The Rent Is Due: Why Small Teams Can Finally Build Instead of Subscribe

A capable mid-tier AI model now costs $2 per million input tokens and $10 per million output, and it scores within a few points of the flagship on most published tests. That price-to-quality ratio changes who gets to build software. For the first time, a freelancer or a three-person shop can offer a small business something close to “built in house” instead of another monthly subscription.

I have run Linux servers since the dial-up days. I watched software go from something you owned to something you rented, one seat at a time. This month, the direction reversed a little, and small businesses should notice.

What changed in September 2026

Two releases, six days apart, moved the price floor for serious AI-assisted work. Anthropic shipped Claude Opus 5.5 on September 22 and Claude Sonnet 5.5 on September 28.

Opus 5.5 got cheaper, which almost never happens with flagship models. It dropped to $4/$20 per million tokens, 20% below Opus 5, and Anthropic says it performs at the level of its more expensive Fable 5.1 on most work (Anthropic).

Sonnet 5.5 kept Sonnet 5’s price but closed most of the gap to Opus. These are Anthropic’s own launch numbers:

Sources: The Next Web, MarkTechPost.

The mid-tier model now beats the flagship on terminal work and ties it on office-style tasks. Anthropic also says it is 30%+ faster than Sonnet 5 and costs up to 30% less per task, because it uses fewer tokens for the same job (Anthropic).

The price of “good enough” collapsed

The quality that cost flagship money a few months ago now sells at mid-tier rates. List prices per million tokens, in US dollars:

Sources: Anthropic, BenchLM, Alphacorp.

A simple illustration at list price, before caching discounts: a project that consumes 10 million input and 2 million output tokens costs about $40 on Sonnet 5.5, $80 on Opus 5.5 and $200 on Fable 5.1. For a small business, that is the difference between a rounding error and a budget line.

The per-token price is only half the story. Independent testing by Artificial Analysis puts Sonnet 5.5 at $0.41 per task on low effort, but $7.60 per task at max effort. Opus 5.5 costs $0.55 per task on low effort and $5.98 at max (Artificial Analysis).

So the cheap model is only cheap when you use it sensibly. The practical rule for small teams:

• Sonnet 5.5 at low or medium effort for daily work: well-scoped features, bug fixes, documents, admin scripts.
• Opus 5.5 for the hard thinking: architecture, tricky migrations, anything where one wrong decision costs a week.
• Never max effort by default. At the top setting, the mid-tier model can cost more per task than the flagship.

You don’t need the API to start, but don’t cheap out on the plan. For professional work, the subscription is the workshop, and running out of capacity mid-project costs more than the plan itself. The $20 Pro plan is for trying things out, not for running a practice.

• Solo freelancers: Max 20x, $200 a month. Serious client work means long agentic sessions most days, and hitting a limit mid-task breaks the flow.
• Small studios, up to about five people: a mix of Max ($100) and Max 20x ($200) per person. The builders who live in the code take 20x. The people who spend more of the day with clients, planning and paperwork take Max. This only works when the team knows each other well enough to spread the workload honestly.

A five-person studio lands between $500 and $1,000 a month (Suprmind). And the model does not replace the human part: the client meetings, the planning and the paperwork are still yours.

From renting to owning

The real shift is not that AI writes code. It is that one competent person with a good model can now deliver what used to need a team.

Most small businesses do not need a platform. They need a booking flow, a stock list that talks to their invoicing, a customer portal, a report that lands in someone’s inbox on Monday. For fifteen years, the only affordable answer was a SaaS product built for a thousand other companies. You bent your process to fit the tool, and you paid per seat, forever.

Now the math looks different. Take a hypothetical 10-person team on a tool at €30 per seat per month: €3,600 a year, every year, at a price you do not control. Compare that with a small custom tool, built with AI assistance by a freelancer, running on open-source components and a modest EU server:

This is not free. Someone still has to design it, secure it, back it up and patch it. But that someone can now be a local freelancer on a fair retainer, instead of a shareholder three time zones away.

The landlords

Large software vendors rarely behave like suppliers to a small business. They behave like unelected authorities: they write the rules, change them overnight, and you get no vote.

Every small business owner I work with knows the pattern:

• A price increase arrives by email, effective next billing cycle.
• The plan you chose is “retired” and the replacement costs more.
• A feature you rely on moves to a higher tier.
• Your data is exportable in theory, in a format nothing else reads.
• The terms of service change, and continuing to log in counts as consent.

None of this is illegal. It is simply what happens when one side holds the code, the data and the switching costs. That is not a free market. A free market needs customers who can walk away.

Ownership restores that. When the business owns its code and data, the developer, the host and even the AI model all become replaceable suppliers again. That is capitalism working the way the textbooks promise: competition on quality and price, not on how hard it is to leave.

The honest part

An article about escaping big vendors, built on a big vendor’s model, owes you some caveats.

The AI company is also a landlord. Anthropic sets its own rules too. Sonnet 5.5, for example, routes higher-risk cybersecurity requests back to the older Sonnet 5 (Help Net Security). The difference is where the dependency sits. Use the model to build, not as a runtime your client’s business cannot live without. The code it writes stays yours, and you can switch models tomorrow.

Benchmarks are not your workload. Most figures above come from the vendor’s launch material. In a hands-on test by DataCamp, Sonnet 5.5 and Sonnet 5 cost almost the same to finish one task: $0.56 versus $0.57. Test on your own work before you promise savings to a client.

Cheap code is not cheap software. A model can produce a working app in an afternoon. Security, backups, updates and monitoring are still a professional’s job, and now they are your job.

Keep the model out of production. The model writes the software, the automation and the infrastructure as code, from the first line to fully managed servers. It works only in development and test environments, on test data. Production is deployed and operated by the contractor, running reviewed runbooks themselves. Client data stays away from the model, and a human answers for every production change. That is the safer design for security, accountability and data protection alike; the data protection side deserves its own article.

Not everything should be rebuilt. Email, payroll and regulated accounting are better bought from specialists. Own what makes the business different; rent the commodity.

A fairer market starts small

This is not a revolution against capitalism. It is a return to a version of it where the small player can compete.

The tools that once needed a funded team are now available to anyone willing to learn them properly, for $100 to $200 per person per month. That puts real power back in the hands of freelancers, small agencies and the businesses they serve. The corporations did not lose their power this month. They just lost part of their monopoly on capability.

If you run a small business, ask one question at your next renewal: what would it take to own this instead? If you are a freelancer or a small team, learn to answer that question well. Do it with open-source foundations, EU hosting, honest pricing and code your client keeps. That is how the independent side of tech gets its turn.

Sources

Prices are list prices in US dollars as of 29 September 2026. Benchmark figures are mostly vendor-reported unless marked as independent.

• Anthropic: Introducing Claude Sonnet 5.5
• Anthropic: Claude Opus
• The Next Web: Sonnet 5.5 launch and cyber limits
• MarkTechPost: Sonnet 5.5 benchmarks and pricing
• Help Net Security: Sonnet 5.5 safeguards
• Artificial Analysis: Sonnet 5.5 (independent)
• Artificial Analysis: Opus 5.5 (independent)
• DataCamp: Sonnet 5.5 hands-on test (independent)
• BenchLM: Sonnet price history
• Alphacorp: Claude 5.5 lineup pricing
• Suprmind: Claude Max 5x and 20x pricing