Planning a business around an AI industry that disagrees about its own pace

In July 2026, 1,386 people who work on frontier AI signed a public statement asking the US government to support an international effort to develop “the technical and governance tools needed to deliberately pace the frontier of automated AI development” [1]. Seven months earlier, the White House had issued an executive order declaring it the policy of the United States “to sustain and enhance the United States’ global AI dominance through a minimally burdensome national policy framework for AI,” and gave the Attorney General 30 days to establish a task force whose “sole responsibility shall be to challenge State AI laws” [2]. Both documents are live. Neither is going to talk the other out of anything.

If you run a one-person business, or a small team, on Claude, ChatGPT and Gemini, there are two tempting ways to handle this and both are wrong. One is to pick a side and feel informed. The other is to file it under industry noise and skip it. The useful move is narrower: treat the disagreement as evidence that the pace of change at your vendors is contested rather than settled, then go find the parts of that pace your vendors have already committed to in writing. This guide is not for people who build frontier models or work on AI policy, who need the argument on its own terms. It is for the person whose client research, invoicing summaries and first drafts pass through somebody else’s model, and who wants to know which parts of this touch an ordinary working week.

The argument runs through the companies you already pay

The statement is more specific than “slow down,” and it does not ask for a pause. It says the world’s leading AI companies “believe they could be close to automating AI research,” that there is “a real risk that capability development rapidly accelerates beyond our ability to understand or control,” and that each company and country is under “intense competitive pressure not to unilaterally slow acceleration” [1]. What it asks governments for is tooling that does not exist yet, which the statement says the world currently lacks [1]. Signatories are listed at OpenAI, Anthropic, Google DeepMind, Meta AI, Safe Superintelligence and Thinking Machines [1]. That list includes people at all three of the vendors this guide is about, asking for a brake they do not currently have.

The executive order points the other way with equal specificity. It directs the Secretary of Commerce, within 90 days, to “publish an evaluation of existing State AI laws that identifies onerous laws,” the FCC to open a proceeding on whether to adopt a federal reporting and disclosure standard for AI models, and the FTC to issue a policy statement on the application of the FTC Act’s prohibition on unfair and deceptive acts [2]. It is a plan to reduce friction, not add it.

You are not going to resolve that. What you can do is notice what it means for planning. When the people closest to the technology publicly disagree about whether the next two years should be faster or slower, any roadmap of yours that assumes one answer is a bet, not a plan. The good news is that you do not have to forecast the frontier to run a business on it. You only have to know how fast things change one layer down, at the API and the subscription, and that number is published.

Your vendors already tell you how fast they will move under you

Every major vendor publishes a deprecation policy. It is the least-read page they operate and the most useful one for a small business, because it is the vendor’s own written answer to the question the pacing debate is really about: how much warning you get before the thing you built on stops existing.

OpenAI commits to “at least 6 months” for generally available models and “at least 3 months” for specialized variants such as chat, Codex and deep research builds, and warns that preview models “may be retired with much shorter notice, such as 2 weeks” [3]. Anthropic states that it notifies customers with active deployments, “providing at least 60 days’ notice before model retirement for publicly released models” [4]. Google is the loosest of the three. Its Gemini page says “the shutdown dates listed in the table indicate the earliest possible dates on which a model might be retired,” with the exact date communicated to users later [5]. No fixed floor is promised.

Read those three side by side and you have a real, checkable spread: 6 months, 60 days, and an earliest-possible date. That is the pacing debate as it actually reaches you. If a workflow of yours would take a month to rebuild, a 60-day notice period is comfortable and a 2-week preview model is not a place to put it. This is a decision you can make today, without any view on automated AI research.

Retirement is the ordinary case, not the emergency

The second thing those pages tell you is that production models are being retired on a cycle of a little over a year, in all three houses, right now, while the argument about pacing is still unsettled.

Claude Sonnet 4 was deprecated on 14 April 2026 and retired on 15 June 2026, with claude-sonnet-4-6 named as the replacement [4]. Counting from the date carried in its own model id, claude-sonnet-4-20250514, that is 13 months of service. OpenAI has gpt-5-2025-08-07 and o3-2025-04-16 shutting down on 11 December 2026 in favor of gpt-5.6-sol, with gpt-3.5-turbo-0125, gpt-4-0613 and gpt-4-turbo going dark earlier, on 23 October 2026 [3]. That is 16 months for GPT-5, on the same counting. Google lists gemini-2.0-flash as released on 5 February 2025 and shut down on 1 June 2026, replaced by gemini-3.6-flash: 16 months [5].

Releases run faster than retirements. Google’s own table shows five Flash generations arriving between 17 December 2025 and 2 September 2026, from gemini-3-flash-preview to gemini-3.8-flash [5], and shows gemini-3-pro-preview shut down on 9 March 2026, less than four months after its 18 November 2025 release [5]. Preview tiers churn; the models you are meant to build on last about 13 to 16 months.

None of that is a crisis, and none of it waited for a policy answer. It is the background rate. Whether Washington ends up building pacing tools or dismantling state rules, the model you send your prompts to is going to be replaced roughly once a year, and you will get somewhere between two weeks and six months of warning depending on which vendor and which tier you picked [3][4][5].

A migration changes more than the model name

The reason this matters more than a search-and-replace is that swapping models changes behavior, parameters and cost at the same time.

Parameters break. Anthropic’s deprecation page says temperature, top_p and top_k return “a 400 error when set to a non-default value on Claude 4.7 and later models,” and recommends omitting them and steering through prompting instead [4]. If you tuned temperature into a Zapier or n8n step two years ago and forgot about it, the migration is not a rename, it is a failed run.

Cost moves in ways list prices do not show. Anthropic’s pricing page says Claude 4.7 and later models “use a newer tokenizer” and that “this tokenizer produces approximately 30% more tokens for the same text,” with the exact increase depending on the content [6]. Same document, same list price, larger bill. Current Anthropic list prices are $5 per million input tokens and $25 per million output for Claude Opus 5, $2 and $10 for Claude Sonnet 5, $3 and $15 for Claude Sonnet 4.6, and $1 and $5 for Claude Haiku 4.5 [6].

Prices also move in the direction nobody plans for. Sonnet 5 launched at $2 and $10 as introductory pricing announced through 31 August 2026, and Anthropic’s page now records that this is the standard price and that “the previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur” [6]. That one went your way. The general lesson does not depend on the direction: a budget built on a rate with a date attached to it is a budget with an expiry date in it, and the date is on the vendor’s page, not in your spreadsheet.

So the honest unit of planning is not “will AI slow down.” It is hours of forced migration work per year, and you can put your own numbers on that.

Regulatory dates are forecasts, and the important ones move

The other half of the pacing question is whether rules will slow your vendors down. Here the record is instructive, because Europe ran the experiment in public.

The EU AI Act, as adopted, says it “shall apply from 2 August 2026,” with earlier dates for specific chapters [8]. Those earlier dates held: prohibited practices and AI literacy obligations entered into application on 2 February 2025, and the governance rules and obligations for general-purpose AI models became applicable on 2 August 2025 [7]. The general date did not hold. An AI Omnibus entered into force on 27 July 2026, and the Commission’s current page puts the rules for systems used in certain high-risk areas at 2 December 2027, and for systems integrated into products such as lifts or toys at 2 August 2028 [7]. Transparency rules are still listed for August 2026 [7].

The pattern is worth internalizing because it repeats. Obligations already in application tend to stay. Obligations that require standards, guidance or infrastructure that does not exist yet tend to slip, and they slip by quarters, not weeks. When you read that some rule will change how your vendors behave 18 months from now, treat it as a forecast with a wide error bar. When you read that something applied last year, treat it as fact. The same discipline applies to the pacing statement itself: it asks for tools that do not exist, to be built by governments, through an international effort [1]. That is a long-dated forecast by construction.

Four things worth keeping portable

The work that survives either outcome is unglamorous and takes an afternoon.

Keep your prompts as plain text you own, in a file or a repo, rather than only inside a Notion page, a Zapier step or a custom GPT. Prompts are the part of your setup with the longest useful life, and the part most often trapped in whichever tool you wrote them in.

Keep an eval set. Ten real inputs from your actual work, with the outputs you consider correct, saved next to the prompts. That is the entire cost of being able to answer “did the new model get worse at this” in 20 minutes instead of finding out from a client.

Keep model names in one place. If claude-sonnet-4-6 or gpt-5.6-sol is typed into nine automations, a retirement notice becomes nine jobs. If it is typed once, in a variable or a single Make or n8n scenario the others call, it becomes one.

Keep one promise ungrounded in model behavior. If what you sell to a client depends on a specific model producing a specific style, you have taken on the vendor’s release schedule as a business risk without being paid for it. Sell the outcome and the judgment. Let the model underneath be replaceable, because it will be replaced.

What still goes wrong

The deprecation pages are honest but incomplete. They tell you when a model disappears; they do not tell you when a model that stays gets quietly updated behind an alias, which is the more common way output changes under a small business. An eval set catches that. Reading the docs does not.

Notice periods are floors, not guarantees of comfort. OpenAI’s policy adds that if “safety or compliance concerns require us to retire a model sooner, we will provide as much notice as reasonably possible” [3]. Google publishes earliest-possible dates rather than a committed minimum [5]. If your business genuinely cannot absorb a two-week surprise, no amount of quarterly auditing fixes that. Redundancy across two vendors does, at the cost of maintaining two sets of prompts, which is real work and not always worth it.

And the honest limit on the pacing debate itself: nobody in it can tell you what happens next, including the people who signed the statement. They wrote that it is “hard to predict exactly how much this will accelerate AI progress” [1]. That is not evasion, it is the actual state of knowledge. Anyone selling you a confident timeline for how fast your tools will change over the next three years is working from the same public documents you now know how to read, and adding conviction that is not in them.