LLM Pricing Models Compared: Token vs Consumption vs Subscription

AI pricing was easy to ignore when LLMs were still side projects. A few thousand dollars in token spend could disappear into a vague “innovation” budget, attached to experiments that never survived contact with real users. But that happy-go-lucky phase is over.

Once agentic support workflows, coding assistants, or AI features inside a SaaS product reach production scale, pricing stops being a procurement detail. Choose the wrong model, and a successful rollout can quickly become an unplanned six-figure expense—or a deal-breaker at renewal.

The cautionary tales are already emerging. In early 2026, Uber reportedly burned through its annual AI tools budget in just four months after rolling out agentic coding assistants. The tools had done exactly what they were supposed to do. The problem was that the financial model around them had not kept pace. The dream of large-scale automation had run headfirst into the bill.

Most overruns follow the same pattern on a smaller scale. A support team models “cents per ticket” during a pilot, then ends up with a five-figure monthly bill once multi-step agents, retrieval, monitoring, and safety checks are added in production. A SaaS vendor includes AI in its premium tier, only to discover that a small group of power users is quietly consuming most of the token budget and eating into margins.

In each case, the technology worked, the contract was where the wheels came off.

This article breaks down the pricing models AI vendors actually use in 2026: token-based pricing, broader consumption models, subscriptions, committed spend, and hybrids that combine all three. It explains how to evaluate them against your own usage patterns—and which questions help distinguish a fair, scalable contract from one that comes back to bite you at renewal.

In 2023, most enterprise AI contracts had one shape: pay per token, watch the meter, budget accordingly. Since then, that shape has split into several commercial models.

Vendors rarely explain which one a contract is. OpenAI, Anthropic, Google, and AWS have each layered token rates, consumption tiers, commitments, and bundled subscriptions on the same compute. The label on the invoice usually obscures more than it reveals.

The same workload, at the same volume, can cost wildly different amounts depending on which commercial model sits underneath it. A support team processing 50,000 tickets a month might pay a few hundred dollars on a pay-as-you-go token meter. It might pay several thousand on a consumption-priced bundle, or more on a subscription with unused capacity.

These numbers reflect different bets on how predictable your usage will be. Vendors are betting on your behalf unless you read the terms closely.

Understanding LLM pricing models is part of the technical decision itself, not a finance-team afterthought bolted on afterward. The engineer choosing which model tier to route a workload to is making a budget call as much as an accuracy call. The procurement lead signing an annual contract is making an architecture call as much as a commercial one.

The Traditional Model: Token-Based Pricing

Token-based pricing is the original commercial model for hosted LLMs, and it remains the substrate underneath almost everything else in this article. Vendors charge per token processed, quoted per 1,000 or per 1 million tokens, with input and output priced separately.

Output tokens cost more, typically two to six times the input rate, because generation is more compute-intensive than reading a prompt. A 1,000-word document runs roughly 1,300 tokens, though that shifts for code and non-English text.

The clearest way to see how far this model has moved is to look at the same market three years apart. When OpenAI launched GPT-4 in March 2023, it priced the model at $30 per million input tokens and $60 per million output tokens: numbers that made enterprise adoption a high-value-use-case-only decision.

Claude 2, Anthropic’s flagship model at the time, launched in July 2023, priced at $8 per million input tokens and $24 per million output tokens: cheaper than GPT-4, but the same broad range. By July 2026, Anthropic’s lineup runs from $1/$5 per million tokens on Claude Haiku 4.5 up through $3/$15 on Claude Sonnet 5 (standard rate) and $5/$25 on Claude Opus 4.8, with the Fable-class frontier tier at $10/$50. OpenAI’s spread runs from sub-$1 nano-class models up to premium reasoning tiers in the $15-30 range.

Frontier-model pricing has fallen roughly tenfold since 2023, driven by hardware efficiency and direct competition between providers.

Token-based pricing’s advantage is transparency. The unit economics are visible. There is no minimum spend, and a team can start experimenting immediately at whatever volume it needs.

The disadvantage shows up at scale, particularly under agentic workloads. Multi-step agents burn tokens far faster than a simple chatbot: long contexts, tool-call schemas, and reasoning loops routinely push per-interaction cost 30x above conversational use. A metered rate that looked negligible in a prototype can become a five- or six-figure monthly line item once a workflow goes into production.

Reasoning models compound this further. Hidden “thinking” tokens can make the effective cost three to nine times the headline output price.

Token-based pricing makes the most sense for prototyping, for workloads with genuinely variable or bursty volume, and for any team that does not yet have a predictable usage baseline to negotiate against. It’s the right default until you have real production data, and often the wrong choice to stay on indefinitely once you do.

The New Model: Consumption-Based Pricing

Consumption-based pricing takes the same underlying token economics and repackages them around a unit of work rather than a raw token count: price per API call, per inference, per document processed, or per feature invocation.

The distinction from pure token-based pricing is subtle but real. Token-based pricing charges for compute. Consumption-based pricing charges for an outcome that happens to be built on that compute, often bundling several metered components (model tokens, a tool call, a retrieval lookup, a safety check) into what looks like a single line item.

Amazon Bedrock illustrates this well. Its on-demand pricing bills per 1,000 input and output tokens at rates matching each provider (Claude pricing on Bedrock matches direct Anthropic API pricing). Layered on top: Guardrails content-filter evaluations at $0.15 per 1,000 text units, Knowledge Base retrieval calls, and Agents orchestration overhead.

A single “request” to a Bedrock-hosted agent can trigger four or five differently priced meters running simultaneously. Google has moved in a similar direction by folding Gemini access into Workspace’s per-seat plans rather than charging a separate line item for every AI-assisted action. The AI cost is still there, just consumed as part of a broader feature bundle rather than itemized per call.

The appeal of consumption-based pricing is that it maps more naturally to a business conversation. “$0.03 per ticket resolved” is a number a support director can put in a board deck. “1,850 input tokens and 240 output tokens per ticket at $3/$15 per million” doesn’t land the same way. This framing also gives vendors room to price different features at different margins without touching the headline model rate.

The trade-off is auditability. A bill built from five metered components is harder to reconcile than one built from a single token rate. Vendors can quietly reprice an individual component, a guardrail check, a retrieval call, without touching the number a buyer actually remembers from the sales deck.

Aggregate spend still scales with usage exactly as it does under token-based pricing. Consumption-based pricing changes the unit you see, not the underlying variability.

This model fits teams with a well-defined, repeatable workflow where cost genuinely tracks a business unit: documents processed, tickets resolved, calls handled. It suits those teams far better than those still exploring what an LLM-powered feature should even do.

Enterprise Model: Subscription/Commitment Pricing

Subscription and commitment pricing moves the bet in the other direction. Instead of paying for exactly what you use, you pay upfront, monthly, or annually, for a defined envelope of usage, in exchange for predictability and, usually, a discount off list.

This is the subscription pricing AI vendors have increasingly leaned into once a deal gets big enough to negotiate.

Two variants show up in enterprise deals. The first is a per-seat subscription. ChatGPT Enterprise has no published list price, but reported 2026 contracts cluster around $50-60 per user per month, with a 150-seat minimum and a mandatory annual prepaid term: a roughly $108,000 floor before any usage above the included baseline.

Anthropic’s equivalent tiers are less rigid. Claude for Work runs from Team Standard at $25/seat and Team Premium at $125/seat up to Enterprise pricing custom-quoted from around $20/seat plus usage, without the same hard seat floor.

The second variant is committed API spend. Organizations that commit to $250,000-$500,000 in annual API volume typically see 15-20% off list pricing. $500,000-$1,000,000 commitments reach 20-30%, and commitments above $1-5 million can unlock 30-40%+ discounts along with reserved capacity.

AWS Bedrock‘s Provisioned Throughput works on the same logic at the infrastructure layer. You reserve a dedicated model unit at an hourly rate, with 1-month and 6-month commitment tiers priced progressively lower than no-commitment rates.

The advantages are real: a single predictable invoice, negotiated discounts unavailable to pay-as-you-go customers, guaranteed throughput where it applies, and the security and data-handling commitments compliance teams need regardless of price.

The disadvantages are just as real, and less often quoted by sales teams. Seat-based commitments routinely go underutilized. One procurement analysis found a 2,000-seat deal at 35% activation, wasting roughly $390,000 a year on unused licenses.

Annual terms lock in spend even if usage drops mid-year. And “unlimited” bundles rarely are: ChatGPT Enterprise’s base seat covers standard usage, but heavy Codex or agentic workloads bill separately once a baseline is exceeded.

Subscription and commitment pricing make sense once usage is large enough and steady enough to negotiate against, and once procurement simplicity, guaranteed capacity, or compliance requirements matter as much as the per-token rate.

Hybrid Models

Most 2026 enterprise AI contracts are, in practice, hybrids, whether or not either side calls them that. The typical pattern is a committed base (a subscription fee, a seat license, a minimum spend) with metered overage on top once usage crosses an included threshold.

A second pattern bundles several products (a chat interface, an API, an agent platform) under one price so light and heavy users share a single line item.

Google’s Workspace pricing is the cleanest recent example of the second pattern. Through early 2025, Gemini was a $20-30 per-user add-on customers could switch on selectively.

Google then folded it into Business and Enterprise tiers with no opt-out, raising the base price roughly 17-22% to absorb the cost. Teams that wanted Gemini now pay less than the old add-on. Teams that never wanted it now pay more regardless.

OpenAI’s ChatGPT Enterprise follows the first pattern. A flat per-seat price covers a Codex baseline, with additional agentic usage metered separately once that baseline is exhausted. Anthropic runs a version of the same idea: Claude Pro and Max subscribers get Claude Code usage included in their seat price, while the identical models accessed through the API bill per token with no subscription involved.

The appeal of hybrid pricing is real on both sides of the table. It smooths a vendor’s revenue while giving buyers one number to budget against, with metering only kicking in for genuine outliers.

The confusion points are exactly what make hybrid models the hardest to audit. What counts toward the included baseline is rarely documented with the same precision as the overage rate.

The same underlying model can be free on one product surface and metered per token on another. That makes “how much does this model cost” a question with no single answer until you specify which access path you mean.

Pricing Model Comparison: Side-by-Side

The commercial model that wins depends almost entirely on two variables: how much you use and how predictable that usage is. Three scenarios, each using representative 2026 rates, make the trade-offs concrete. Each one uses a vendor’s published or reported pricing rather than a rounded average, so the numbers can be checked against the sources directly.

Scenario 1: Small company, variable usage

A 15-person team runs internal tools and occasional customer-facing chat, totaling roughly 50 million tokens a month (35M input, 15M output), with volume that swings 3-4x month to month as projects start and stop.

Winner: token-based. At this volume, a seat minimum or committed spend buys nothing this team needs. Pure metering is both the cheapest and the most forgiving option when usage is this uneven.

Scenario 2: Growing company, predictable usage

An 80-person company has steady internal adoption plus one production customer-facing feature, running about 600 million tokens a month (400M input, 200M output) with usage that has flattened month over month.

Winner: negotiated consumption/token pricing. This is the awkward middle. $50K a year sits below the roughly $250K threshold where committed-spend discounts start, so this company is too small for meaningful commitment pricing but too steady to justify a rigid seat-based Enterprise plan built for a much larger headcount. The pragmatic move is to negotiate volume pricing with a smaller vendor tier, not the flagship enterprise contract.

Scenario 3: Enterprise, large committed volume

A 2,000-seat organization runs agentic coding assistance and customer-facing automation at a scale that would run into the billions of tokens a month at list price.

Winner: subscription/commitment. Above roughly $1 million in annual spend, negotiated discounts of 30-40% become available. AWS Bedrock’s Provisioned Throughput can beat on-demand pricing at high, steady volume, but the exact break-even point depends on the model and traffic pattern, so it needs case-by-case modeling, not a single rule of thumb. At this scale, the guaranteed capacity and procurement simplicity are worth as much as the discount itself.

Across all three scenarios, this AI pricing comparison points to the same conclusion. The model that wins is rarely the one with the lowest headline rate. It’s the one whose commitment structure matches how predictable your actual usage is, and that match is worth checking before any contract gets signed, regardless of company size.

Hidden Factors That Change Your Bill

Even inside a single pricing model, several factors routinely move the real bill 20-50% away from the number on the pricing page.

• Context window size: Some vendors tier pricing by prompt length rather than charging a flat rate regardless of size. Google’s Gemini 3.1 Pro, for instance, charges $2/$12 per million tokens for prompts up to 200,000 tokens, then $4/$18 above that threshold. A document-heavy workload can silently cross into the higher band.
• Model variants: The three-tier split vendors settled into (a reasoning-heavy flagship in the mold of OpenAI’s o1, a fast mid-tier workhorse like GPT-4 Turbo, and a stripped-down “basic” or “mini” tier for high volume) still shapes pricing today. The gap between a vendor’s cheapest and most expensive model routinely spans 20-30x per token, and reasoning variants add a further three-to-nine-times penalty.
• Region and data residency: Restricting inference to a single geography carries a real premium. Anthropic’s US-only routing option, for example, applies a 1.1x multiplier across input, output, and cache pricing versus its default global routing.
• Fine-tuning vs. base model: Customizing a model changes the pricing shape, not only the rate. Fine-tuning typically means paying for training tokens, an ongoing monthly storage fee for the resulting model, and, on platforms like AWS Bedrock, a mandatory move from on-demand pricing to Provisioned Throughput, since customized models generally cannot run in pay-per-token mode at all.
• Rate limits and overage pricing: “Unlimited” plans commonly mean “reasonable use,” a threshold vendors rarely publish. Exceeding an included tier can trigger throttling, a forced upgrade, or metered overage charges that only appear once the invoice arrives.
• Caching and storage. This is the largest lever most comparisons skip. Prompt caching typically cuts cached input cost by up to 90%, and batch processing adds a further flat discount, commonly 50%, for workloads that can tolerate asynchronous processing.

Stacked together, these two levers can turn a five-figure monthly workload into a fraction of that. List price is close to meaningless for any cache-heavy, batchable application, regardless of which commercial model wraps around it.

How to Evaluate a Vendor’s Pricing

A rate card tells you what a token costs. It doesn’t tell you what your workload will actually cost, or whether the contract in front of you is competitive. A short list of questions and checks closes that gap.

• Questions to ask: What counts toward an “included” baseline, documented in writing rather than described verbally by a sales rep? Is the price per token, per call, or per seat? Can you get an actual rate card, not just a demo quote?
• Do negotiated discounts apply to every model in the family, or only to specific tiers? What happens to unused committed spend at contract end? Is there a true-down right if seat activation lands below commitment?
• How to calculate true cost: Model your own measured workload, not a vendor’s benchmark numbers: token volume by model tier, cache-hit rate, share of traffic eligible for the batch API. Multiply against real rates, not the pricing-page headline, and run the same math across two or three vendors.
• What to negotiate: A longer term for a lower per-seat or per-token rate. A true-down right tied to activated users, not provisioned seats. Price protection if the specific model version you built on is deprecated mid-contract. A guarantee that new model generations arrive at equivalent-or-better economics.
• Red flags:  No rate card at any commitment size (reasonable for a genuinely custom enterprise deal, a warning sign for anything smaller). “Unlimited” usage with no published threshold. Contracts pinning a single model version with no migration guarantee. Bundled AI features that quietly raised the base subscription price for every seat with no opt-out.

Reading a vendor pricing strategy correctly says almost as much about a company’s priorities as its product roadmap does. A vendor pricing per outcome is betting on the outcome; one pricing purely per token is pricing the compute.

Choosing Your Pricing Model

Four commercial models, one underlying resource. Token-based pricing charges for compute directly and suits variable or early-stage usage. Consumption-based pricing repackages that compute around a business-relevant unit and suits well-defined, repeatable workflows.

Subscription and commitment pricing trades flexibility for predictability and negotiated discounts, paying off once volume is large and steady enough to justify it. Hybrid pricing (a committed base plus metered overage, or several products bundled under one price) is where most real 2026 contracts land, deliberately or not.

The decision framework is simpler than the pricing pages make it look. If usage is small or swings 3x or more month to month, default to token-based pricing and revisit once you have a real baseline. If a workflow is repeatable, consumption-based pricing is worth the extra audit effort.

If usage is large and steady, negotiate a commitment and use the leverage volume gives you.

Next steps are about running the numbers on your actual contract, then negotiating rollover rights, true-down clauses, and model-deprecation protection. Vendor selection matters just as much as pricing: two vendors on the same commercial structure can still differ sharply on quality and support.