How to read Microsoft's AI pivot without rebuilding your stack

The announcement lands on a Monday. A large vendor reorganises, cuts a division, folds three teams into one, renames its AI platform, and stands up a new group to sell deployment services. By lunchtime there are four explainers telling you what it means for the industry. You read two of them and arrive back at the only question that concerns you, which is whether anything on your desk has to change this week. Most of the time the honest answer is no. Occasionally it is yes, and the part that mattered was never in the coverage. It was a date on a documentation page that nobody linked to.

This guide is a method for telling those two cases apart. It uses Microsoft as the worked example, because Microsoft publishes its lifecycle commitments as policy pages anyone can open [1][3], and because a company that size reorganises often enough to give you plenty of practice. It is written for a solo operator or small team running real work through one or two AI products, with no appetite for following vendor strategy as a hobby. It is not for anyone whose stack is fixed by a procurement or compliance process, and it will not help you decide what to think about the vendor as an investment. Prices and policies below were fetched on 4 September 2026.

A pivot reaches you as a retirement or an invoice

Reorganisations are mostly internal. Reporting lines move, budgets shift, people leave. None of that crosses the boundary into your working week by itself. There are only two channels through which a vendor’s strategy actually arrives at your desk: something you use stops existing, or something you use changes price or packaging.

Hold the July 2026 Microsoft cuts against that test. The company announced it would eliminate 4,800 jobs, about 2.1% of its workforce, of which 1,600 were in the Xbox division, and four Xbox game studios were transferred to independent or new ownership [8]. It was the third round in about fourteen months, after more than 6,000 in May 2025 and approximately 9,000 that July [8]. If you run a two-person consultancy that drafts proposals in Word and calls a model through an API, the correct operational response to all of that was nothing at all. No product you depend on stopped existing. No price you pay moved.

Meanwhile, in the same period, two things did reach desks, and neither made a headline. Models with retirement dates already on the calendar moved closer to those dates. And the Microsoft 365 Copilot Business add-on was listed at $18.00 per user per month on annual billing, reduced from $21.00 [7]. One of those is a migration you have to schedule. The other is money. Both are findable on a Tuesday afternoon if you know which pages to open.

So the working rule is unglamorous. Read the headline for context and nothing else. Read the documentation for dates. The strategy commentary tells you where a vendor is putting its attention over the next three years, which is interesting and rarely actionable. The lifecycle pages tell you what is being taken away from you and when, which is boring and always actionable.

The notice you are owed is published, and it varies by more than an order of magnitude

The vendors here all state, in writing, the minimum warning they will give before removing something. These numbers are wildly different from each other, and almost nobody reads them until the email arrives.

OpenAI commits to at least 6 months of advance notice before retiring a generally available model, at least 3 months for specialised variants such as chat, Codex and deep research versions, and says preview models may be retired with much shorter notice, such as 2 weeks [4]. Anthropic states that it notifies customers with active deployments and provides at least 60 days’ notice before model retirement for publicly released models [5]. Microsoft’s model platform commits to at least 60 days before retirement for generally available models and at least 30 days for preview models [1]. Google’s practice, as recorded in its own changelog, has run tighter still in places: the shutdown of gemini-robotics-er-1.6-preview on 31 August 2026 was posted on 30 July [6].

Set those against Microsoft’s Modern Lifecycle Policy, which governs its products more broadly and promises a minimum of 12 months’ notification before ending support where no successor product or service is offered, excluding free services and preview releases, with notification up to three years ahead for select Azure services [3]. That is the spread you are working with. Twelve months at one end, two weeks at the other, for things that feel to you like the same category of dependency.

The sharpest version of this is worth sitting with. The same model family can carry different notice depending on which door you came in through. OpenAI’s own commitment for a generally available model is at least 6 months [4]. OpenAI models delivered through Microsoft’s platform sit under a 60-day commitment [1]. Nothing improper is happening. It is simply that your runway is a property of your contract, not of the model, and you cannot infer one from the other.

Retirement dates exist before anyone announces them

This is the part that changes how you work, and it is one sentence on a Microsoft documentation page. Generally available models have their retirement date set programmatically at launch to 18 months out, and there is no separate announcement [1]. The same page states plainly that retirement dates are not extendable [1].

Read that again in terms of your own calendar. The date you will eventually be emailed about already exists. It was created the day the model shipped, and it is exposed through the Models API as well as the published schedule [1]. You do not have to wait to be told, and waiting to be told costs you sixteen of the eighteen months. The retirement schedule is a public page you can open this afternoon and read like a calendar.

What is on it right now is instructive. The gpt-4o version dated 2024-05-13 retires on 1 October 2026 with gpt-5.1 named as its replacement [2]. o1 retires on 21 October 2026 with gpt-5.6-sol listed as its replacement, and o3-mini on 1 October 2026 with o4-mini [2]. That replacement column identifies Microsoft’s recommended successor and, where supported, the target for automatic upgrades at retirement [2]. And gpt-5.1, the successor being pointed at, already carries its own retirement date of 15 May 2027 [2]. Nothing about that is a surprise or a punishment. It is the 18-month clock running exactly as documented.

Two details decide whether the migration is silent or manual. For Global Standard, Data Zone Standard and Standard deployment types, Microsoft manages automatic upgrades when a model version is retired [1]. Provisioned deployments are not auto-upgraded, and provisioned customers must manually migrate to the replacement model [1]. If you are on a standard deployment, the risk is not an outage, it is that your outputs change one morning and you had no reason to look. If you are provisioned, the risk is an outage and it is on the calendar.

The last detail is that vendors keep separate calendars for the same model. Microsoft’s schedule lists claude-sonnet-4-5 and claude-haiku-4-5 with a retirement date of 19 October 2026 and no replacement named [2], while Anthropic’s own deprecation page lists claude-haiku-4-5-20251001 as active with a tentative retirement no sooner than 15 October 2026 [5]. Same model, two pages, two dates to track. Whichever page governs your bill is the one that governs your Tuesday.

The useful question is not whether the notice period is generous. It is whether it is longer than the time you personally need to move. That is arithmetic you can do once.

If the answer comes out lower than the number of workflows you actually run, you have found the real problem, and it is not the vendor. It is that you have more dependencies than your notice period can absorb, which is a thing you can fix on a quiet week rather than a forced one.

Repackaging is the half that people miss

Retirement is the loud channel. Pricing and packaging is the quiet one, and it moves more money.

The current Microsoft 365 Copilot business page shows the shape of it. The Copilot Business add-on is $18.00 per user per month on an annual subscription, reduced from $21.00 [7]. Microsoft 365 Business Standard with Copilot is $23.50 per user per month and Business Premium with Copilot is $32.00, both paid yearly [7]. The versions of those plans sold without Teams are $20.30 and $28.80 [7]. Monthly billing costs more than annual across the range, with Business Premium listed at $38.40 per user per month [7].

The durable pattern there has nothing to do with those specific numbers, which will move again. It is that the AI capability stopped being a thing you decided to buy and became an ingredient in a plan you already had. When that happens, your invoice stops containing a line item that says AI, and your ability to answer “what am I spending on this” quietly degrades. A price going down is not automatically good news either, because a cheaper add-on that is now bundled into a more expensive base plan is a price increase wearing a hat.

The same repackaging happens at the API surface, where it is easier to miss because it does not appear on a bill at all. Google’s changelog entry for 21 July 2026 records that the sampling parameters temperature, top_p and top_k are now deprecated [6]. No model was retired in that entry and no price moved. What changed is the interface your code uses to steer output: settings you tuned once are now on a removal path, and a release note is the only place that says so. That class of change never makes the strategy coverage, and it is the one most likely to change your output without changing your invoice.

One page listing what you would have to rebuild

All of the above only becomes useful if you can answer, in under a minute, what you actually depend on. Almost nobody can, and the fix takes an hour.

Write one line per recurring workflow. Name the vendor, the exact model or product string rather than the family, where the call is made from, what breaks if it stops, and how many hours it would take to move. The version string matters more than it looks. “We use Claude” and “we use Gemini” are not dependencies you can check against a retirement schedule. claude-haiku-4-5-20251001 is, and it resolves to a specific date on a specific page [5].

Next to each line, write the published retirement date, looked up rather than guessed. For anything on Microsoft’s platform, that date already exists whether or not you have been told about it [1]. For OpenAI and Anthropic, the deprecation pages carry both the announcement date and the shutdown date, so you can see how much warning was actually given in practice rather than in policy. OpenAI’s 26 August 2026 notice on whisper-1 and its gpt-4o transcription models set shutdown for 26 February 2027 [4], and Anthropic deprecated claude-opus-4-1-20250805 on 5 June 2026 for retirement on 5 August 2026 [5]. Both met their stated minimums. That track record is worth knowing before you need it.

Then keep the prompts and reference material that drive each workflow in files you own, outside any single product’s saved projects. A prompt you cannot paste into a second tool is a prompt you cannot migrate under a deadline. And keep a small set of real inputs with the outputs you were happy with, so that when a model is swapped underneath you, silently or otherwise, you have something to compare against other than your memory of how it used to feel.

Review the page once a quarter. Fifteen minutes, four times a year, is enough to catch every date on this list well before it catches you.

What still goes wrong

The notice is only as good as the address it goes to. Microsoft says retirement notifications are sent automatically to subscription owners with active deployments [1], and Anthropic says it notifies customers with active deployments [5]. If the account is in a founder’s name, or the billing email routes to an inbox nobody has opened since setup, the policy was honoured and you still found out late. That failure is cheap to prevent and easy to overlook.

Automatic upgrades are a mercy and a trap at the same time. A managed upgrade on a standard deployment [1] means you will not get an outage, which also means you will not get a signal. Your summaries, classifications and drafts start coming from a different model, the change is usually an improvement, and occasionally it is not, in a way that is specific to your material and invisible in aggregate. Nothing in a vendor’s policy protects you there. Only a saved set of your own inputs and outputs does.

And published minimums are minimums, not promises about the whole category. Preview models can go with 30 days on one platform [1] and as little as two weeks on another [4], which means anything built on a preview is a prototype no matter how well it works. Beyond that, the deeper limit is that none of this tells you what a vendor will do next. The 2025 and 2026 restructurings [8] were genuinely large and, for almost every small operator, operationally irrelevant. The retirement schedule was small, boring, and the thing that actually moved work. Reading the pivot well mostly means resisting the pull of the interesting document in favour of the useful one.