The parts of an AI vendor you can hold it to

You picked a model for a reason you could say in one sentence. It was the cheap one. It was the fast one. It was the one with the generous free tier. That reason went into a config file a long time ago and has been quietly true ever since, which is the same as saying you stopped checking it.

Then the reason stops being true. On 16 August 2026, DeepSeek raised peak-hour pricing on DeepSeek-V4-Flash from $0.28 to $1.32 per million output tokens, and on DeepSeek-V4-Pro from $0.87 to $3.96 per million tokens, with off-peak hours at half the peak price, saying it was revising pricing “to allocate resources more reasonably” [8]. Nothing was broken and nobody was misled. A vendor whose appeal was its price decided to charge more, which was always something it was allowed to do. This guide is about the difference between what a vendor is currently like and what a vendor has actually committed to, and about rebuilding your choices on the second thing. It is written for a solo operator or small team choosing between models on public terms. If you have a negotiated contract with committed spend, your commitments live in that contract and you should read that instead.

What you bought was a position, not a property

“Cheapest” is not a feature of a company. It is a place the company is standing, relative to other companies, on a day you happened to look. Standing there costs the vendor something, and when the cost changes or the strategy changes, it moves. The same is true of “fastest”, “most generous free tier”, and “the one that lets you do anything”.

You can watch this happen in advance if you read pricing pages for their dates rather than their numbers. Google’s Gemini API pricing page currently lists Gemini 3.8 Flash at $0.75 per million input tokens “through December 31, 2026” and $1.50 “starting January 1, 2027”, with output going from $3.75 to $7.50 on the same day [6]. That is not a warning sign. It is a vendor being unusually straightforward about the fact that a price is a decision with a term attached. Most vendors do not print the second number, which does not mean it does not exist.

So when the position changes, the honest reaction is not betrayal. It is the recognition that you were relying on something the vendor never promised. The useful question is what it did promise, and where that is written.

A property is something you could show a third party

Here is the test that separates the two. Imagine explaining your vendor choice to somebody who does not trust you. You have to point at a sentence on a page the vendor published, and that sentence has to say the same thing to them as it says to you. A price satisfies that test only until the page changes, and pricing pages change without ceremony. A licence, a notice period, a data-use clause and a storage jurisdiction satisfy it more durably, because they are written as commitments rather than as offers.

There are roughly five of these worth recording for any model you depend on. What licence the weights or the service ship under. How much notice you get before a model you are calling is retired. Whether the inputs you send are used to train the vendor’s models. Which country your data sits in. Whether the interface you call is one that something else also speaks. None of them is about quality, and that is deliberate. Quality you can test in an afternoon. These five you cannot test at all, which is why they have to be read.

Notice periods are written down and they are not the same

The clearest example of a real property is the retirement policy, because Anthropic and OpenAI both publish one and the two are not equivalent.

Anthropic describes four lifecycle states for a model, active, legacy, deprecated and retired, and commits to “at least 60 days’ notice before model retirement for publicly released models”, with deprecated models given a recommended replacement and a retirement date [4]. That policy has been exercised in public: claude-opus-4-1-20250805 was notified of deprecation on 5 June 2026 and retired on 5 August 2026, with claude-opus-4-8 named as the replacement, and requests to a retired model fail [4].

OpenAI publishes a longer floor and a wider spread. Generally available models get “at least 6 months”, specialised variants such as chat or Codex models get “at least 3 months”, and preview models can be retired with much shorter notice, “such as 2 weeks” [5]. Its deprecation log shows what that looks like in practice: the June 2026 entry deprecating the gpt-5-2025-08-07, o3-2025-04-16 and o3-pro-2025-06-10 snapshots names gpt-5.6-sol as the replacement for all three and a shutdown date of 11 December 2026 [5].

Sixty days and six months are both real commitments and they are different commitments. If you have a workload that needs a quarter’s notice to move, one of those numbers works for you and the other does not, and no amount of per-token savings changes that arithmetic. The same log tells you something else worth knowing, which is that retirement happens to models on every price tier. You will be migrating regardless. The only variable you get to choose is how much warning is in writing.

Your tier is part of the terms, not just the price

The tier you are on is usually presented as a price difference. It is often a terms difference wearing a price difference as a costume.

Google’s Gemini API terms are explicit about this. On the free tier, Google “uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products”, and “human reviewers may read, annotate, and process your API input and output”, although the terms say this includes “disconnecting this data from your Google Account, API key, and Cloud project before reviewers see or annotate it” [7]. On the paid tier, Google “doesn’t use your prompts (including associated system instructions, cached content, and files such as images, videos, or documents) or responses to improve our products”, and logs prompts and responses for a limited period solely to detect and prevent violations of the Prohibited Use Policy [7]. Same API, same model names, materially different arrangement.

Jurisdiction sits in the same category. DeepSeek’s privacy policy, last updated 10 February 2026, states that “we directly collect, process and store your Personal Data in People’s Republic of China”, and lists among its purposes “to improve and develop the Services and to train and improve our technology, such as our machine learning models and algorithms”, while giving users “the right to opt-out of using your Personal Data for training our models or optimizing our technologies” [3]. Every one of those is checkable, none of them moved when the prices did, and all of them matter more than the peak-hour move from $0.28 to $1.32 per million output tokens [8] if you handle client material under a confidentiality clause.

The practical consequence is that “we moved to the free tier to save money” is not a budget decision. It is a change to what you have agreed to, and if you resell the output or process somebody else’s data, it is a change you may have promised not to make.

Portability is a property of the interface, not of your good intentions

Everyone plans to stay portable. What makes portability real is not the plan, it is whether the thing you call speaks a shape that something else also speaks.

DeepSeek’s own documentation is a clean example. It states that “the DeepSeek API uses an API format compatible with OpenAI/Anthropic” and that “by modifying the configuration, you can use the OpenAI/Anthropic SDK or softwares compatible with the OpenAI/Anthropic API to access the DeepSeek API”, with base URLs at https://api.deepseek.com and https://api.deepseek.com/anthropic [1]. That is a structural property. It means the cost of leaving, or arriving, is a base URL and a model name rather than a rewrite, and it is still the wording on the docs page today, after the August repricing [1][8].

Open weights are the same kind of property one layer down. The DeepSeek-V4-Pro model card states that the repository and the model weights are licensed under the MIT License [2]. You may never run a 1.6 trillion parameter model with 49 billion activated parameters yourself [2], and that is fine. The point is that a published licence is not revocable by a pricing decision, so the model remains available from somebody, somewhere, at some price, in a way that a closed model on a discontinued tier does not.

So the portability question to ask about any vendor is not “could I leave if I had to”. It is “what exactly would I edit”. If the answer is a line of configuration, you have a property. If the answer is a weekend, you have an intention.

Re-pick on the written list and let price break the tie

The method is dull and takes about twenty minutes per vendor. For each model you actually depend on, write five lines: licence, notice period, training use on your tier, storage country, and what a competitor’s endpoint would need in order to receive the same calls. Link each line to the page you got it from. Date the record, because every one of those pages can be revised and the date is how you know when your note stopped being evidence.

Then choose. Rule out anything whose written properties do not fit the work, which for most people is a short list and a quick pass. Among what survives, compare price, and treat that comparison as the tiebreaker it is rather than the criterion it has been pretending to be. This inverts the usual order, and the inversion is the whole point: you can absorb a price change in a month, and you cannot absorb the discovery that a year of client material was processed under terms you never read.

Re-run it when something forces you to look anyway, which in practice means a repricing, a deprecation notice, or a new client asking where their data goes. Do not put it on a calendar. Calendar reminders for this get dismissed.

What still goes wrong

Written properties are more durable than positions, not permanent. Policies get revised, and the revision date is usually the only clue: DeepSeek’s privacy policy carries a 10 February 2026 update [3], which tells you the current text is current and nothing about the previous text. A notice period is a promise about process, not about quality, price or the company continuing to exist. Reading all five lines correctly still leaves you exposed to a vendor that simply stops.

The properties also do not rescue you from the thing they describe. A published deprecation policy guarantees warning, not that the replacement model will behave like the one you tuned your prompts against, and Anthropic’s own page tells you a retired model’s requests fail rather than degrade [4]. An MIT licence on a 1.6 trillion parameter model [2] is a real right that most solo operators cannot exercise on their own hardware. Compatible API shapes get you a config change rather than a rewrite, and still leave you to re-test the outputs.

And this whole approach costs you something. Picking on written properties rather than price means sometimes paying more than the cheapest option to stay on terms you can defend, and the difference is real money you could have kept. If your AI use is a personal chat subscription with no client data in it, the five lines are overkill and the cheapest thing that works is the right answer. The moment somebody else’s material passes through the API, the order flips.