2027 planning guide · Pricing

AI API pricing in 2027: flat rate, unlimited plans and fair use

A practical guide to per-token, flat-rate and private AI pricing, and the capacity limits that decide what you actually pay.

Short answer. AI API pricing in 2027 comes in four shapes: per-token metering, flat-rate subscriptions with fair-use limits, committed enterprise contracts and private deployment licences. The lowest sticker price rarely decides total cost. Concurrency, retries, context length, embeddings and support usually do. Compare providers on cost per completed task at your peak load, not on cost per million tokens.

Quick answers to the questions buyers ask first.
QuestionOne-sentence answer
What is a flat-rate AI API?A subscription that bills a fixed monthly or annual price instead of metering each token, usually bounded by concurrency and requests-per-minute limits.
Is unlimited the same as infinite capacity?No. Unlimited means no monthly token cap and no per-token overage; fair-use concurrency and RPM limits still apply.
Which is cheaper for a chatbot?Flat rate is usually cheaper at steady, high utilisation; per-token is cheaper for spiky, low-volume or experimental workloads.
What hidden costs matter most?Retries, long-context prompts, embedding and reranking calls, egress, support tiers and the engineering time spent managing rate limits.
How do I compare an RFP quote?Normalise every quote to cost per 1,000 completed tasks at your peak concurrency, then add support, private deployment and exit costs.
Do annual discounts lock me in?They reduce the price but shift risk; check termination, refund and price-change clauses before prepaying for a year.
What limits should I load-test?Concurrent requests, RPM, input and output ceilings, queue behaviour at peak, and how the API signals HTTP 429.
When is per-token better?It is better when usage is unpredictable, total volume is low, or rare long-context jobs dominate the bill.

AI API pricing is easy to quote and hard to compare. One provider sells tokens, another sells a monthly plan, and a third sells a committed pool with a private network attached. The invoice lines differ because the providers are pricing different risks.

This guide is for the buyer who has to defend a number: a founder modelling run-rate, an architect sizing concurrency, or a procurement lead normalising three RFP responses. It covers the four pricing models, the precise meaning of unlimited, the total-cost lines that get missed, and the clauses to fix before signature.

The four pricing models buyers actually see

Many AI API quotes reduce to one of four models. They differ less in headline price than in who carries the risk of unused capacity.

How does per-token AI API pricing work?

Per-token pricing bills each request for the input tokens it sends and the output tokens it generates, often at different rates. The meter usually includes system prompts, tool schemas, retrieved context and images, not just the user's message. Reasoning models may also bill reasoning tokens that the user never sees. The commercial consequence is that the invoice measures consumption, not completed work: a timeout followed by a retry can be billed twice while producing one answer. Ask whether cached input, failed requests and reasoning tokens are billed, and whether rates differ by region. For current rates, see the pricing page.

What is a flat-rate AI API subscription?

A flat-rate AI API charges a fixed monthly or annual fee and does not meter tokens on the invoice. Capacity is bounded by other means: concurrent requests, requests per minute, daily allowances, and per-request input and output ceilings. Flat pricing suits predictable, steady workloads where the subscription costs less than metered usage for the same completed-task volume. It is a poor fit for projects that run for one week a month and sit idle afterwards. The test is utilisation, not the sticker price: a plan used at 5% of its capacity can cost more per task than a metered bill. See how unlimited API access is defined.

How do committed enterprise contracts differ from self-serve plans?

Committed contracts trade a volume promise for a lower unit rate, reserved capacity or private deployment terms. They often add a minimum spend, an annual true-up, price protection and a service schedule. The risk shifts to the buyer: if usage falls below the commitment, the difference is still payable. Ask what happens at half utilisation, whether unused commitment rolls over, whether overage is billed at list or a pre-agreed rate, and how exit works. A committed quote is only comparable after those terms are normalised against a self-serve plan. The provider comparison sets out the trade-offs.

What unlimited means in an AI API contract

Unlimited is the most misread word in AI pricing. It needs a definition in the contract, not a badge on a pricing page.

Does unlimited mean no token cap?

Only when the contract states that there is no monthly token cap and no per-token overage. That is Plugsky's canonical definition: unlimited means no monthly token cap and no per-token overage, not infinite capacity. Fair-use concurrency and requests-per-minute limits still apply, and each model's context window still applies to each request. On Builder, that means 10 concurrent requests and 600 RPM; on Scale, 16 concurrent requests and 1,200 RPM. A provider that advertises unlimited but throttles to a handful of concurrent calls is selling a different product. Put the definition and the remedy in the contract.

Which limits replace tokens on a fair-use plan?

Concurrency, requests per minute and per-request input and output ceilings become the binding constraints. They determine throughput under load, queue behaviour and the shape of a burst. A plan with no token cap but four concurrent requests can feel slower than a metered plan with twenty. Test at two to three times expected peak, watch for HTTP 429 responses, and measure time to first token and total completion time. The uptime and capacity schedule matters as much as the token clause.

Plugsky self-serve plans, checked 20 November 2026.
PlanPriceRequests/dayToken allowanceConcurrentRPMInput / output ceiling
Free$010050,000/day——Micro, Lite
Hobby$8/mo1,0002M/day412064K / 8K
Starter$20/mo2,0005M/day6240128K / 16K
Builder$60/mo—No monthly cap10600256K / 32K
Scale$120/mo—No monthly cap161,200512K / 32K

Builder and Scale are governed by concurrency and RPM rather than a published daily request allowance, with no monthly token cap and no per-token overage. New free accounts from 2 November 2026 receive 100 requests and 50,000 tokens per day; older accounts retain 500 requests and 250,000 tokens per day until migration. Annual billing is 20% off. Paid plans include a 7-day money-back guarantee, and new monthly customers get 30% off the first month, capped at $20. The Scale input figure is a plan ceiling; the individual model context still applies. All figures checked 20 November 2026.

How to compare total cost of ownership

Total cost is not the sum of token prices. It is the cost of producing a correct outcome, including the work that never reaches the model.

Why is cost per million tokens a poor comparison?

Because it prices an input, not the output your product needs. Two providers can charge the same per million tokens and deliver very different cost per resolved support ticket, because one needs longer prompts, more retries or a stronger model for the same quality. The metric also hides embeddings, reranking, storage and human review. Use cost per million tokens only as a unit rate inside a wider model. Instrument one real workflow for two weeks and count calls, input and output tokens, retries and successful completions before you trust a headline rate.

What belongs in an AI API total cost of ownership model?

At minimum: request volume, input-to-output ratio, retry rate, embedding and reranking calls, vector storage, egress, support tier and engineering time. Add one-off migration, evaluation and compliance work, and the cost of private deployment if you need it. Separate fixed from variable cost: a flat plan is mostly fixed, a metered plan is mostly variable, and a committed contract is fixed with a variable tail. The output should be cost per completed task at median and peak load, not a single monthly average. That is the number to compare in an RFP.

How do you calculate the break-even point for a flat plan?

Break-even tasks equal the monthly plan price divided by the measured metered cost per completed task. Run the calculation at median and peak load, because retries rise under pressure. Then add the value of predictable budgeting and subtract the engineering cost of managing rate limits and spend alerts. A flat plan can be worth a premium for a product with steady traffic and a hard budget; it is rarely worth it for a prototype that runs for one week a month. Recalculate after any model or prompt change, since both move the cost per task.

What to verify in an AI API contract

Pricing terms belong in an exhibit, not a slide. These are the clauses that change the number you actually pay.

  • Unit definition. What counts as a token, whether input and output rates differ, and whether cached input, images and reasoning tokens are metered.
  • Retry billing. Whether failed, timed-out or duplicated requests are charged, and how the provider identifies a retry.
  • Caps. Daily and monthly allowances, concurrency, RPM and per-request ceilings, with soft versus hard limits stated.
  • Overage. The rate, the notification threshold, and whether the provider can stop service instead of billing.
  • Price protection. Notice period, any cap on increases, currency and tax treatment.
  • Commitment. Minimum spend, true-up, rollover and the consequence of lower usage.
  • Termination. Refund terms, data export, deletion and transition assistance.
  • SLA. The uptime target, measurement method, exclusions and credit mechanism. Hobby is not in the SLA table.
  • Model changes. Deprecation notice, model-ID stability and whether routing can change the serving engine.
  • Data terms. Training use, retention, subprocessors and the DPA that governs the processing.

A quote is not a price until the unit, the cap and the remedy are written down. Ask for the pricing schedule as an exhibit, not a slide, and have counsel review the final terms before signature.

What we did not verify. We did not audit any provider's invoicing system, negotiate an enterprise quote, or test billing under failure. Competitor list prices change frequently, so this page links to official pricing pages instead of restating rates that may be stale. The Plugsky figures are published self-serve terms checked on 20 November 2026 and are not a quote.

Frequently asked questions

Is a flat-rate AI API cheaper than per-token pricing?

Not automatically. Flat rate wins when steady utilisation is high enough that the subscription costs less than metered usage for the same completed work. Per-token wins when volume is low, spiky or dominated by occasional long-context jobs. Model the break-even point with your own two-week usage sample, including retries and embedding calls, before choosing. A plan that looks cheap per month can be expensive per task if you use a small fraction of its capacity.

What does no per-token overage mean?

It means the invoice does not add a charge for tokens used above an allowance. On Plugsky Builder and Scale there is no monthly token cap and no per-token overage; capacity is governed by concurrency, RPM and per-request input and output ceilings. Read the fair-use clause as well, because it can still throttle or queue traffic at peak. The phrase should appear in the pricing schedule, not only in marketing copy.

How is a token counted?

Token counts depend on the tokenizer, so the same paragraph can produce different counts on different providers. Ask whether the count includes system prompts, tool schemas, images and reasoning tokens. For budgeting, measure with the provider's own token-counting endpoint or the usage fields in the response rather than estimating from characters. Keep a sample of real requests and re-measure after any model or prompt change. The usage object returned with each response should be the authoritative record when you reconcile an invoice.

Are failed or retried requests billed?

That depends on the contract and the failure mode. A request rejected before inference may not be billed; a request that times out after generation may be. Agent loops multiply this risk because one user task can trigger many calls. Ask for the billing rule in writing and make tool side effects idempotent so retries do not duplicate work. Instrument retries per completed task; it is one of the most commonly missed lines in a cost model.

Does a larger context window cost more?

Usually yes, because you send more input tokens, even if the per-token rate is unchanged. Long-context requests also consume more memory and can reduce throughput for everyone on the same capacity. The cheaper pattern is retrieval: send the few passages that matter instead of the whole corpus. Check the per-request input ceiling as well as the model's advertised context window, because a plan can cap input below the model limit.

What is a fair-use policy in AI pricing?

A fair-use policy is a set of operational limits that protects shared capacity when a plan has no token cap. It usually covers concurrency, requests per minute, burst behaviour and queueing. Fair use is not a licence to throttle arbitrarily; the limits and remedies should be written in the contract or published plan terms. Ask what happens at 80% and 100% of a limit, and whether the provider queues, throttles or fails the request.

Do embeddings count toward usage?

On metered plans, embeddings are usually billed by input tokens and should be included in the cost model. On flat plans, they count toward the same fair-use capacity as chat requests. RAG pipelines can generate far more embedding calls than chat calls during a backfill, so budget for re-indexing separately from steady-state query traffic. A model change that alters embedding dimensions forces a full re-embed, which is a cost event as well as an engineering one.

How do I budget for retries and agent loops?

Measure calls per completed task, not calls per user request. Add a retry allowance for timeouts and rate limits, cap agent steps, and log token usage per step. A workflow with a fixed three-step chain is easier to forecast than an open-ended agent that can call tools until it decides to stop. Put a hard step limit and a per-task spend ceiling in the application, not only in the provider plan.

What should an enterprise AI API quote include?

It should define the unit, committed volume, overage rate, concurrency, support tier, SLA credits, price protection, termination and data-processing terms. Ask for the private deployment scope and any per-region surcharge in writing. A single headline rate without those terms is not comparable to another quote. Request the pricing schedule and the DPA together, because the data terms can change the deployment architecture and therefore the cost. Ask every bidder for the same schedule so the comparison is like for like.

Can I cap spending on a flat-rate plan?

A flat-rate plan already caps the subscription fee, but not the operational cost of retries, queueing or engineering. Ask for usage alerts, concurrency controls and a hard stop for non-production keys. If you need a hard financial ceiling, a metered plan with a spend cap can be easier to govern than an unlimited plan with fair-use limits. The right answer depends on whether your finance team needs a predictable invoice or a hard maximum.

When should we renegotiate AI API pricing?

Renegotiate when your usage profile changes materially: sustained higher concurrency, a new region, a private deployment, or a shift from chat to agent workloads. Bring measured cost per completed task and peak RPM data. Renewal is also the moment to fix price-protection, exit and model-deprecation terms. Start the conversation before the renewal window closes, because the leverage is highest while the contract is still competitive. Check the notice period in the current agreement first, so a late request does not roll the term automatically.

Sources

Primary sources and official documentation for this 2027 planning guide — evidence checked 20 November 2026.

  1. OpenAI API pricing — checked 20 November 2026.
  2. Anthropic API pricing — checked 20 November 2026.
  3. Google Cloud Vertex AI pricing — checked 20 November 2026.
  4. Amazon Bedrock pricing — checked 20 November 2026.

Last checked 20 November 2026.

Related reading

Price the workload, not the token

Model cost per completed task with your own traffic, then compare flat and metered plans on the same unit.

Start Free → Read the docs