What to Do When an AI Tool Raises Its Price or Kills a Model

Do nothing for an hour, then run the drill: confirm what actually changed, check whether it breaks a running workflow or only a budget line, and migrate only what breaks. Most price and model announcements need one config change and a note in your calendar. A few need work that same week. The skill is telling them apart, and the difference is always the same question — does something stop running, or does something get more expensive?

Both happen constantly now. In the last ninety days, one vendor cut prices 80%, another announced an increase without publishing a number, and two flagship models were removed from an API outright. None of that is unusual anymore. It’s the operating condition.

Two kinds of change, and they are not the same emergency

A break stops work. The model ID your automation calls returns an error. The app you exported to is gone. There is no version of tomorrow where the workflow runs unchanged.

A repricing costs money. The workflow still runs. Your margin moved.

Breaks have deadlines set by someone else. Repricings have deadlines you set. Treat them the same and you’ll either panic over a $40 line item or get caught flat when an endpoint dies at midnight. Here is what each has looked like in practice this year:

Date	What happened	Kind
May 12, 2026	OpenAI removed dall-e-2 and dall-e-3 from the API; GPT Image 2 is the replacement	Break
Jul 8 & Jul 30, 2026	Runway retired Gen-3 Alpha, then Gen-3 Alpha Turbo and Gen-4 Aleph — calls naming those IDs now fail	Break
Jul 30, 2026	OpenAI cut GPT-5.6 Luna 80% to $0.20/$1.20 per M and Terra 20% to $2/$12	Repricing (down)
Aug 31, 2026	Claude Sonnet 5 intro API pricing ends — $2/M input becomes $3/M in, $15/M out	Repricing (up, dated)
Sep 24, 2026	OpenAI’s Sora API is discontinued; the web and app experiences already ended April 26	Break (scheduled)
Announced Aug 6, 2026	DeepSeek says API pricing is going up, calls the increase significant, publishes no rate and no date	Repricing (up, undated)

Look at the right-hand column. Four of those six were on a published calendar before they landed. The vendors told anyone who was reading.

The drill

1. Confirm it at the source. Roundups lag and aggregators guess. Read the vendor’s own changelog, pricing page, or API docs. Today’s DeepSeek story is a clean example: the reporting is real, and the official docs say a significant increase is planned — but they publish no rate and no effective date, so anyone quoting a specific new price is inventing it. Write down what the vendor actually said, not what a blog inferred.

2. Classify it. Break or repricing. One question: does something stop running?

3. For a break, find the deadline and the successor. Vendors almost always name the replacement in the same notice — Runway pointed Gen-3 Alpha Turbo users to Gen-4.5 or Gen-4 Turbo, and Gen-4 Aleph users to Aleph 2.0. Put the deadline in your calendar with a two-week buffer, then find every place the dead identifier appears. That’s a grep, not a project.

4. For a repricing, do the math before you have the feeling. Multiply the new rate by last month’s actual volume. Most operators discover the increase is real and small. If it’s real and large, you now have a number to decide with instead of a mood.

5. Migrate what breaks. Schedule what costs. Breaks get done this week. Repricings get a calendar entry at the price-change date and a decision then, with a full month of data.

The part that matters more than the drill

Every one of those six events is survivable in minutes if your stack was built for it, and a bad week if it wasn’t. Three habits do almost all the work:

• Name models in config, never in code. If swapping gen4_aleph for aleph2 means editing one line in one place, a retirement is a chore. If the string is scattered across nine automations, it’s an outage. This is the single highest-leverage thing on the list.
• Know which seats are commodity and which are load-bearing. A cheap API doing bulk classification is a commodity — it should be swappable by design, because that’s exactly the seat where price moves hardest. Your thinking seat and your brand-voice seat are load-bearing; you’d absorb a price increase there rather than rebuild. Decide which is which before an announcement forces it.
• Never let one vendor’s price be the reason a workflow exists. If a workflow is only viable at $0.14 per million tokens, it isn’t a workflow — it’s a bet on someone else’s pricing page. Build it so it still makes sense at triple.

That third one is the whole lesson of this year. DeepSeek’s price floor was genuinely the cheapest credible option in the market, and it still is today. It’s also the floor the vendor has now publicly said it intends to raise. Both things are true. The operators who’ll shrug at that announcement are the ones who treated cheap tokens as a nice margin instead of a foundation.

What to do this week

1. Grep your automations for hardcoded model identifiers. Move every one into config.
2. List your top five AI line items and mark each commodity or load-bearing. If everything is load-bearing, you haven’t decided — you’ve just accumulated.
3. Open the changelog page for the three tools your work actually depends on. Subscribe if they offer it. Fifteen minutes now beats a broken pipeline later.
4. Put the dated changes above into your calendar: Aug 31 for Sonnet 5’s intro pricing, Sep 24 for the Sora API.

The vendors will keep changing the terms. That isn’t a betrayal, it’s a market repricing itself in public. Your job isn’t to predict it — it’s to build a stack where the announcement is a config change, not a crisis.

A clearer stack beats a bigger stack — and a swappable one beats both.

Next: run the stack audit to find out which of your tools have no job at all, or check what you should be spending before the next price change makes the decision for you.