AI subscription, API or open model: what costs your small business less?

In 60 seconds

For a B2B business of 5 to 50 people, the cheapest AI is the option that completes your task at the required quality and controls for the lowest total cost. If a person works in a chat interface, evaluate subscription seats. If you need to process orders or update a CRM without copying and pasting, try an API. An open model accessed through an API still needs integration; hosting it yourself adds infrastructure operations. Compare the same tasks over one or two weeks and measure cost per accepted task. Revolut provides context about scale, not a savings promise for your business.

Verdict by profile

• If sales and management occasionally consult documents or prepare drafts, start with a subscription and review which seats get used. Choose data and administration controls before buying extra capacity.
• If operations needs an agent to read orders and prepare ERP changes, compare a closed-model API with an open-model API. Both need someone responsible for integration.
• If you have sustained, predictable demand, a team to operate the service and a specific cost or control requirement, investigate dedicated GPUs or self-hosting with a complete budget.
• If you do not yet know which data you may send or who reviews errors, settle that before buying another option.

These are editorial starting points for an evaluation, not findings from a Kiia benchmark.

What not to copy from Revolut

In his AI Engineer NYC 2026 talk, David Abitbol reports roughly 2,000 Revolut engineers, 56 B200 GPUs, a blended USD 0.30 per million tokens and a bill one-tenth of its previous Anthropic and OpenAI spending. He describes tuning, a sandbox with human oversight, zero retention and telemetry; weights run on a US-based service. Claude Code and Codex remained available as exceptions. These are Revolut’s deployment and measurements, checked against the transcript on November 4, 2026.

The transcript contains inconsistent GLM version names. The version served today and GPU ownership remain unconfirmed; we do not describe the machines as hardware Revolut bought. It also provides no breakdown that would let us include staffing, tuning and on-call work in that USD 0.30, or establish a comparison at the same scope. Do not extrapolate the figure to a company of 20 people. The internal productivity metrics do not isolate the model’s effect either.

Separately, the official GLM-5.1 model card publishes weights under MIT. That establishes this version and its license, not the version in Revolut’s current deployment. For weights and licensing, see open-weight models for your small business rather than repeating its hosting routes here.

Comparison at the same scope

Compare one outcome: for example, an order draft a person accepts after checking its fields. There are four payment routes within the title’s three categories. A subscription buys access for people; an API buys consumption; hosting reserves capacity you pay for even when idle.

Criterion	Subscription per person	Closed-model API	Open model through an API	Open model you host
Verified entry price	Seat fee; official examples below. Extra usage depends on the plan.	Tokens and tools; examples below.	Tokens and provider services; examples below.	No universal tariff: obtain infrastructure and operations quotes.
Where processing and data live	In the contracted service; check the plan’s region and retention.	In the contracted endpoint and your logs.	At the inference provider and in your logs; open weights do not mean local data.	In your chosen infrastructure; check backups and external connections too.
MCP status	Not determined by payment type; verify product and plan support.	Not determined by pricing; verify model, client and connector.	Open weights do not guarantee it; verify the implementation.	Depends on the client and servers you deploy; hosting a model does not establish support.
Authentication	Individual identities; manage onboarding and offboarding.	Service credential as required by the API; separate environments.	Provider credential; the same integration work.	You must implement endpoint authentication and access control.
Respects user permissions	Check each connector’s permissions.	Your application must enforce them.	Your application must enforce them.	Your application and infrastructure must enforce them.
Integrations	Check the plan’s native integrations; a seat fee does not build a custom workflow.	You maintain connectors, validation and recovery.	You maintain that same work.	You maintain integration and inference serving.
Limits	Plan quotas, models and usage windows.	Rate limits, context and budget.	Provider limits and the version’s capabilities.	Memory, concurrency and capacity availability.
LATAM support	Verify country, billing and support for the exact plan.	Check the API’s regional list and billing.	Confirm country, actual processing region and contracting entity.	Confirm location, support and on-call coverage; GPUs do not create local support.

These rows are purchasing and design checks. They do not certify a provider’s connectors, permissions or residency. OAuth authentication, for example, does not establish that an integration respects each user’s permissions.

Official prices: verified on November 4, 2026

Web references in USD, not quotes for every country. Each row links to the official source checked that day. Annual rates are monthly equivalents with an annual commitment; they do not buy a single month.

Subscription	Reference price and billing	Official source · checked
ChatGPT Plus / Pro	Plus: USD 20/month. Pro: USD 100, 200 or 500/month.	OpenAI · Nov 4, 2026
ChatGPT Business	From USD 25/user/month billed monthly or USD 20 billed annually; 2-user minimum.	OpenAI · Nov 4, 2026
Claude Pro / Max / Ultra	Pro: USD 20/month or USD 200/year. Max: USD 100/month. Ultra: USD 200/month.	Anthropic · Nov 4, 2026
Claude Team	Standard: USD 25/seat/month or 20 billed annually. Premium: USD 125 or 100, respectively. For teams of 2 to 150.	Anthropic · Nov 4, 2026
Google AI Pro, personal account	USD 19.99/month on the US page. For business accounts, check the Workspace offering.	Google, account type · Nov 4, 2026

For APIs, these are examples of text rates, not models of equivalent quality or provider recommendations. Prices are per million tokens, at standard consumption rates without batch discounts; tools and other services cost extra.

API and scope	Input / cached input / output, USD	Official source · checked
OpenAI GPT-6.1 Sol, up to 272,000 input tokens	2.00 / 0.10 / 10.00. Cache writes: 2.50 for 30 minutes; 5.00 for 12 hours.	Pricing, model conditions · Nov 4, 2026
Anthropic Claude Sonnet 5.5	2.00 / 0.10 / 10.00. Cache writes: 2.50 for 5 minutes; 4.00 for 1 hour.	Anthropic · Nov 4, 2026
Google Gemini 4 Argon	2.00 / 0.10 / 10.00 through Nov 30, 2026. From Dec 1: 4.00 / 0.20 / 20.00. Implicit caching; no explicit caching.	Google · Nov 4, 2026
Z.ai GLM-5.1, open weights	1.40 / 0.26 / 4.40. Cache storage is free for a limited time.	Z.ai · Nov 4, 2026

For GPT-6.1 Sol, exceeding that context changes pricing for the whole request; regional processing carries a premium where available. Gemini output includes thinking tokens. Check the linked conditions before projecting usage. Revolut’s blended figure is not a fifth API offer comparable with this table.

Anthropic excludes applicable taxes from displayed prices. For the others, verify taxes, final currency and local invoicing at checkout or in the contract; we do not convert currencies or assume shared tax treatment across LATAM and Spain. The official API lists from OpenAI, Anthropic and Google include Spain and LATAM countries, but check your country and the exact product. Access from a country does not mean processing there. Z.ai billing and support terms for your entity: verify before purchasing.

Some subscriptions already grant API credits. OpenAI documents credits for eligible Pro and Business subscriptions, subject to activation, conditions and expiry; Anthropic advertises them for Max and Ultra. Check your actual grant and avoid counting it twice in the budget. Our guide to OpenAI API credits with Business and Pro explains how to claim them. For seats, consult ChatGPT Pro versus Business and which Claude subscription to choose, alongside today’s official prices.

Decision criteria

Start with who uses AI. A person preparing an email can work in a subscription interface. An order-processing queue needs an integration, even if its first calls use credits included in a plan.

Then consider volume and predictability. Count tasks, peaks, context, retries and human review. Idle GPUs still cost money; API spending can vary with demand. Buying fixed capacity to remove a variable bill changes risk, rather than demonstrating savings.

For data, identify sensitive fields and the countries used for processing, storage and support. Review retention, subprocessors and terms with the person responsible in your business. Self-hosting can also send data through connected tools or telemetry.

Finally, identify who maintains the service. In a small business, diverting the only developer to inference-engine updates, security or on-call work has a cost. A permissive weights license does not cover that work. Both an open-model API and a closed-model API must first pass the same data and quality review.

Why the token price is not the cost

Reuse the formula from The real cost of an AI agent:

Total cost = model + tools + channel + infrastructure + supervision + maintenance + error

Cost per correct result = total cost / completed and accepted tasks

Include integration amortized over an explicit period, updates and failure recovery. A model with cheap tokens may need more retries or corrections. Define operations’ acceptance criteria before measuring; a generated response is not yet a correct order. If no tasks are accepted, do not calculate a unit cost: the option has failed the trial.

When to choose neither

Keep the current process if no alternative meets your permissions, quality or budget requirements. Without a technical owner, self-hosting is usually a poor first purchase for a small business: it adds serving, capacity and security work before proving the workflow.

Revisit that decision when you have measured sustained demand, complete API costs, specific control requirements and a team able to maintain availability. Then request a concurrency, quality and recovery trial on the proposed infrastructure, with operations and migration budgets. There is no universal token threshold or savings percentage we can promise.

If prices rise or limits change

An opinion about the end of subsidized subscriptions is not a pricing announcement. Work from official terms: Gemini 4 Argon’s published change date in the table is a verifiable example, not a prediction about every provider.

• Set a budget and owner for each workflow. Check whether the provider’s limit stops calls or only alerts; add an application cutoff and bound retries and concurrency.
• Put provider-specific details behind an adapter: credentials, model names and formats should be changeable without rebuilding the business process.
• If you use OpenAI-compatible endpoints, test tools, structured output, errors and streaming. The Z.ai quickstart shows client compatibility; it does not guarantee identical features or results.
• Run the same tasks on an alternative before you need it. Document migration ownership, how to export instructions and evaluations, and how to return to the manual process.

Next step: a one- or two-week trial

Choose anonymized real tasks and common exceptions. A synthetic distributor example: extract order fields and prepare a draft, without writing to production. Define mandatory fields, errors that invalidate a result and who accepts it.

Try those same inputs in a subscription, a closed-model API and an open-model API. Keep instructions, available data and acceptance criteria consistent. Record copying and pasting time in chat and API integration effort so the comparison covers the full workflow. You do not need to rent GPUs for this first trial.

Record for each task	Purpose
Route, plan, model and date	Reproduce what was evaluated.
Input, cache and output tokens, tools and retries	Reconcile usage with the bill.
Preparation, review and correction time	Include human work in the cost.
Accepted or rejected result, with reason	Measure quality using the same criteria.
Limits, failures and recovery	Assess whether the workflow can be sustained.

Apply the formula above, allocate the relevant share of subscriptions and integration, and separate promotional credits from recurring costs without them. Compare against the measured manual process too. Decide by task and profile; you can keep chat seats for some people and an API for one workflow. Take the results to your budget review and use the comparisons hub to investigate the chosen option.