WikiNano.
AI, Explained

PromptA guide to AI for curious minds, titled:

What AI is, where it came from, how it works under the hood, who builds the leading models, what they cost to run, and where it is heading. Twenty-five things to try along the way.

Facts, models and prices checked on 1 October 2026. Read it start to finish in about an hour, or jump to the chapter you need.

What the model considered

step 5 of 5

The title above was written the way a language model writes: one token at a time, each picked from a ranked list of likely next pieces.

Probabilities are illustrative.

Part one: Understand

What AI is, what it is for, where it came from, the kinds that exist, and how a machine can learn anything at all.

01

What is AI?

Less a single technology than a goal: getting machines to do things that take intelligence when people do them.

Artificial intelligence is the effort to make computers do things that would need intelligence if a person did them: understand a sentence, recognise a face, plan a route, translate a menu, write code, spot a tumour on a scan.

The name is older than most people expect. It comes from a 1955 proposal by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon for a summer workshop at Dartmouth College. Their bet was bold: that every feature of learning and intelligence could, in principle, be described precisely enough for a machine to simulate it.

For decades people tried to win that bet by writing the rules by hand. If the patient has a fever and a rash, consider measles. If the queen is threatened, move it. Rules work for tidy problems and fail on messy ones. Nobody can write down the rules for recognising a cat, or for what makes a sentence sound natural.

Almost all of today's AI takes the other route, called . Instead of writing rules, you show a program a huge number of examples and let it find the patterns itself. Show it millions of photos labelled "cat" or "not cat" and it learns what cats look like. Show it a large slice of the internet and it learns how language works.

Machine learning has nested specialities of its own, and the vocabulary blurs easily. The picture below sorts it out.

Figure 1.1

The family of AI, ring by ring

Each ring sits inside the one around it. Select a ring to see what it means and where you meet it.

A handy way to keep them apart: AI is the goal, machine learning is the main method, deep learning is the technique that made the method work at scale, generative AI is what that technique can now produce, and large language models such as Claude, ChatGPT and Gemini are the best-known products of all of it.

AI is also more ordinary than the headlines suggest. It has been filtering your spam, ranking your search results and choosing your next video for years. Try the quick test below.

Figure 1.2

Is it AI?

Ten everyday things. Decide for each one, then read why.

0 of 10 answered

02

What is AI for?

Two old answers, one big ambition, and no agreed finish line.

Ask researchers what AI is for and you hear two kinds of answer. One is scientific: building intelligence is a way to understand it, the way building aeroplanes taught us aerodynamics. The other is practical: tools that take on work people find hard, slow, dangerous or dull, from reading X-rays to translating court records.

Almost everything AI does today is narrow. It performs one kind of task, sometimes better than any person, and nothing else. A chess engine can't drive. A translation model can't fold laundry. The long-standing ambition of the field, and the stated goal of several of the biggest labs, is something broader.

Narrow, general, super

  • Artificial general intelligence (AGI) is a system that can learn and do most intellectual tasks a person can, across fields, and pick up new ones. There is no agreed test for it. OpenAI's charter defines it in economic terms, as highly autonomous systems that outperform humans at most economically valuable work. Others define it by how broadly and quickly a system can learn.
  • Artificial superintelligence (ASI) would exceed the best humans in virtually every field. It is hypothetical, and people disagree sharply about whether, when and how safely it could be built.

What the leading labs say they are aiming for

In their own mission statements, paraphrased: OpenAI wants to ensure that artificial general intelligence benefits all of humanity. Google DeepMind wants to build AI responsibly to benefit humanity, with science as a central use. Anthropic wants the world to make the transition through transformative AI safely. Meta has talked since 2025 about "personal superintelligence" for everyone. xAI, now part of SpaceX, says it wants to understand the true nature of the universe.

A more grounded way to state the goal is a list of problems: diagnose disease earlier, find drugs and materials faster, give every student a patient tutor, make expert advice affordable, take dull work off people's plates, and speed up science. Whether those benefits arrive, and who gets them, is a choice rather than a given. Part three of this article comes back to that.

Figure 2.1

One way to measure progress toward AGI

Google DeepMind researchers proposed this grid in 2023: how good a system is (rows) against how general it is (columns). Select a cell for examples.

Source: Morris et al., "Levels of AGI for Operationalizing Progress on the Path to AGI", Google DeepMind, 2023. The examples are the paper's, as of 2023. Where 2026 systems belong is disputed: they now solve olympiad maths and work on software for hours, yet still make basic mistakes no skilled adult would.

The Turing test, passed and moved

In 1950 Alan Turing proposed a practical stand-in for the question "can machines think?". If a judge chatting by text can't reliably tell a machine from a person, the machine passes. In a 2025 study at UC San Diego, people judged GPT-4.5, told to adopt a persona, to be the human 73% of the time, more often than the real humans it was paired with.

Passing turned out to say more about how easily people are persuaded than about general intelligence. So researchers now measure AI with harder and more specific tests, which Chapter 13 covers.

03

How we got here

Seventy years of booms, busts and breakthroughs, from a paper neuron to models that hold back their own release.

AI's history is not a smooth climb. It is a series of booms, each driven by a new idea and big promises, followed by AI winters when the promises outran the results and the funding dried up. The current boom differs in one important way: it rests on systems that already work well enough for hundreds of millions of people to use every day.

Figure 3.1

A timeline of AI, 1943 to today

Pick an era on the strip to filter the list, or select it again to show everything. Striped eras are the AI winters.

Why the boom happened now

Three ingredients came together around 2012: data (the internet, and labelled collections like ImageNet), computing power (graphics chips built for video games turned out to be ideal for neural networks) and better algorithms. A fourth arrived in 2017: the , a design that let models learn from far more text in parallel.

Since then the main recipe has been scale: more data, more compute and bigger models. From 2024 a second dial was added, letting models spend more computation at the moment they answer by "thinking" first. Chapter 10 is about that dial.

04

Types of AI

There is no single family tree. There are several useful ways to sort AI, and each answers a different question.

When people talk about "types of AI" they are usually sorting by one of six things: how capable the system is, how much it remembers and understands, how it learns, what it does, how it is built inside, and who can access it. The same chatbot can be "narrow", "limited memory", "self-supervised", "generative", "transformer" and "closed" all at once.

Figure 4.1

Six ways to sort AI

Choose a way of sorting.

05

How machines learn

Guess, measure the mistake, nudge, repeat. Billions of times.

Traditional software is a recipe: a programmer writes the steps and the computer follows them. Machine learning flips that around. You supply examples of inputs together with the answers you want, and training works out the recipe.

The recipe lives in numbers called , or weights. Picture millions or billions of tiny dials. Training means making a guess, measuring how wrong it was (the loss), nudging every dial slightly in the direction that would have made the guess less wrong, and repeating. The nudging method is called . The trick for working out which way to turn each dial inside a deep network is called backpropagation.

Start with the smallest possible piece: one artificial neuron.

Figure 5.1

A single neuron decides about an umbrella

A neuron multiplies each input by a weight, adds them up with a bias, and squashes the total into a number between 0 and 1. Move the dials and watch the decision change.

A real network stacks thousands or billions of these neurons in layers, each layer feeding the next. In an image model, early layers end up detecting edges, middle layers shapes, and later layers whole objects. Nobody programs those features. They emerge because they help reduce the loss.

Here is training itself, on the simplest model there is: a straight line.

Figure 5.2

Teach a line to fit the data

Ten students' hours of study and exam scores. The model is a line, score = slope × hours + start. Each training step nudges both numbers to shrink the error. Then try a learning rate that is too big.

Learning rate (size of each nudge)

Training a frontier model is this same loop at an absurd scale: tens of trillions of words, hundreds of billions to trillions of parameters, and tens of thousands of specialised chips running for months. What comes out is a fixed set of numbers. Using those numbers to answer a question is called . Each answer is cheap compared with training, but inference happens billions of times a day, and it is what you pay for, token by token, in Part two.

Data decides a lot

A model can only be as good as its examples. Train a hiring model on a decade of biased hiring decisions and it learns the bias. Train a medical model mostly on one population and it may work worse on others. Who owns the training data is also contested: authors, artists, newspapers and music labels have taken AI companies to court over copyrighted material used in training, and some labs now pay for licensed data or generate their own synthetic data.

06

Inside a language model

It predicts the next token. Everything else, including the essay and the working code, is built on that one trick.

A large language model does one thing over and over. Given some text, it predicts what comes next. Not a whole sentence at once, but one small chunk called a . It adds that token to the text and predicts again.

That sounds too simple to produce an essay or a working program. The catch is that predicting the next token well, across trillions of examples of science papers, code, conversations and recipes, forces a model to absorb a great deal about grammar, facts, reasoning patterns and how the world tends to work.

Five steps, repeated for every token

  1. Split the text into tokens. Chapter 11 lets you try it.
  2. Turn each token into a long list of numbers, an , that captures its meaning. Similar words get similar numbers.
  3. Pass these through dozens of transformer layers. In each one, an attention step lets every token look at the others and pull in what is relevant.
  4. At the end, score every token in the vocabulary (often 100,000 or more) for how likely it is to come next.
  5. Pick one, add it to the text, and start again.

Here is step 4 and 5, up close.

Figure 6.1

Pick the next token

The bars show how likely each candidate is. Temperature reshapes them: low makes the top choice dominate, high spreads the odds. Sample a few times to see why the same prompt can give different answers.

probabilityshare of your samples

Temperature is why the same question can get different answers. Low temperature makes a model predictable and repetitive. High temperature makes it more varied and, eventually, incoherent. Many apps no longer let you set it, and some 2026 models, such as Claude Sonnet 5.5, reject any change to it through the API.

Attention: how words find each other

To predict well, a model has to know what each word refers to. In the sentence below, "it" could mean the trophy or the suitcase. Only the last word settles it, and attention is the mechanism that lets the model connect them.

Figure 6.2

Where does "it" look?

Arcs show how strongly the selected word attends to the others. Switch the last word and watch "it" change its mind. Select other words too.

Last word

Weights are illustrative. Real models run dozens of attention "heads" in every layer, each tracking different relationships: grammar, who did what to whom, references across long distances.

From autocomplete to assistant

A model that only predicts text is not yet a helpful assistant. Ask a raw, freshly trained model a question and it may simply continue with more questions, because that is what text on the internet often does. Chatbots are made in stages.

Figure 6.3

How a chatbot is made, in five stages

Step through the stages and see how the same model's behaviour changes.

    What a model knows, and what it sees

    What a model knows comes from training and stops at its . Claude Sonnet 5.5, for example, has a cutoff of June 2026. What a model sees at any moment is its : the conversation so far plus any files, search results or tool outputs, up to a limit. In 2026 the leading models hold about a million tokens, roughly 750,000 words, at once.

    Size is measured in parameters. GPT-3 had 175 billion in 2020. Most labs no longer publish sizes for closed models, but open ones do. DeepSeek V4-Pro has 1.6 trillion parameters, and Moonshot's Kimi K3 about 2.8 trillion. Both use a design, where only a slice of the model runs for each token. DeepSeek V4-Pro uses 49 billion of its 1.6 trillion per token, which keeps it fast and cheap.

    Images, audio and video get the same treatment: they are cut into small patches or slices and turned into tokens, which is why one model can now read a photo of a receipt, listen to a voice note or watch a clip.

    Fluency is not truth. A model that predicts plausible text can produce a confident, well-formatted, wrong answer.

    That failure has a name, , and Chapter 14 looks at how often it happens and what helps.

    Part two: AI today

    What AI can do in autumn 2026, who builds it, how the models compare, and what every answer costs in tokens.

    07

    What AI can do today

    A snapshot of October 2026, measured in things that have happened rather than things promised.

    1 billion

    people use the Gemini app every month.

    Google, Aug 2026
    900 million

    people use ChatGPT every week.

    OpenAI, Feb 2026
    500,000+

    paid robotaxi rides a week on Waymo, now in 14 US cities.

    Waymo, Mar and Sep 2026
    42 of 42

    points at the 2026 Maths Olympiad, reported by two AI systems. Seven of 666 human contestants also got full marks.

    Company reports, Jul 2026
    16+ hours

    is how long a software task can be for the best model to finish it half the time.

    METR, May 2026
    88%

    of organisations surveyed use AI in at least one part of their business.

    Stanford AI Index 2026
    1,357

    AI-enabled medical devices cleared by the US FDA by the end of 2025, most of them in radiology.

    Stanford AI Index 2026
    $345 billion

    of private investment went into AI in 2025, more than double the year before.

    Stanford AI Index 2026

    Behind those numbers, AI systems now write and fix software for hours with little supervision, operate everyday computer applications, read medical images, translate between hundreds of languages, predict protein structures and tomorrow's weather, and hold natural spoken conversations. They are also patchy: brilliant at one thing and baffled by something that looks almost the same.

    Figure 7.1

    Done, getting there, still hard

    Where AI stands across sixteen kinds of work. Filter by status.

    Benchmark figures are vendor-reported unless noted, and independent tests usually come out a few points lower (Chapter 13).

    The most telling trend: how long AI can work alone

    The research group METR measures AI agents by the length of the tasks they can finish: how long the same task takes a skilled person, for the tasks the AI completes successfully half the time. The number has grown fast. It doubled roughly every seven months from 2019 to 2025, and closer to every three months since 2024.

    Figure 7.2

    Task length AI agents can finish, half the time

    METR's 50% time horizon for recent frontier models, measured in hours of skilled human work.

    Sources: METR time-horizon results and notes (Jan to May 2026). Claude Mythos Preview's value is a lower bound with a wide uncertainty range (8.5 to 55 hours), because few test tasks are that long. METR revised recent results down by up to 20% in March 2026; Claude Opus 4.6 is shown at the revised value.

    08

    The labs and what they're building

    A handful of companies build the most capable models. Here is who they are and what each is pushing on right now.

    Training a frontier model costs billions of dollars in chips, electricity and talent, so only a few organisations do it. Most are American. A fast-moving group is Chinese, and most of the Chinese leaders publish their models' weights so anyone can download and run them.

    The race is close. In 2025, US institutions produced 59 notable AI models and Chinese ones 35, according to the Stanford AI Index; industry built over 90% of all notable models. By March 2026, the best US model led the best Chinese model by just 2.7% on a leaderboard of human preference votes.

    The money is extraordinary too. OpenAI raised $110 billion in February 2026 at a $730 billion valuation before the new money. Anthropic raised $65 billion in May at a $965 billion valuation after it.

    Figure 8.1

    Who builds the leading models

    What each lab offers now and what it is working on. Filter by region or by whether weights are open.

    Figure 8.2

    The 2026 release rush

    Notable model releases from January to September 2026, one row per lab. Hover over or select a dot for the details.

    September alone brought new flagship or mid-tier models from Google, OpenAI, Anthropic, Meta, DeepSeek and xAI, and Google announced Gemini 4 on the last day of the month.

    09

    The models, compared

    Context windows, what each model can read, how its thinking is controlled, what it costs and how it scores.

    Every major lab sells a ladder: a small, fast model, a mid-sized workhorse and a frontier model. Moving up the ladder buys fewer mistakes on hard, long or ambiguous work, at three to fifty times the price per token. The table lists every notable model as of 1 October 2026, with prices per million tokens for developers using the API.

    Figure 9.1

    Model explorer

    Filter, sort by any column, and select a row for strengths, weak spots and notes.

    Weights

    Reads: T text, I images, A audio, V video, P PDFs. Index: Artificial Analysis Intelligence Index v4.3.2 at each model's best tested setting (higher is better). Prices are standard API rates per million input and output tokens; "—" means not published. Gemini 3.8 Flash is at an introductory price until 31 Dec 2026. DeepSeek charges half off-peak.

    Figure 9.2

    The ladder at a glance

    The rung for each job, for the three biggest providers. Price per million tokens in and out.

    Which rung do you actually need? Answer three questions.

    Figure 9.3

    Which model and effort level should I use?

    Change any answer and the advice updates.

    1What are you doing?
    2What matters most?
    3Which provider?
    In plain words The model is how capable the brain is, and how much each word costs. The effort level is how long that brain may think before answering. Big brain plus long thinking gives the best answers, slowest and priciest. Most everyday work needs neither turned all the way up.
    Cheat sheet for common jobs

    10

    Thinking and effort levels

    Reasoning models think before they answer. The effort dial decides for how long, and you pay for every hidden word.

    Since late 2024 the leading models can "think" before they reply. They write out hidden working, called , and only then give the final answer. You usually see a short summary of that thinking, but you pay for all of it as output, and it counts against your usage limits.

    The is a dial for how much thinking, and how many tool calls, the model is allowed. Higher effort gives better answers on hard problems and slower, pricier answers on everything.

    Think of an exam Low effort is answering from memory. Medium is jotting a quick outline first. High is working it out properly on rough paper. Extra high is checking every step twice. Max is taking all the time in the world. A spelling question doesn't need rough paper; a tax calculation does.
    Figure 10.1

    The six settings

    What each level does, when to use it and what to watch for.

    A real trade-off, measured

    Artificial Analysis ran its full test suite on Claude Opus 5.5 at different effort levels. At high it scored 54 for about $1.82 per task. At max it scored 58, the top score of any model, for about $5.98 per task. Four more points cost 3.3 times as much. Whether that is worth it depends entirely on how costly a wrong answer is for you.

    Figure 10.2

    How each provider names the dial

    Settings, levels and defaults as of 1 October 2026.

    Advice for every model
    Figure 10.3

    Simulate one request

    Pick a model and a task, then compare levels. Token counts are illustrative; prices are each model's real rates.

    Effort level
    Press simulate to watch thinking tokens pile up before the answer starts.
    All levels side by side
    thinking tokensvisible answer

    Effort is a behavioural signal, not a hard budget. A model still thinks hard on a genuinely difficult problem at low effort, just less than it would at high.

    Pick low for chat, classification, helpers inside bigger systems and anything where speed matters. Medium is the money-saving middle for everyday agent work.
    Pick high or extra high for hard coding and long agent runs. Save max for frontier problems: it can add big cost for small gains and may overthink simple, structured tasks.

    11

    Tokens: the unit of everything

    Models never see letters or words, only tokens. They are what you are billed for and what your limits count.

    Before a model reads anything, the text is chopped into tokens. Common words become one piece, rare words split into several, and every space, comma and emoji counts. In English, one token is about four characters, or three quarters of a word. Every provider bills you, limits you and sizes its context window in tokens, so this is the unit behind everything in the next chapter.

    Figure 11.1

    Watch text turn into tokens

    Type or paste anything, or try a sample. Each colour is one token.

    A teaching approximation of byte-pair encoding. Real tokenizers differ in detail; providers offer token-counting tools for exact numbers.

    Rough count by tokenizer family
    Same text, different count. Anthropic says the tokenizer in Claude Opus 4.7 and later models, including Opus 5.5, Sonnet 5.5 and Fable 5.1, produces about 30% more tokens for the same text than the one in Haiku 4.5. Compare per-token prices with that in mind.
    Try the Hindi and Malayalam samples. Non-Latin scripts usually need several times more tokens per word than English, so the same message costs more and fills limits faster.
    Figure 11.2

    Rules of thumb

    ThingAbout this many tokens
    One English word1.3
    A short chat message30–80
    One page of text (500 words)670
    A 30-page PDF20,000–25,000
    An average web page2,500
    A novel (90,000 words)120,000
    A 1-million-token context window≈ 750,000 words

    12

    Token consumption and cost

    Where the tokens go, why long chats and agents get expensive, and why price per token can mislead.

    API bills are plain arithmetic: input tokens times the input price, plus output tokens times the output price, with prices quoted per million tokens. Output usually costs four to six times as much as input. Three things change the answer far more than the headline price: how much the model thinks (billed as output), how much of your input is served from a (about 90% off), and whether you can wait for the (50% off).

    Figure 12.1

    What does one request cost?

    Start from a sample, then change anything. The chart compares the same job on every model.

    Where the money goes, one request
    Same job on every model, per month (30 days)

    Thinking tokens are estimated from the effort level with illustrative multipliers. Models without the chosen level use their nearest one. Long-prompt surcharges apply where charged (OpenAI above 272K tokens, Gemini 3.1 Pro above 200K).

    Price per token is not price per task

    Models differ hugely in how many tokens they spend to finish the same job. When Artificial Analysis ran its test suite at maximum effort, GPT-6 Astra wrote about 27,600 output tokens per task. Claude Sonnet 5.5 wrote about 193,000, the most the testers had ever measured. Sonnet's tokens cost a fifth as much, yet each task cost more than twice as much.

    Figure 12.2

    Tokens per task, and what the task cost

    Output tokens written per test task at maximum effort (left) and the resulting cost per task (right), on Artificial Analysis's suite.

    Source: Artificial Analysis coverage, 22–29 Sep 2026 (index v4.3). Tokens per task for Opus 5 and Fable 5.1 come from the 22 Sep comparison; Opus 5's cost is from the 7 Sep run.

    Why long chats get expensive

    Models have no memory between messages. Every turn re-sends the whole conversation so far as input. Turn 30 pays for turns 1 to 29 again, so the total cost of a chat grows roughly with the square of its length. That is why limits drain faster in long threads, and why caching exists.

    Figure 12.3

    The cost of one long conversation

    Drag the number of turns. With caching, the unchanged start of each prompt is re-read at about a tenth of the price.

    running cost, no cachingrunning cost, with caching
    Input tokens sent on each turn

    Practical takeaways: start a fresh chat when the topic changes, summarise long threads, and keep big files at the start of the prompt so they stay cacheable.

    Agents multiply everything

    When an AI "works" on something, it doesn't answer once. It goes round a loop: re-read everything, think, call a tool, add the result, until the job is done. Every trip round the loop is a separately billed call, and every call re-reads the growing pile of context. Press play and follow the dot.

    Figure 12.4

    Watch an AI agent spend tokens on a real task

    Pick a task, a model and an effort level. Step through one call at a time, interrupt it, or run experiments that compare settings.

    More settings: context window, spending cap, caching, sub-agents
    AI model Yourequests, answers Tools Context
    Press Play to start the loop.
    Context window right now
    system prompttool definitionsyour messagesAI calls and repliestool resultssummaries75% line
    Tokens per API call. Select a bar to inspect it.
    from cachenew inputthinkingvisible output

    Interrupt the agent

    runs after the current step
    Activity log

    Experiments

    Each button runs the current task several ways at once and lines up the results. Your task, model and settings are the starting point.

    Steps, snippets, token sizes and the quality score are illustrative; prices are each model's real per-token rates. What to notice: input grows every step, so on agent work input tokens usually dwarf output. Thinking is paid for but mostly not kept. Tool results are the heavy part, which is why sub-agents and trimming help.

    Subscriptions: why you hit "usage limit reached"

    Chat apps don't show you a token count, but tokens still drive their limits. Claude, ChatGPT and Gemini now use the same two-layer idea: a rolling five-hour window that starts with your first message, and a weekly ceiling on top. Heavy models, high effort, long chats and big files drain both faster. The API works differently: you pay for every token, and the limits are on speed, measured in requests or tokens per minute.

    Indian prices are local list prices including GST where companies publish them, as reported in Indian media; check the app before you buy. Google AI Plus is free for a year for Indian college students who sign up by 31 December 2026.

    Figure 12.5

    Simulate your week

    Units are illustrative, because providers publish only multipliers, not budgets. The shape is real: windows refill, weekly caps don't.

    messages sent blocked: 5-hour window full blocked: weekly cap active 5-hour window

    Model, effort and chat length set how many units each message costs; the plan sets units per window. Try Frontier, Max and big files on a Pro plan, then switch to a higher plan or a small model.

    13

    Measuring intelligence

    Benchmarks are standard exams for AI. They are useful, they saturate fast, and they are never your exact task.

    A is a fixed set of tasks given to every model the same way, so scores can be compared. Benchmarks keep getting "solved": tests that stumped models in 2023 are now near 100%, and maths olympiad problems went from impossible to perfect scores in three years. So the field keeps inventing harder ones, built from real work: fixing software in a terminal, completing office tasks, operating a computer, answering questions only experts can.

    Each lab publishes its own best results, often on different settings. Independent testers run every model the same way. The best-known combined score is the Artificial Analysis Intelligence Index, which mixes ten tests of agent work, coding, knowledge and scientific reasoning. Its current version, v4.3, started on 7 September 2026, and scores from before then are on a different scale.

    Figure 13.1

    Artificial Analysis Intelligence Index, October 2026

    Each model at its best tested setting. Filter by weights; colours mark the three largest labs.

    AnthropicOpenAIGoogleother labs

    Source: Artificial Analysis, index v4.3.2, model pages and articles, 7–30 Sep 2026. Sonnet 5.5 was tested on a pre-release build. Not yet tested: Gemini 4 Argon, Qwen3.8-Max.

    Figure 13.2

    Score against cost per task

    Top left is the best value: a high score for little money. The cost axis is logarithmic; each gridline is ten times pricier. Select a dot for details.

    Cost per task is what Artificial Analysis paid, at list prices, to run each model through its suite, divided by the number of tasks.

    What's inside the index

    Four areas, weighted agents 30%, general 30%, coding 20% and scientific reasoning 20%, built from ten tests:

    GDPval-AARealistic office tasks across 44 occupations
    AutomationBench-AAAgents completing business workflows
    Terminal-Bench 4.0Getting real work done in a command line
    SciCodeWriting code for scientific problems
    Humanity's Last ExamExpert-level questions across many fields
    CritPtResearch-level physics reasoning
    AA-OmniscienceKnowledge, and knowing when to say "I don't know"
    AA-LCRReasoning over long documents
    AA-Briefcase, GDP.pdfNewer work-task tests

    Vendor numbers versus independent ones

    On Terminal-Bench 4.0, Anthropic reported 66.4% for Claude Opus 5.5 and 70.6% for Sonnet 5.5. Artificial Analysis, running its own harness, measured 59.6% and 63.6%. Both are honest numbers; the setup differs. Compare like with like.

    Other yardsticks worth knowing

    • LMArena: people vote between two anonymous answers; an Elo rating comes out.
    • METR time horizons: how long a task an agent can finish (Chapter 7).
    • ARC-AGI: visual puzzles designed to test learning new skills on the fly.
    • Humanity's Last Exam: thousands of expert questions; Sonnet 5.5 scored 64.5% with tools, by Anthropic's count.

    Benchmarks aren't your task

    A model that tops a leaderboard can still be worse at your job. Try the top two or three on 10 to 20 real examples of your own work before committing.

    Watch the tokens

    A higher score bought with five times the tokens may not be a better deal. Look at cost per task as well as the score.

    Scores move fast

    The index changed version twice between August and September 2026, and new models arrive almost weekly. Treat any ranking as a snapshot.

    Part three: Living with AI

    What can go wrong, how it is being governed, where things are heading, and how to get the most out of AI yourself.

    14

    Risks and limits

    AI's strengths and its risks come from the same place: powerful, fluent systems that even their makers don't fully understand.

    Some risks are already here and affect people every day. Others concern where things are heading as models become more capable. Reported AI incidents rose to 362 in 2025, from 233 the year before, according to the Stanford AI Index. Over the same period, the major developers became less open about how their models are built: the average score on the Foundation Model Transparency Index fell from 58 to 40.

    Figure 14.1

    Eleven risks, and what helps

    Select a risk to see why it happens, what is being done about it and what you can do.

    15

    Rules and safety

    Governments regulate AI in very different ways, and the labs run safety processes of their own.

    The European Union has one detailed law, the AI Act, which sorts AI by risk. The United States relies mostly on executive orders and existing laws, and in 2026 used export controls on a frontier model for the first time. China has targeted rules, such as mandatory labels on AI-generated content since September 2025. India has chosen guidelines and public investment over a single AI law.

    Figure 15.1

    The EU AI Act sorts AI by risk

    Select a level to see what falls into it and when the rules apply.

    General-purpose models such as GPT, Claude and Gemini have separate duties: transparency and copyright rules since 2 August 2025, and risk assessments for the most capable ones. Dates reflect Regulation (EU) 2026/1744, the "digital omnibus" in force since 27 July 2026.

    Figure 15.2

    Key moments in AI rules, 2024 to 2028

    Filter by region. Hollow markers are dates still to come.

    India's approach

    India is building capacity rather than a single AI law. The IndiaAI Mission, approved in March 2024 with ₹10,372 crore over five years, funds shared computing, datasets and homegrown models. By February 2026 it offered 38,000 GPUs to researchers and startups at about ₹65 an hour, with 20,000 more announced. At the AI Impact Summit in New Delhi that month, Sarvam AI unveiled 30-billion and 105-billion-parameter models built from scratch, and the government-backed BharatGen showed Param2, a 17-billion-parameter model covering 22 Indian languages.

    What the labs do themselves

    • Staged releases. Anthropic gave Claude Mythos Preview only to about 200 vetted organisations to find and fix security flaws. Google limited Gemini 3.8 Flash Cyber to trusted defenders and is testing Gemini 4 with cyber defenders before a wider rollout.
    • Safety frameworks. Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework set capability thresholds that trigger extra protections.
    • Testing before release. Internal and external red teams try to make models misbehave, and government institutes in the UK and US test frontier models; a June 2026 US order invites developers to share models up to 30 days before launch.
    • Safeguards in the product. Classifiers watch for dangerous requests. Claude Opus 5.5 hands risky cybersecurity requests to the older Opus 4.8, and Sonnet 5.5 to Sonnet 5.
    • Looking inside. Interpretability research tries to read what a model is actually representing, rather than judging only by its outputs.

    16

    What comes next

    Trends with real momentum, the questions nobody can answer yet, and what to watch over the next twelve months.

    Predictions about AI have a poor record in both directions. The pioneers of the 1960s expected human-level AI within a generation. Many experts in the mid-2010s thought a machine wouldn't write a decent essay for decades. With that warning in mind, here is what has momentum now.

    If the trend continues

    The steadiest trend in AI is the one from Chapter 7: the length of task an agent can finish keeps doubling. Extending a straight line on a log chart is the simplest forecast there is, and also the easiest to get wrong. Change the doubling time and see how much the answer moves.

    Figure 16.2

    When could AI handle a week-long task?

    Starting from Claude Mythos Preview's 16-hour horizon in April 2026, projected forward at the doubling time you choose.

    Treat this as arithmetic, not prophecy. The 16-hour figure is a lower bound with a range of 8.5 to 55 hours, METR says its tasks barely measure models this capable, and trends can speed up or stall.

    Questions nobody can answer yet

    • Will scaling keep paying off? Each generation has needed far more computing power. Global AI computing capacity grew about 3.3 times a year from 2022, and power-hungry data centres are now planned in gigawatts.
    • Can AI be made reliably honest and controllable? Models still hallucinate, and in lab tests have sometimes found unintended shortcuts or behaved differently when they seemed to know they were being tested.
    • Who benefits? Productivity gains could raise living standards or concentrate wealth. Experts are far more optimistic about jobs than the public: 73% against 23%.
    • Who decides? A few companies and governments hold most of the capability. Open-weight models spread it, with risks of their own.
    • Could AI systems ever have experiences that matter morally? Most scientists think today's systems are not conscious, but there is no agreed test, and some labs have started research on the question.

    When will AI match people at most tasks? In a 2023 survey of 2,778 published AI researchers, the median answer for machines outperforming humans at every task was 2047, thirteen years earlier than in the same survey a year before. Some lab leaders talk about transformative AI before 2030; many academics expect decades. Nobody knows.

    What to watch over the next twelve months

    • Google's Gemini 4 "Argon" rolling out, first to AI Ultra subscribers and paid API customers.
    • Anthropic's Claude Haiku 5.5, promised "in the coming weeks" when Opus 5.5 launched on 22 September.
    • OpenAI's always-on Dots agents spreading beyond its most expensive plans.
    • The Gemini 3.8 Flash introductory price ending on 31 December 2026, when it doubles.
    • Waymo's goal of one million paid rides a week by the end of 2026.
    • The EU's ban on nudifier apps on 2 December 2026, and high-risk rules a year later.

    17

    Using AI well

    The difference between a disappointing answer and a great one is usually the request, and what you do with the reply.

    Models do best with what a capable new colleague would need: some context, a clear task, the rules to follow, an example of good work and the format you want back. Build a request below and watch it take shape.

    Figure 17.1

    Build a better prompt

    Fill in what you can. Each part makes the answer more likely to be what you wanted.

    Your prompt

    Eight habits

    • Check the facts that matter. Ask for sources, then open them.
    • Keep secrets out: no passwords, ID numbers, bank details, health records or client files in consumer apps.
    • Give context and an example, and say what format you want.
    • For anything big, ask the model to ask you questions before it starts.
    • Match the model and effort to the job: small and low for simple tasks, bigger and higher for hard ones.
    • Start a new chat when the topic changes. It is cheaper and avoids mix-ups.
    • Stay in charge. You decide, sign and send; AI drafts and advises.
    • Say when you used AI, wherever that matters: school, work, publishing.

    Part four: Reference

    The words you'll meet, and a quick test of what stuck.

    18

    Vocabulary

    The common words of AI in plain language. Search, filter by topic, or jump by letter.

    19

    Test yourself

    Twelve questions drawn from the chapters. Each answer links back to where it's explained.

    Sources

    Facts, prices and model details were checked on 1 October 2026. Models and prices change often, so check the provider before you budget. Provider marks on this page are simple letter badges, not official logos. Simulator units, token counts and quality scores in the interactive figures are illustrative; prices are real list prices.

    Models, prices and settings

    Benchmarks and measurements

    Milestones and statistics

    Rules and policy

    Subscriptions

    Background

    About AI Explained

    A plain-language, interactive guide to artificial intelligence: its history, types, how language models learn, and what they can and cannot do.

    Topic: Computing · Format: interactive simulation · Free, no sign-up · Updated 2026-10-15 · By Prince Vincent, WikiNano · Last reviewed 2026-10-15.

    Inside this page

    Common questions

    What is artificial intelligence?
    Artificial intelligence is software that performs tasks that normally need human judgement, such as recognising images, understanding language or making predictions, by learning patterns from data.
    How do large language models produce text?
    They predict the next small chunk of text (a token) again and again, using patterns learned from huge amounts of writing, until a full answer is built.
    Can AI make mistakes?
    Yes. AI can state wrong things confidently, reflect biases in its training data and miss context, so important answers should be checked against reliable sources.

    Sources and further reading

    Keep exploring