damage the broader category's trust position with enterprise buyers regardless of which specific vendor was involved.

The aggregate effect is that per-token pricing has reached the end of its useful life as the primary commercial wrapper for AI features. The companies that recognized this earliest — Intercom Fin in 2024, Cursor and Sierra in 2025 — moved to outcome-based pricing ahead of the curve. The companies still relying on per-token resale into 2026 are losing competitive ground.

What the Outcome Tax Actually Looks Like

The "outcome tax" framing is useful because it captures both the structural and the commercial dynamics of the new pricing model.

Structurally, outcome-based pricing decouples the customer's spend from the underlying inference cost. The customer pays for measurable business outcomes — resolved support tickets, shipped pull requests, completed agent interactions, finalized clinical notes — rather than for tokens consumed. The vendor absorbs the cost-of-goods risk on the underlying inference and captures the margin between value-priced outcome and cost-to-deliver.

Commercially, the "tax" framing reflects that customers pay a defined percentage of the value the AI generates. A customer support team that previously paid $35 per resolved ticket through a combination of headcount and tooling can pay $8 per AI-resolved ticket and still capture meaningful savings while compensating the vendor at a margin level traditional per-token resale cannot achieve. The tax rate — the vendor's share of customer value — typically lands in the 15-25% range across the named examples, which is far better economics for the vendor than per-token resale and still favorable economics for the customer.

The product implications of the outcome tax structure are significant.

Definition precision becomes critical. "Resolved support ticket" requires specific measurable resolution signals — customer-confirmed resolution, no follow-up within a defined window, sentiment threshold above some bar. Loose outcome definitions produce dispute risk and customer trust erosion that mirrors the per-token surprise-invoice problem. The best outcome-priced products define success criteria with the rigor of an SLA.

Telemetry investment increases. Vendors must build telemetry that observes outcomes rather than relying on AI self-reporting, because customer trust in outcome billing depends on independent verification. The engineering investment in telemetry, attribution, and reporting tooling is meaningful — often 15-25% of total engineering investment in early-stage outcome-priced products.

Fail-safes become a product feature. When the AI reports an outcome that was not actually achieved (or that the customer disputes), the vendor needs documented processes for review, refund, and resolution. The fail-safe design is increasingly a competitive differentiator, with the most sophisticated vendors offering machine-readable outcome dispute APIs and automated refund logic.

The Examples That Are Working

Five companies have established what successful outcome-based AI pricing looks like at scale. Each illustrates a different facet of the new pricing model.

Intercom Fin (per-resolution). Intercom's Fin product charges customers per resolved customer support conversation, with resolution defined by customer confirmation or absence of customer follow-up within a defined window. The pricing transition from per-message billing to per-resolution billing through 2024-2025 produced a meaningful improvement in customer adoption metrics, because resolution is the outcome customers actually care about. The model has become the canonical example of outcome-based AI pricing in customer support.

Cursor (per-shipped-PR enterprise tier). Cursor's enterprise pricing introduced a tier that charges based on pull requests shipped through Cursor-assisted development, with shipped PRs defined as merged-to-main with passing tests. The tier complements the standard per-user model and gives enterprise customers a way to pay for AI productivity in a unit that matches their business outcomes. The pricing has been particularly attractive to engineering organizations transitioning from headcount-based productivity measurement to output-based measurement.

Sierra AI (per-successful-interaction). Sierra's customer-facing AI agents are priced per autonomous customer agent interaction with defined success criteria — typically including customer-stated resolution, sentiment threshold, and task completion verification. The model has been particularly effective for vertical applications (retail returns, telecom service requests, healthcare scheduling) where the success criteria are clearly definable.

Harvey (per-matter completion). Harvey's legal AI platform charges legal services firms per matter completed rather than per query, with matter completion defined by the law firm's internal billing milestones. The pricing has been particularly aligned with law firm economics because matter-based billing is already the dominant revenue structure in legal services.

Abridge (per-clinical-note). Abridge's clinical AI documentation charges healthcare providers per finalized clinical note rather than per recording minute or per query. The pricing aligns with healthcare billing economics and has driven faster enterprise adoption than per-query alternatives.

The pattern across these examples: outcome unit selection drives commercial success more than any other pricing decision.

The Operating Playbook for Pricing Transition

For AI SaaS companies operating on per-token or per-seat pricing in 2026 and considering the transition to outcome-based pricing, five operating moves carry disproportionate weight.

1. Audit your customers' actual success metrics, not your usage metrics. The biggest mistake in outcome-priced pricing design is choosing a unit that is easy to measure rather than the unit the customer actually budgets against. Customer success metrics — resolved tickets, shipped features, closed deals, completed matters — sit in different systems than the AI's usage telemetry. The pricing-design audit needs to look at customer financial systems, not your product analytics. Spend two weeks interviewing customer-side finance and ops leaders before designing the outcome unit.

2. Run hybrid pricing for at least four quarters before fully transitioning. Pure outcome-based pricing creates revenue volatility that early-stage SaaS companies often cannot absorb. The transition should mix outcome billing with platform fees, retainer minimums, or annual commitments that provide revenue floor stability while the outcome metering scales. Most successful outcome-priced AI products operate hybrid models for two to four years before fully transitioning.

3. Invest in outcome telemetry as a first-class product feature. Customer trust in outcome billing depends on transparency about how outcomes are measured. The most successful outcome-priced products have built customer-facing outcome dashboards that show the full audit trail of each billed outcome, the resolution signals that triggered the billing, and the dispute path for outcomes the customer disagrees with. The dashboard is often the most-used product feature for the customer's procurement team.

4. Design dispute resolution as a customer success differentiator. Outcome disputes will happen. The vendor's response to disputes — speed, fairness, transparency — becomes a major retention and expansion driver. The companies that build best-in-class dispute resolution become the trusted standard in their category; companies that handle disputes defensively destroy customer trust over time.

5. Price the outcome relative to the customer's alternatives, not your cost. Per-token pricing failed partly because vendors priced based on inference cost markup, which is a thin economic story. Outcome pricing succeeds when the vendor prices relative to the customer's alternative cost of achieving the same outcome — internal headcount, alternative tools, manual processes. The "tax rate" of the vendor should reflect customer alternative cost rather than vendor input cost. This shift in pricing logic is the deeper change beneath the surface change in pricing model.

What the Outcome Tax Means for the Foundation Labs

The transition from per-token to outcome-based AI SaaS pricing has second-order effects on the foundation labs themselves.

For OpenAI, Anthropic, and Google DeepMind, the per-token API revenue model is starting to face pressure from the same dynamics that pushed AI SaaS companies away from per-token pricing. Enterprise customers buying API access at scale are increasingly negotiating outcome-aligned or committed-volume pricing arrangements rather than pure metered consumption. The labs' response so far has been to offer enterprise commitments and reserved capacity, but the longer-term implication may be a shift toward more sophisticated commercial structures that mirror the outcome-based SaaS pattern.

For open-source model ecosystems, the outcome-based pricing transition is structurally favorable. In a per-token world, open-source models compete directly with closed models on cost-per-token, which favors open-source for cost-sensitive deployments. In an outcome-based world, the customer pays for measurable business outcomes regardless of which underlying model produces them, which decouples the model layer from the commercial layer. Companies using open-source models internally can capture the cost savings while still pricing to customers at outcome-relevant levels. This is structurally favorable to open-source models because it removes the per-token commercial competition while preserving the technical contribution.

Signal's earlier analysis of the AI middleware tax covered how the middle layer between foundation labs and AI SaaS companies has been extracting margin from the entire AI value chain. The outcome-based pricing transition reshapes this dynamic by giving AI SaaS companies pricing structures that produce higher margins than per-token resale, which reduces the margin pressure on the middle layer and creates room for more sustainable infrastructure economics across the stack.

The Longer Arc: Pricing as Strategy

The transition from per-seat SaaS pricing to per-token AI pricing to outcome-based pricing is more than a tactical pricing change. It is a strategic shift in how software companies relate to customer value.

Per-seat pricing — the dominant model from approximately 2005 to 2023 — priced software based on the number of users with access. The model was simple, predictable, and worked when software value was a function of how many people in an organization could use it. As AI features made the per-user assumption less meaningful (one user with AI is worth many users without), the model came under pressure.

Per-token pricing — the dominant model from 2023 to 2025 — priced software based on AI inference consumption. The model captured the variable cost of AI inference but failed to align vendor pricing with customer value, which produced the unpredictability and margin problems documented above.

Outcome-based pricing — the emerging model in 2026 — prices software based on measurable customer value the software produces. The model aligns vendor and customer incentives more tightly than either predecessor and produces healthier economics on both sides when the outcome unit is well-chosen. The model is harder to implement, requires more sophisticated telemetry, and creates new categories of operational complexity. The vendors that get it right will define the next decade of AI commercialization.

Takeaway: Per-token AI pricing has reached the end of its useful life as the dominant commercial model. The 2026 transition to outcome-based pricing — the "outcome tax" — replaces variable input-cost meters with fixed-share-of-customer-value billing, producing better economics for both vendors and customers when the outcome unit is well-chosen. Intercom Fin, Cursor's enterprise tier, Sierra AI, Harvey, and Abridge have established the canonical examples. For AI SaaS companies still operating on per-token resale, the operating playbook is to audit customer success metrics rather than your usage metrics, run hybrid pricing for at least four quarters before fully transitioning, invest in outcome telemetry as a product feature, design dispute resolution as a differentiator, and price relative to customer alternatives rather than vendor cost. The pricing model that defines 2026-2030 AI commercialization will not be a tweak to per-seat or per-token — it will be the outcome tax, in increasingly sophisticated forms.