AI 101Beginner19 min read

The 2026 AI year in review: what really changed, and what is still hype

The main stories of 2026 with dates and sources (faster model releases, agents becoming products, open models and the US-China gap, money and jobs, EU rules and courts), each sorted into solid, claimed and hype, plus the limits that did not change and a method for reading AI news.

The 2026 AI year in review: what really changed, and what is still hype

The 2026 AI Year in Review: What Really Changed, and What Is Still Hype

Every few days this year, a headline has told you that AI has changed everything. Some of those headlines were right, many were not, and most did not tell you which was which. This tutorial is a plain-words review of 2026 so far, written on 14 November 2026. It covers the things that actually happened, with dates and sources: faster model releases, agents moving from demo to product, open models catching up, record money and uneven results, and new rules and court decisions. For each, it separates what is solid, what is only claimed by the company that made it, and what is still hype. You will also learn a simple method for reading any AI news the same way, so the next headline does not catch you out.

Difficulty: ★★☆☆☆ (no technical background needed)
Tools needed: None, only a browser if you want to open the sources
Last updated: November 2026, with facts checked on 14 November 2026

Overview

A year in review is a dangerous thing to write about AI, because the field moves faster than any article. By the time you read this, something will have changed. So this tutorial does two jobs. It tells you what the main stories of 2026 were, as of the date above. And it shows you how to weigh each one, so you can do the same for next month's news without needing anyone to explain it.

Here is the method, in three labels that you will see throughout.

  • Solid. Something you can verify in a primary source: an official page, a published report, a regulator's text or a court record. Dates, names and decisions usually fall here.
  • Claimed. Something the company or lab that built the product says about it, usually with benchmark scores. It may well be true, but it is the seller describing the product, and a benchmark is not the same as your own work. Treat it as a lead, not a verdict.
  • Hype. Something with no clear source, a vague superlative ("the smartest ever", "changes everything") or a forecast presented as fact.
  • You will also meet a fourth label, reported, for things that appear in news coverage but that I could not confirm against the primary source. Where that happens, I say so.

    The tutorial has seven steps, one for each theme, then a short guide to using all this. The themes are: how to read a year-in-review, the faster model race, agents becoming products, open models and the US–China gap, money, adoption and jobs, rules and courts, and finally the limits that did not go away. At the end you will write a short "what this means for me" note, which is the part that actually changes what you do.

    One honest warning before we start. Everything here is accurate to the best of my checking on the date above. Some of it will be out of date within weeks, and the table of dates at the end shows what to re-check first. If a statement matters to a decision, open the source and read it yourself.

    Who This Is Useful For

  • People who want a clear picture of 2026 in AI without wading through hype.
  • Professionals who need to explain to colleagues or clients what has really changed.
  • Anyone who wants a method for reading AI news critically.
  • What You Will Learn

  • The main developments of 2026 in AI, with dates and sources.
  • Which parts are solid, which are only claimed and which are still hype.
  • How to read a benchmark claim, a funding headline or a regulation story.
  • What did not change: the limits and risks that remain.
  • A simple method for turning news into a decision about your own work.
  • What You Need

    1. About thirty minutes.
    2. A notes app or paper, for the closing note.
    3. A browser, if you want to open the sources and read them yourself, which I recommend for at least two.

    Stanford AI Index 2026

    The ninth annual report from Stanford's Institute for Human-Centered AI, published in April 2026. A good primary source for numbers on capability, investment, adoption and public opinion.

    hai.stanford.edu
    EU AI Act (European Commission)

    The official page for the EU's AI rules, including the 2026 Digital Omnibus changes and the dates that apply.

    digital-strategy.ec.europa.eu

    Facts checked on 14 November 2026. Product facts are as stated by the companies on their own pages, and benchmark results are the companies' own unless I say otherwise. I give no prices, because they change. Where sources disagreed or I could only find secondary coverage, I say "reported" and give you the primary source to check.

    Step 1 — How to read a year in review

    Before the story, the method. Four habits will serve you here and for every AI headline.

    1. Find who is speaking. A lab announcing its own model is the seller. A regulator publishing a rule is a primary source. A newsletter summarising either is secondhand. All three can be useful, but they deserve different amounts of trust.

    2. Separate the date from the claim. "Launched on 9 July" is a fact you can check. "The best model ever" is a claim. Write them down separately.

    3. Treat benchmarks as a guide, not a grade. A benchmark is a test set. A model can score very high on a test and still fail at your task, and a score can be tuned for. A benchmark result is useful when it comes from an independent source, when the test is named and when it measures something close to your work.

    4. Ask what changed for a person like you. A new model can be impressive in a lab and make no difference to how you draft an email. Ask: what could I do this year that I could not do last year?


    Step 2 — A faster race between models

    What happened (solid, from company pages). The big labs released new models at a pace that made "the latest model" hard to pin down.

  • Anthropic announced Claude Fable 5 on 9 June 2026. Its rollout was interrupted when, on 12 June, the US government applied export controls to the model and Anthropic suspended access. Access was restored from 1 July after the controls were lifted. Anthropic then released Claude Fable 5.1 on 1 September, Claude Opus 5.5 on 22 September, Claude Sonnet 5.5 on 28 September, and, according to its models page, Claude Fable 5.2 and Mythos 5.2 on 9 November.
  • OpenAI released the GPT-5.6 family, in three tiers named Sol, Terra and Luna, with general availability on 9 July 2026 after a limited preview that began on 26 June. OpenAI later changed which model free users get by default, and adjusted prices in July and August. On 3 September it announced GPT-6 Astra, which coverage describes as going first to a limited group of customers, including members of its cybersecurity-focused Daybreak programme, before wider availability. A restricted GPT-6 Cyber followed on 5 November (reported by news outlets; check OpenAI's own pages for what is available to you).
  • Google announced Gemini 3.5 at its I/O event on 19 and 20 May 2026, with a faster Flash model first and a Pro model to follow.
  • What is claimed. Each company reports that its new model leads on coding, reasoning or knowledge-work benchmarks, and each names a rival it says it beats. These are the companies' own tests. They are not wrong to publish them, but you cannot compare them across companies, because the tests, settings and dates differ. The same applies to big framing claims: OpenAI called GPT-6 Astra its most intelligent and best-aligned model, and its co-founder Greg Brockman said the "AGI era" had arrived. Those are the company's own words, and the safety claims came without independent verification at the time of the announcement.

    What is hype. Any statement that a single model is "the best" without saying at what, measured by whom and when. Over a year with this many releases, the "best" label changed hands several times, and the truthful answer for most readers is that several models are very capable and the differences matter more for specialised work than for everyday tasks.

    What it means for you. Three practical points.

    1. Names and versions change quickly. A tutorial that names one model is a snapshot. Learn the kinds of model and tool, and check the current names when you need them.
    2. Access is not guaranteed. The Fable 5 suspension is a reminder that a model can be paused for reasons outside your control, including government action. Do not build a critical process on a single provider without a fallback.
    3. Most everyday tasks do not need the newest model. Drafting, summarising and brainstorming work well on mid-range models. The latest release matters most for hard, long or specialised work.

    Step 3 — Agents move from demo to product

    What happened (solid, from company pages). The word of the year was "agent": AI that does multi-step work for you instead of answering one question at a time.

  • Google I/O 2026 was framed around agents. Google described Gemini Spark, a personal agent in the Gemini app that takes action on your behalf, starting as a beta for subscribers of its top plan in the US. It described a Daily Brief agent that pulls together items from your email and calendar, information agents in Search that monitor topics in the background, an Antigravity platform for building agents, and Managed Agents in its API, in preview.
  • OpenAI described a GPT-5.6 setting called Ultra that coordinates several agents on parallel workstreams, and its GPT-5.6 models are offered in ChatGPT Work and in Codex.
  • Grok Bot is presented in its official 1.0 notes as an AI teammate with its own computer, which can be shared with teams and can join video calls.
  • What is claimed. Demos show agents booking, researching and filing things end to end. Company descriptions focus on what is possible. Independent evidence on how reliably agents complete real, messy tasks without supervision is thinner, and where it exists it is more mixed than the demos suggest.

    What is hype. "Your AI employee" and "set it and forget it". An agent that acts in your accounts can be wrong at a cost, and the people who use them well keep review steps, small permissions and checks.

    What it means for you. Agents are real, they are arriving inside tools you already use, and they are worth learning, carefully. The habits from the agents tutorials apply: start read-only, keep "ask me first" on for anything that sends, spends or deletes, and check the work.

    Step 4 — Open models and the shrinking gap with China

    What happened. Two linked stories.

  • Open-weight models got close to the frontier (reported, with named analysts). In July 2026, Moonshot AI in China announced Kimi K3, a very large open-weight model, with weights published later that month. Coverage by outlets such as Axios and the South China Morning Post, and analyst commentary, described it as performing close to leading closed models on many benchmarks. Estimates of how far open models trail the closed frontier varied: one analyst put the lag at three to five months, and another source citing a government institute put it at four to seven months, both narrowing from about six to ten. These are estimates from named commentators, not a settled measurement.
  • The US–China gap narrowed (solid, from the Stanford AI Index 2026). The report says the top US model led China's best by about 2.7 per cent in March 2026. It also notes that the US builds more notable models, while China leads in publications and citations.
  • What is claimed. Benchmark scores for open models, which come from the model makers and from third-party leaderboards. Leaderboards can be gamed and measure different things.

    What is hype. "China has won" or "open models make the big labs irrelevant". Neither is established. A gap of a few months in capability is not a gap of zero, and cost, reliability, support and safety differ.

    What it means for you. More choice, and more pressure on prices. Open-weight models also make it practical to run capable AI on your own hardware for privacy, which the local-AI tutorial in this track covers. At the same time, before you use any model from any country, check where your data goes and what the terms say. That applies to every provider, not only to one.

    Step 5 — Money, adoption and jobs

    The best primary source here is the Stanford AI Index 2026, published on 13 April 2026. These figures are from Stanford's own summary, so confirm any figure you plan to repeat on its page.

    Solid, from the report:

  • Investment hit a record. Total AI investment in 2025 is reported at about $582 billion, a record, and the report says the United States committed roughly 23 times more private AI investment than China.
  • Adoption is widespread. The report puts organisational adoption of AI at 88 per cent in 2025, and estimates the value consumers get from generative AI in the US at about $172 billion a year by early 2026.
  • Jobs are under pressure in specific places. Employment of software developers aged 22 to 25 fell by nearly 20 per cent since 2024, and about a third of organisations expect AI-driven workforce reductions.
  • Experts and the public disagree. On whether AI will have a positive effect on jobs, the report found 73 per cent of experts said yes against 23 per cent of the public.
  • Claimed. Forecasts of how many jobs will vanish or how much money AI will add to the economy. They depend on assumptions, and the range is wide.

    Hype. "Half of all jobs gone by next year", and the opposite claim that nothing will change. The honest picture in the data is uneven: strong effects in some roles, especially early-career ones in certain fields, and slower change elsewhere.

    What it means for you. Do not treat adoption statistics as proof of benefit to a particular business. If you are early in your career, pay attention to the tasks AI does well in your field and build the skills it does not replace: judgement, checking, communication and domain knowledge. If you hire or manage, the report's point about early-career roles is worth discussing before you decide how to use AI in your team.

    Step 6 — Rules and the courts

    EU rules (solid, from official pages).

  • The EU's rules for general-purpose AI models have applied since 2 August 2025.
  • In 2026 the EU agreed a "Digital Omnibus" to simplify and adjust the AI Act. A provisional agreement was reached on 6 and 7 May 2026, the Council gave its final approval on 29 June 2026, and the Commission's page says the changes entered into force on 27 July 2026.
  • The main change: the dates for most high-risk obligations were pushed back. For stand-alone high-risk systems such as those used in hiring, education and credit, the date moved to 2 December 2027. For AI built into regulated products such as medical devices, it moved to 2 August 2028.
  • The deal also adds a ban on AI systems that create non-consensual intimate imagery or child sexual abuse material, with a deadline of 2 December 2026 for systems to comply.
  • Law-firm summaries report that the transparency duties, such as labelling certain AI-generated content, kept to their original August 2026 timeline, with the machine-readable marking element reported as later. Check the official text for the exact dates that apply to you.
  • Copyright (reported). In the United States, the case Bartz v. Anthropic, brought by authors over books used to train AI, ended in a settlement of about $1.5 billion, which several legal-news outlets report received final court approval in July 2026. Coverage says an earlier 2025 ruling in that case found that training on lawfully obtained books was fair use, while acquiring and keeping pirated copies was not, and that this ruling is one district court's decision and not binding precedent elsewhere. Other cases, including ones about images, were still pending in the sources I found. I could only confirm these through secondary coverage, so check court records or a major legal-news outlet before relying on any detail.

    Claimed. Anything that says "AI is now legal" or "AI is now illegal". The law is a patchwork of rulings, rules and dates, and it differs by country.

    What it means for you. Rules are arriving, but slower than the headlines suggest, and they differ by place. If you use AI for work, check what applies where you and your customers are, and keep it simple: do not feed it material you have no right to use, label AI-generated content where required and keep records. The next tutorial in this module covers privacy, copyright and safe use in practical terms.

    Step 7 — What did not change, and what to do with all this

    The limits that remain.

  • Uneven ability. Stanford's report describes models that can win a gold medal at an international maths olympiad yet still fail to read an analogue clock reliably, and robots that succeed at only a small share of real household tasks. Strong in some ways, brittle in others, and hard to predict from outside.
  • Mistakes still happen. The hallucination tutorial applies as much as ever. Newer models are better on average, and a confident wrong answer is still a confident wrong answer.
  • Safety and transparency are lagging. The report counts documented AI incidents rising to 362 from 233 the year before, and says an index of how transparent model makers are about their models fell on average. In the US, only 31 per cent of people surveyed trusted their government to regulate AI responsibly, the lowest of the countries compared.
  • Costs and access can change overnight. Prices moved, free tiers changed and one major model was suspended for weeks.
  • What to do with all this. A review is only useful if it changes something. Write a short note answering five questions.

    1. One thing that changed this year that I can use now: ___
    2. One claim I was ready to believe that I will now check: ___
    3. One habit I will keep or add because of the limits above (for example: check three claims, keep "ask me first" on): ___
    4. One rule or law that could affect my work, and where I will check the official text: ___
    5. One date I will come back to review this on (for example, three months from now): ___

    Dates to Re-check First

    If you only have time to refresh a few things, start with these, which are the likeliest to have changed.

    TopicWhat to checkWhere
    Model names and versionsWhich models each lab lists as currentThe lab's own models page
    Prices and free tiersWhat your plan includes todayThe tool's official plan page
    Agent productsWhich are generally available and in your regionThe product's official page
    EU AI Act datesThe dates that apply to your useThe European Commission's AI Act page
    Copyright casesThe status of cases that matter to youCourt records, a major legal-news outlet
    AI Index figuresAny number you plan to quoteThe Stanford AI Index page

    Worked Example: using the three labels on a headline

    A colleague forwards a post: "NovaTech's new AI is 100% accurate and will replace half of all analysts by Christmas. It beats every human expert, and users say it changed their lives."

    You apply the labels.

  • Who is speaking? A company post, so the seller.
  • Solid? The post names a product and a release date. Check those on the company's own page.
  • Claimed? "Beats every human expert." Where is the test, who ran it and what exactly was measured?
  • Hype? "100% accurate" with no source is a red flag. No system is, and the AI Index notes that even strong models fail at simple tasks. "Replace half of all analysts by Christmas" is a forecast presented as fact. "Users say it changed their lives" is an anecdote.
  • You reply to your colleague with three questions: "Where is the test, who ran it, and what was it measuring? Which analyst tasks, and on what evidence? And what happens to the errors?" Then you spend ten minutes testing the product on a real task from your own work, in a way that you can check. The result is a decision based on your own evidence, not on the post.

    Common Mistakes to Avoid

    Three habits cause most of the confusion about AI news.

    Mistake #1: Treating a lab's benchmark as a neutral ranking. A company's own score for its own model is a claim, and scores across companies use different tests and settings. Fix: look for independent results, ask what was measured and how close it is to your task, and test the model on your own work.

    Mistake #2: Believing headline numbers with no source. Percentages, jobs forecasts and funding figures spread quickly with the original context stripped away. Fix: find the primary source, such as the report, the regulator's text or the court record, read the sentence around the number and note the date.

    Mistake #3: Chasing every release. Switching tools every week to use "the best model" costs time and attention and rarely changes your results. Fix: pick a small set of tools, review them each quarter, and try new releases on harmless jobs before relying on them. Keep a fallback for anything important.

    Going Further

    Make a personal timeline. In your notes, keep a dated list of the AI events that affected your work this year. In twelve months you will have your own review, which is better than any article.

    Read one primary source in full. Choose a section of the AI Index, or the EU's AI Act summary page, and read it with a pencil. You will find that the primary text is often more cautious than the headlines.

    Compare two summaries of the same event. Read a company post and an independent report of the same release side by side and note what each one emphasises and leaves out.

    What to learn next. The next tutorial covers privacy, copyright and safe use in plain rules, which are the practical side of the legal and safety developments in this review. After that comes a weekly routine for keeping up without burning out.

    Key Takeaways

    In this tutorial you learned:

  • Use three labels. Solid (verifiable in a primary source), claimed (said by the seller, often with benchmarks) and hype (no clear source or a forecast presented as fact).
  • Models arrived faster than anyone could track. Multiple labs released new versions, access can be paused for reasons outside your control, and the "best model" claim changed hands. Learn the kinds of tool and check current names when you need them.
  • Agents became products. They are arriving inside tools you already use and are worth learning, with small permissions, review steps and checks.
  • Open models closed much of the gap. Estimates differ and are not settled, and the US–China difference in the Stanford report was small in March 2026.
  • Money and adoption are large and uneven. Record investment, wide adoption, pressure on early-career roles in some fields, and a gap between experts and the public.
  • Rules and courts moved, slowly and differently by place. The EU delayed most high-risk dates while adding new bans, and a large US copyright settlement was reported, with the fair-use picture still unsettled.
  • The limits remain. Uneven ability, mistakes, lagging transparency and incidents, and changing costs and access.
  • Turn news into a decision. Write the five-line note, set a review date and re-check the dates that go stale first.
  • Before you move on, a challenge. Take one AI headline from this week, apply the three labels and write down who is speaking, what is solid, what is claimed and what is hype. Then answer one question: what, if anything, should this change about how I work? If the answer is "nothing", that is a perfectly good result.

    Learn AI, after work

    Track your progress, earn XP, and unlock more free tutorials in the AfterWork Academy app.

    Open this tutorial in the app

    More AI tutorials