Most businesses signing AI subscriptions have not done the math past the sticker price. The monthly fee is the visible part. Setup, integration, growing seat counts and renewals are the rest. That doesn’t make subscribing wrong. The question is whether renting a tool for one job makes more sense than owning one.
We build custom AI tools through our AI integration service, so weigh that. Every fact below about Anthropic’s prices, terms and model lifecycle was checked against Anthropic’s own documentation in October 2026.
Short answer: Subscribe when the workflow is new, low-volume, generic or not core to revenue. Consider owning when one workflow is stable, frequent and specific to how you work. On the Claude API you hold the code, prompts and keys, and you own the outputs. Prices and models still change, with notice. Model usage at small-business volumes is often cheap, so the costs that decide it are the build and the upkeep.
What Anthropic’s Prices and Terms Say
As of October 2026, from Anthropic’s pricing page, Commercial Terms and model documentation:
| Topic | What Anthropic says |
|---|---|
| Sonnet 5.5 price | $2 per million input tokens, $10 per million output |
| Haiku 5.5 price | $0.10 input, $0.50 output per million, for prompts up to 100,000 tokens |
| Batch API | 50% off input and output, for requests that can run asynchronously |
| Prompt caching | A cache read costs 5% of the input price on Sonnet 5.5 and Opus 5.5 |
| Outputs | The customer keeps rights to its inputs and owns its outputs |
| Training | Anthropic may not train models on customer content |
| Price changes | Published rates can change, effective 30 days after posting or notice |
| Model retirement | At least 60 days’ notice before a public model is retired |
Sources: Anthropic pricing, Anthropic Commercial Terms, Anthropic on model training, Claude model deprecations.
Why the Subscription Default Exists
Subscribing is the default for good reasons. It’s fast to start, someone else maintains it and the cost is predictable. A badly scoped custom build can be slow and expensive. But a subscription moves the risk rather than removing it, to the vendor’s price list, roadmap and terms.
The Costs Nobody Mentions in the Demo
The sticker price is the seat cost, rarely the whole cost. List the lines a demo doesn’t show:
- Integration: connecting the tool to your CRM, inbox or database, if it connects at all.
- Seats: per-seat pricing grows with headcount, not usage.
- Usage limits: credits or caps, and what happens when you go over.
- Admin: setup, access management and checking it’s used.
- Overlap: other tools that already do part of the job.
The share these add depends on the tool and your team, so put your own number next to each line. The total is your real subscription cost.
The Annual Price Hike Problem
Subscription prices change. Read the terms for how much notice you get and whether your rate is fixed. By the time a price changes, your data, workflows and habits are wired into the tool, so moving is the expensive part.
This is vendor lock-in in practice. It looks like a pricing email you almost delete.
Owning doesn’t remove price changes. Anthropic can update its published rates, effective the earlier of 30 days after posting or notice to you (source). The difference: per-token prices are public, and because the code is yours, you can test another model without rebuilding the workflow.
What Owning Actually Means
“Build” doesn’t mean hiring AI engineers or training a model. For most small businesses, a custom AI tool is one workflow with one defined output, built on a model API such as Claude’s. You hold the code, the prompts and the API keys.
Anthropic’s Commercial Terms say the customer retains all rights to its inputs and owns its outputs, and Anthropic may not train models on customer content (source). Its privacy centre says the same for the API, unless you choose to share data, for example by sending feedback (source).
What you don’t own is the model. Anthropic gives at least 60 days’ notice before retiring a publicly released model, and lists Claude Sonnet 4.5 as deprecated on September 30, 2026, retiring November 30, 2026 (source). Someone has to move your tool to a newer model and test it.
Who owns the IP when an agency builds a custom AI tool?
It varies, so put it in the contract before work starts: the full codebase, every prompt, any data pipelines, and API keys in your own accounts. If the builder keeps the IP and charges an ongoing fee to use it, you’re subscribing again under another name.
The Third Option: Commission a Custom Build
Between subscribing and hiring an in-house team, you can commission one tool from a partner and own it. The cost depends on the workflow, the systems it connects to and the testing it needs, so there’s no honest standard price. Get a written scope first: one job, defined inputs and outputs, and what the tool won’t do.
It works when the workflow is frequent and the scope is tight. It breaks down when the scope drifts, the most common failure on both sides of this decision.
When Subscribing Is Still the Right Answer
Not a caveat. An actual condition.
If you don’t know what success looks like for a use case, you have no basis for scoping a build. A subscription that teaches you what you need beats a build that solves the wrong problem.
Subscribe when any of these is true:
- Volume is low.
- Requirements are unclear.
- The domain changes fast and a vendor tracks it for you.
- The workflow isn’t core to revenue.
- Nobody on your side will own the tool after launch.
When Building Wins, Concretely
This is our own worked example, not a client case. Volumes and token counts are assumptions; prices are Anthropic’s published rates (source).
A firm drafts about 1,000 client update emails a month from CRM data. Assume 3,000 input and 500 output tokens per email.
- Claude Sonnet 5.5: 3 million input tokens ($6) and 500,000 output tokens ($5), about $11 a month.
- As a batch: if the drafts can wait, the Batch API halves that to about $5.50.
- Claude Haiku 5.5: $0.30 input and $0.25 output, about $0.55 a month, if its drafts pass your quality check.
At that volume, model usage isn’t the deciding cost. The build, the time spent reviewing drafts and upkeep when your CRM or the model changes are. Compare those with two or three years of the subscription at your real seat count. Building wins when no subscription does the job your way, the workflow runs daily and owning comes out cheaper over that period. If the numbers are close, keep the subscription.
The Defensibility Question Most Owners Don’t Ask
A subscription gives you what your competitors can also buy. A tool built into your workflows, using your own data, is harder to copy, though that isn’t guaranteed. Many custom AI tools are a thin layer over a model anyone can call. In our view, a tool is defensible only with at least two of: data only you have, workflow depth, domain expertise built into the prompts and checks, and full ownership. If the builder keeps the IP, you don’t have the last one.
The Audit Step Everyone Skips
Before any build or subscribe decision, audit what you already pay for:
- Pull every AI subscription from your billing records.
- Check each admin dashboard for last month’s active users.
- Mark tools that overlap.
- Note each contract’s renewal date and price-change terms.
Cancel what nobody uses. Then pick the one workflow where your tools fall short. That’s the candidate for a build, or a better subscription.
Related Posts
- Build vs buy AI tooling: a decision framework for business owners
- The full cost of an AI integration: what the build quote doesn’t cover
- Ownership vs dependency in AI contracts: what the contract should say
- AI vendor lock-in risks: what to check before you sign
FAQ
How do I know if my business should build a custom AI tool or keep subscribing?
Compare two or three years of total cost on each side. A stable, frequent workflow that subscriptions serve poorly makes a strong build case. If the scope is unclear, subscribe first.
What does owning a custom AI tool actually mean in practice?
You hold the code, prompts and API keys in your own accounts. On the Claude API, Anthropic’s terms say you own the outputs and it may not train on your content. You pay per token plus maintenance, and you still plan for model upgrades.
How much does it actually cost to build a custom AI tool for a small business?
It depends on the workflow, the integrations and the testing, so we don’t quote a standard price. Get a written scope and price that. Running costs can be estimated from published per-token prices once you know the volume.
What are the real hidden costs of AI subscriptions that vendors don’t mention?
Integration, seat growth, overages on usage-priced plans, price changes at renewal and the admin time for overlapping tools. Read the pricing page and contract before you sign.
Can I switch from a subscription AI tool to a custom build later?
Yes, with a switching cost if your workflows are shaped around the vendor’s tool. Check you can export your data. Keep the subscription while the build is tested, and switch once it’s stable.
What’s the minimum workflow volume that justifies a custom build?
There’s no universal threshold. It depends on the staff time involved, the subscription cost at your seat count and the build and upkeep cost. A workflow that runs rarely usually stays off-the-shelf, because there’s little to save.