Stay current on AI in 30 minutes a week

AI news arrives faster than any working person can read it, and almost none of it will ever touch your work. The launches you see three times a day are mostly about models you do not run, benchmarks you do not care about, and companies whose products you cannot buy. Reading more of it does not make you current. It makes you tired and slightly anxious, which is not the same thing.

Meanwhile, the changes that can genuinely break your Tuesday are published quietly, on a schedule, with dates attached, on pages nobody shares. A model you call by name gets a retirement date. A parameter you have always passed starts returning an error. A promotional price expires. A compliance deadline arrives. All of it is announced in advance, in public, by the vendor, and almost none of it trends. This guide is a 30-minute weekly routine built on those pages instead of the feed. It is not for people whose job is the frontier itself. If you are an ML researcher, a fund analyst or a journalist covering the sector, 30 minutes is not your budget and this is not your method.

The changes that can break your week are published on a schedule

Look at what actually shipped in the last few months and ask which items would have reached you through a feed.

On 1 September 2026 Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. The launch was visible everywhere. Buried in the same release note: on those two models, tool_choice types any and tool are not supported and return a 400 error, while auto and none are unchanged [2]. If you had that in production, the interesting part of the announcement was the sentence nobody quoted. Separately, temperature, top_p and top_k are marked deprecated on Claude Opus 4.7 and later, and return a 400 error when set to a non-default value [1]. That is a change to code you wrote a year ago, delivered as a table row.

Google shut down gemini-2.0-flash, gemini-2.0-flash-001, gemini-2.0-flash-lite and gemini-2.0-flash-lite-001 on 1 June 2026 [7]. On 11 June 2026 OpenAI notified developers using older GPT-5 and o3 model snapshots that those snapshots would be removed from the API on 11 December 2026, among them gpt-5-2025-08-07, gpt-5-mini-2025-08-07, gpt-5-pro-2025-10-06 and o3-2025-04-16 [4]. Every one of those is a hard date on which something you built stops working. Every one was published months ahead on a page with no comment section.

Notice periods turn keeping up into a scheduling problem

Once you know how much warning you are entitled to, staying current stops being a reading discipline and becomes a diary discipline.

Anthropic says it notifies customers with active deployments on models with upcoming retirements, giving at least 60 days’ notice before retiring a publicly released model, and that impacted customers are notified by email and in the documentation [1]. OpenAI publishes a tiered version: at least 6 months for generally available models, at least 3 months for specialised variants such as chat, Codex and deep research models, and much shorter notice for preview models, which it says may be retired with notice such as 2 weeks [4]. OpenAI adds that it does not recommend preview models for business-critical production workloads unless you can migrate on short notice, and that if safety or compliance concerns force an earlier retirement it will give as much notice as is reasonably possible [4].

That last tier is the one to internalise. If you have wired a preview model into something a client depends on, your buffer may be as short as the 2 weeks OpenAI gives as its own example [4]. If you are on a generally available OpenAI model you have half a year, which is enough time to do nothing for five months and still be fine [4]. The two situations need completely different amounts of your attention, and no amount of scrolling tells you which one you are in. The deprecation page does, in about 90 seconds.

So the first thing you do, once, is read the deprecation table for every model name that appears in your own code, your automations or your tool settings, and put the dates in your calendar. Anthropic currently lists claude-sonnet-4-5-20250929 with a tentative retirement date of not sooner than 29 September 2026, and claude-haiku-4-5-20251001 not sooner than 15 October 2026 [1]. “Not sooner than” is a floor, not a promise, but a floor is enough to plan against.

Three feeds, chosen for the tools you already pay for

Pick sources the way you pick suppliers: few, deliberate, and tied to money you actually spend. One primary channel per vendor you pay is the whole portfolio.

Anthropic makes this easy. The platform release notes have an RSS feed at https://platform.claude.com/docs/en/release-notes/feed.xml, covering model releases and deprecations, API features and breaking changes, SDKs, console changes, and pricing and rate limit changes in one place [3]. Put that in whatever reads feeds for you and the vendor pushes to you instead of the other way round. Anthropic also points to separate release notes for the consumer apps in its help centre and a changelog for Claude Code, if those are what you use [3].

OpenAI and Google are pages you visit rather than feeds you subscribe to. OpenAI’s API changelog carries dated entries, and the deprecations page is maintained separately [5][4]. Google’s Gemini API changelog does the same job, listing new releases alongside the shutdown dates for older models [7]. Neither page advertises a feed [5][7], so they become two bookmarks you open during your slot. Two clicks is not a hardship. Believing you will remember to check them without a calendar entry is the actual failure mode.

Audit the set twice a year. A source that has not changed one of your decisions in six months is entertainment. Entertainment is fine, and it comes out of a different budget than these 30 minutes.

One question decides whether an item survives

For every item that reaches you, ask whether it changes what you can do, what it costs, or what you are obliged to do. Three buckets. Anything that fits none of them gets closed.

What you can do is the bucket most people over-weight, because it is the fun one, but it does contain real items. Gemini 3.8 Flash became generally available on 2 September 2026 [7]. On 3 September 2026 OpenAI released GPT-6 Astra, which it describes as its most capable model, built for the hardest end-to-end work [5]. Claude Fable 5.1 supports a 1 million token context window by default and 128k maximum output tokens [2]. Those matter only if you are currently blocked by the thing they fix. If you are not blocked, note it and move on.

What it costs is the bucket most people ignore, and it is where the money is. On 10 August 2026 Anthropic made Claude Sonnet 5’s introductory pricing of $2 per million input tokens and $10 per million output tokens the standard price, and confirmed that a previously scheduled 1 September increase to $3 and $15 would not occur [2]. That is a bill that did not go up, and you would only know from the release note. In the other direction, OpenAI’s pricing page states that GPT-5.6 Sol’s promotional pricing is available at least through 21 November 2026 [6]. That is a date on which your costs may change, sitting in plain sight on a pricing page, and it belongs in your calendar rather than your memory. Fable 5.1 also prices prompt cache reads at $0.25 per million tokens, 0.025 times the base input price against 0.1 times on other models [2], which is the kind of detail that quietly rewrites whether caching is worth the engineering.

What you are obliged to do is the bucket that arrives without warning if you are not looking. Claude Fable 5.1 and Mythos 5.1 both require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorises it [2]; if you sold a client on zero retention, that is a model you cannot assume you can use, and no headline will tell you. On the regulatory side, the remainder of the EU AI Act starts to apply from 2 August 2026, and providers of systems generating synthetic audio, image, video or text have until 2 December 2026 to comply with the Article 50(2) transparency rules [8]. Those are fixed calendar items, published years in advance, and they do not care whether you read the news.

Everything else fails the test. Funding rounds, executive moves, demos of products with a waitlist, benchmark arguments about models you do not run, and predictions about 2030. Interesting is not the bar. Actionable is the bar.

The thirty-minute shape

Same slot every week, on the calendar, treated like a meeting with someone you respect.

The first 10 minutes are scanning. Open your feed reader and your two bookmarks, read headlines and dates only, and mark the handful of items that survive the three-bucket question. Most weeks that is two or three items, and some weeks it is zero. Zero is a result, not a failed session.

The next 15 minutes are reading properly, and reading the primary source rather than the coverage of it. A release note is usually shorter than the thread about the release note, and it is the version that contains the 400 error you needed to know about. If an item involves a date, put the date in the calendar now, while you are looking at it, because a date in a notes file is a date you will meet by surprise.

The last 5 minutes are writing one sentence per surviving item: what changed, and whether you will do anything. Most weeks the honest answer is “nothing yet”. Keep the log anyway. It is the only evidence you will have, six months from now, that you knew about something before it hit you, and it is what makes the twice-yearly source audit possible.

One experiment a month keeps the reading honest

Reading alone leaves your knowledge theoretical, and theoretical knowledge of a tool is close to worthless when you have to decide whether to move a client onto it. Budget one hour a month to actually use one new thing on a real task from your own week, with a fair comparison against how you do it now and a decision at the end to keep it or drop it.

One honest hour teaches you more than twenty reviews, because it tests the tool against your work rather than against a demo script. It also inoculates you against the two failure modes at either end: assuming everything new is hype, and assuming everything new is urgent. Declining to try the other twenty-nine tools that month is not falling behind. It is the discipline that keeps the 30 minutes at 30.

What still goes wrong

Thirty minutes a week will not make you an expert, and it is not meant to. It makes you reliably informed about the tools you pay for and reliably ignorant about everything else, which is the correct trade for someone with a business to run. You will still miss things. The bet is that what you miss is cheaper than what the reading costs.

The dates are softer than they look. Anthropic’s table heads its column “Tentative retirement date” and the cells say “not sooner than”, which is a floor and not a commitment [1]. OpenAI is explicit that safety or compliance concerns can pull a retirement forward, in which case it promises only as much notice as is reasonably possible [4]. Notice periods also assume the vendor can reach you: Anthropic notifies customers with active deployments by email and in the documentation [1], which does nothing for a model you use through a third-party tool that never tells you which model it calls. If a product will not name its model, you have no deprecation visibility at all, and that is a reason to ask before you depend on it.

The method is also structurally conservative. It filters on the tools you already use, so it will make you late to a genuinely new category rather than a new version. That is what the monthly experiment is for, and it is an imperfect patch. Finally, the consumer apps move faster and document less than the APIs do; separate app release notes exist in Anthropic’s help centre [3], but the platform notes are the detailed ones, and the gap between “the interface changed this morning” and “there is a note about it” is real. For anything you rely on daily, the interface is the changelog, and you find out by using it.