Rent the model, own everything around it

You do not own the model your business runs on. You rent access to it on terms the vendor publishes, and the lease has a notice period written down in a document most people never open. Anthropic deprecated Claude Opus 4.1 on 5 June 2026 and retired it on 5 August 2026, 2 months later [1]. OpenAI shuts down gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o3-mini and o4-mini on 23 October 2026 [2]. Neither of those is a scandal. Both are the product working exactly as sold.

The useful response is not to panic about vendor risk, and it is not to go build your own model. It is to know which parts of your setup are rented, which parts you already own, and what a forced move would cost you in hours. This guide is for a solo operator or small team with at least one workflow that would visibly break if a named model changed under it. It is not for anyone who wants a fine-tuning tutorial, and it is not for a team with a procurement process and a vendor contract that says something different from the public policy. Every price and date below was read from the vendor’s own page on 4 September 2026.

The notice period is the only guarantee you actually have

Read the deprecation policy of any model you depend on, because it is the closest thing to a contract term you get on a consumer or standard API plan. Anthropic’s commitment is that it “notifies customers with active deployments for models with upcoming retirements, providing at least 60 days’ notice before model retirement for publicly released models” [1]. OpenAI’s is longer at the top and shorter at the bottom: at least 6 months for generally available models, at least 3 months for specialised variants such as chat, Codex and deep research versions, and for preview models “much shorter notice, such as 2 weeks” [2].

That spread matters more than the headline numbers. If a workflow of yours calls a preview model because it was the only one that did the thing, you have a 2-week lease [2]. If it calls a flagship, you have months. Same code, same invoice, very different exposure, and the difference is a line in a public document rather than anything you negotiated.

The second thing worth copying into your calendar is the earliest-retirement floor. Anthropic publishes one for active models: Claude Opus 5 not sooner than 24 July 2027, Claude Haiku 4.5 not sooner than 15 October 2026 [1]. A floor is not an announcement, and a model can outlive it by years. But a floor about 6 weeks out is a different planning object from one nearly a year out, and it costs nothing to know which one you are sitting on.

Models are not the only thing that gets retired. OpenAI shut down the Assistants API on 26 August 2026 in favour of the Responses and Conversations APIs [2]. Anthropic deprecated the temperature, top_p and top_k parameters for Claude Opus 4.7 and later, and setting them to non-default values now returns a 400 error [1]. Interfaces, parameters and defaults move on the same clock the models do. If your automation hard-codes any of them, it is renting those too.

The four things you can actually own

You cannot own a frontier model. You can own four things around it, and all four are cheap.

The words are first. Every instruction you have tuned over months is an asset, and it is worth exactly nothing to you if it lives only inside a vendor’s custom-instructions box. A text file, a Notion page, a repository, anywhere you control, moves to a new model in a paste. The same words saved inside an account do not move at all.

The records are second. Outputs, decisions, the numbers a workflow produced last quarter. If that history only exists as chat history, the vendor holds your archive and you hold a login.

The workflow is third, and it is the one people underrate. The sequence of steps that turns a form submission into an invoice is your business logic. It should live somewhere that treats the model as one replaceable step rather than as the platform.

The fallback is fourth. A second model you have actually run your real work through, recently enough to remember what it got wrong.

Owning the weights themselves is a real option now, and it is a narrower one than the pitch suggests. OpenAI’s gpt-oss-120b and gpt-oss-20b are released under the Apache 2.0 licence, with the 120b model running on a single 80GB GPU and the 20b model in 16GB of memory, both at up to 128k context [5]. Ollama will run open models on your own machine for free, or on its cloud, and states that prompts are never tracked or trained on [6]. That is a genuine answer for data you cannot send anywhere and for behaviour you need frozen. It is also a machine you now administer, a model that is not the one your prompts were shaped against, and an evening you spend when it stops responding.

Renting is cheap, which is why ownership is not about money

Look at what you are actually paying. Claude Pro is $20 per month, or $17 per month billed annually at $200 up front, and Max starts at $100 per month [3]. On the API side, Claude Sonnet 5 is $2 per million input tokens and $10 per million output, Haiku 4.5 is $1 and $5, and Opus 5 is $5 and $25 [3]. OpenAI’s GPT-5.6 line runs $5 and $30 per million for Sol, $2 and $12 for Terra, and $0.20 and $1.20 for Luna [4]. Ollama’s Pro plan is $20 per month with $60 of usage credits included, while local models stay free [6].

For a small operator, those numbers are not the problem. At the prices above, a workflow processing a few hundred documents a month on a mid-tier model costs less than lunch [3]. The expensive line item is your attention when something changes, and that cost does not appear on any pricing page.

So be honest about what ownership buys. It does not save you money at this scale. It buys three things: predictability, because a model you host behaves the same next quarter; privacy, because data that never leaves your infrastructure skips a category of review entirely; and leverage, because a vendor whose pricing you can walk away from prices differently to you than one whose you cannot. Decide which of those you need. If the answer is none of them, rent, and spend the saved evenings on the business.

Own the workflow before you think about owning the weights

The cheapest ownership move available is to keep your automation somewhere you can export. n8n Cloud Starter is €20 per month billed annually for 2,500 workflow executions and Pro is €50 for 10,000, and a standard self-hosted version of n8n is available on GitHub [7]. Running that yourself makes you the person who notices at 11pm that nothing has fired since Tuesday, which is a cost that does not show up as a plan price. But even on the cloud plan, a workflow you can export as a file is a workflow you can move.

Build the workflow so the model is one step with a name you can change. Prompt in one node, model call in the next, output written to your own record. When the model behind it is retired, you edit one step and re-test. When the logic lives inside a vendor’s own agent builder, chained from that vendor’s assistant to that vendor’s storage, you rebuild the whole thing, and you rebuild it under time pressure because the retirement date is what reminded you.

The same rule applies to context. Keep the source material the model reads in a store you own, and hand it to the model at run time. Uploading your knowledge base into one vendor’s memory is convenient for about a year, then it is a migration.

A fallback you have never run is not a fallback

Most people’s fallback plan is the name of another company. That is not a plan, it is a brand you have heard of. A fallback is a second model that has processed your actual inputs, produced output you have read, and failed in ways you can describe.

The habit that makes this cheap: once a quarter, take the one workflow that would hurt most if it broke, run 10 real inputs through a second model, and read both sets of output side by side. It takes an hour. What you get for the hour is a number, which is the number of things you would have to fix, and a judgement, which is whether the second model is merely different or actually worse. Both are worth more when you decide them calmly than on the morning a retirement notice lands.

Write down what you find, with the date. The value of that note is that 6 months later, when a price changes or a model is deprecated, you are picking up an assessment rather than starting one.

What still goes wrong

Ownership is being sold as a product, and some of that is ahead of what it can deliver. River AI raised $1.1 billion in August 2026, led by General Catalyst, on an API that lets developers fine-tune open models with reinforcement learning and low-rank adaptation, pitched as training open models “into ones that are truly yours” [8]. A round that size is a thesis with capital behind it, not a finished tool for a two-person business, and a fine-tuned model you own is still a model that ages while the rented ones improve around it. Owning the weights freezes behaviour, which is the point, and freezing is not always what you want.

The numbers here are a snapshot. Prices, notice periods and policies were read on 4 September 2026 and vendors move all of them, sometimes in the direction you would like. Re-read the two deprecation pages before you make a decision based on this guide rather than trusting the figures in it [1][2].

The quieter failure is the checklist becoming the work. It is possible to spend a weekend making a stack portable that produces nothing anyone pays for, and to feel productive doing it. Portability is insurance, priced in hours, and insurance you cannot afford in time is worse than the risk. If a workflow earns you little, let it depend on one vendor and accept that you will rebuild it if it breaks. Spend the hour on the one that would actually hurt.