The AI Model Timeline

The current recommended workhorse: near the top of the range, at a price that makes sustained use practical.

A genuinely different recipe: breed models from existing ones instead of building each from nothing.

The leading open image lab moved to video — and, notably, did not release these weights openly.

Every major lab now wants to own the full voice pipeline, not just the text in the middle.

Seven significant models in seven days — an unprecedented concentration.

Specialised variants for specific industries are becoming standard.

Labs now ship families rather than models, so there is something for every price point at once.

Open weights used as a distribution strategy, not an ideology.

One lab covering every modality in three days — the scale of the current arms race.

Voice is now its own product line rather than a feature of the main model.

Release cadence compressed to days, which is genuinely difficult to keep up with.

The same pattern again: build an audience with open weights, then close the very top.

A brand-new company opened with near-trillion-scale open weights — the barrier to entry has moved, but not as far as it looks.

The safety review is the story. Governments now sit between a frontier model being finished and being released.

Meta's closed strategy settling into a regular cadence.

Agents, not chatbots, are now what frontier models are designed for.

Learns how software actually gets built — including the debugging and backtracking.

Another major platform choosing openness to attract developers.

Last quarter's frontier at half the price — the clearest illustration of how fast costs fall.

Frontier-scale training is now within reach of any large technology business.

Predictable, frequent open releases became a competitive strategy in itself.

Free weights aimed squarely at the most commercially valuable task.

Notable for what came with it: the most capable variant was restricted rather than sold, a first for a mainstream lab.

The chip maker gives away models to make its hardware the obvious place to run them.

Purpose-built for agents that run for hours, not chat sessions that run for minutes.

Cheap multimodal AI keeps getting cheaper.

Quality over speed as the out-of-the-box choice, reflecting cheap enough inference to afford it.

A dual strategy: give away the strong models, sell the strongest.

The premium tier stopped being where the interesting capability lives.

Application companies started building their own models rather than renting someone else's.

Europe's leading lab held the line on open weights as US labs closed up.

Specialisation is now how labs differentiate, since raw capability has largely converged.

A 1.6T model given away free — an unthinkable proposition eighteen months earlier.

Paying more for extra thinking time became a standard product tier.

Frontier-scale open releases became routine rather than exceptional.

Worth knowing if you pay per token: identical inputs bill differently across this boundary.

Strong reasoning cheap enough to run at consumer scale.

The company that made open weights mainstream stepped back from it — a real loss for the open ecosystem.

Open models tuned specifically for doing things, not just answering.

Removed the custom terms that had made Gemma awkward for commercial use.

Frontier AI stopped being confined to AI companies.

'Small' now describes running cost, not knowledge. These two numbers have fully decoupled.

Frontier releases moved to a roughly two-month rhythm, from roughly annual two years earlier.

Sovereign AI became real. Countries increasingly want models trained on their own languages and values.

Long-context work stopped being a premium feature.

The strongest weights anyone can download and use commercially without asking permission.

Weights that would have been the world's most valuable secret two years earlier, given away.

Multimodal AI at a price point that suits high-volume consumer products.

Whole codebases fit in one conversation, and you no longer configure how hard the model thinks.

Software work is now the highest-value use of these models, and the release schedule reflects it.

Trillion-parameter models stopped being the exclusive property of the largest US labs.

Even the most open-friendly labs started keeping their very best work proprietary.

By the end of 2025 the best free weights were within touching distance of the best paid APIs.

The company that sells the hardware also giving away the models is a hedge: NVIDIA earns nothing directly from Nemotron, but every open model that runs well on its GPUs makes the GPUs harder to replace.

Coding became important enough to justify its own dedicated frontier model.

Long documents got cheaper to process, not just possible.

Image models had been judged on single showpiece outputs. Getting the same face and the same logo to survive across a whole set is what a real design or marketing pipeline actually needs.

Frontier capability moved within reach of small teams and individuals.

The only reasoning models researchers can inspect end to end to understand how the ability forms.

Google retook the lead, and answers started arriving as interactive tools rather than paragraphs.

Faster on easy questions, more careful on hard ones, without you having to say which is which.

An agent re-reads and re-answers dozens of times to finish one task, so the cost per step is what decides whether it is affordable to run at all. Cutting the active size is what moved agents from demo to routine.

A good illustration of the pace: today's budget option is last spring's state of the art.

Open models kept closing the gap on the closed frontier, quarter after quarter.

AI video went from a professional tool to a consumer feed, with all the identity problems that implies.

Agents that work overnight on a real task became a practical proposition.

One model to host instead of two, which halves the cost of running your own.

You stopped needing to know which model to use. The reception was mixed — many people missed being able to choose.

A steady refinement rather than a leap.

A notable reversal from the company that had led the industry into secrecy — largely a response to Chinese open releases.

Strong agent capability at a fraction of frontier pricing.

It became the fastest-downloaded model on Hugging Face within a day of release. The largest open-weight model to date was also one of the most permissively licensed, which set the terms every later Chinese open release had to match.

Spending more compute per question, rather than per model, as a way to buy accuracy.

Long context had been a premium feature of closed models. Handing it to anyone with the hardware to run it made whole-codebase and whole-archive work an open capability.

Made autonomous coding agents genuinely practical — you can hand over a task and come back later.

Complete audiovisual scenes from a sentence. The point where AI video became hard to dismiss.

Frontier-adjacent reasoning became free and self-hostable across every hardware budget.

Thinking and acting merged. The model can go and check rather than guessing from memory.

Legible small text is the difference between a picture of a poster and a usable poster. It is what moved image generation from illustration into commercial design work.

Built for developers wiring models into software, not for people chatting.

Received a notably mixed reception, and marked the beginning of Meta's retreat from open releases.

Thinking before answering became a standard feature rather than a premium option.

Serious multimodal capability for anyone with a single graphics card.

Quietly marked the end of an era: making models bigger stopped being the way forward. Everything after this reasons instead.

You stopped having to pick the right model for the job. The model decides how hard to think.

Demonstrated that the physical build-out, not the research, had become the bottleneck.

Google's long-context, natively multimodal stack became something businesses could depend on.

TikTok's owner has more places to put a model in front of people than almost any lab on earth. Its arrival at the frontier tier is why Chinese model prices fell across the board through 2025.

Wiped roughly $600 billion off NVIDIA's value in a single day. It suggested frontier AI might not need the compute everyone had priced in.

It shipped the same day as DeepSeek-R1. Two Chinese labs independently matching o1 within hours of each other said the reasoning recipe had stopped being a secret.

Enough to hold an entire technical library in view at once.

Broke the assumption that only companies spending billions could reach the frontier. The financial markets noticed a month later.

Taking direction in the vocabulary a director already uses is what makes a video tool something a crew can brief, instead of a slot machine you re-roll until a clip looks right.

Reinforced that how you train matters more than how big you build.

Made continuous, real-time AI assistants cheap enough to deploy widely.

Video generation became a real product, with immediate consequences for stock footage and advertising.

Cloud providers started treating models as a commodity to bundle rather than a premium product.

Most 'open' models release only the finished weights, which lets you run the model but not check how it was made. This one could be audited and rebuilt from scratch — the difference between free software and a free binary.

The shift from AI that answers to AI that acts. Rough at first, but it is the direction everything has gone since.

Open models arrived on-device, where nothing you type leaves your handset.

Became the most-adapted open model family of its generation.

A second way to make AI smarter that has nothing to do with size: let it think longer at the moment you ask. This is the origin of every 'reasoning' model below.

xAI went from newcomer to genuine competitor in under a year.

Took the open image-generation crown and has held it since.

The first time free, downloadable weights genuinely reached the commercial frontier. Estimated cost to train: hundreds of millions of dollars.

Running a capable model became a single-machine decision rather than a cluster decision — the practical start of open weights being a deployment choice instead of a compromise.

Became the default coding assistant for roughly a year, and showed the interface matters as much as the model.

Models started being used to teach other models — an important step as high-quality human text runs short.

Chinese labs became a default choice for developers worldwide, not just domestically.

Real-time spoken conversation with AI became normal, and frontier-grade capability became free.

AI that works with no internet connection and no data leaving your device.

Spend more once on training to save on every query forever. Standard practice now.

Kept efficient open models within reach of the frontier.

Aimed at businesses who need answers grounded in their own documents with a source attached, not clever prose.

More specialisation per expert, and briefly the strongest open model available.

Music had been the last creative field where the AI output was obviously unusable. This is where that stopped being true, and where the argument over training on copyrighted recordings stopped being hypothetical.

The first time a company other than OpenAI clearly held the top spot, which turned a one-horse race into a real market.

Even the most closed labs concluded they needed an open offering to stay relevant with developers.

Changed what people build. Rather than indexing documents into a search system, you can now just paste everything in.

'Open' usually means only the weights. This is the version researchers can actually audit and reproduce.

Data quality turned out to substitute for raw scale — which is why capable models now fit on a phone.

The trick that lets today's models be enormous and still affordable to run. Nearly every large model since uses it.

Google's structural answer to GPT-4 — and the reason your phone can now describe what the camera sees.

The most credible attempt to replace the Transformer. It did not win — attention still holds the frontier — but the ideas turn up inside hybrid models, and it is worth knowing the field did try alternatives.

Easy to overlook at the time. This lab would go on to halve the industry's assumed cost of frontier training.

The first mainstream model positioned on real-time information and fewer refusals rather than raw capability.

Removed the need to learn 'prompt engineering' for images — you just describe it badly and the model fixes it.

Made small models respectable. A 7B model you can run on a laptop stopped being a toy.

Open image generation caught up with the paid tools on quality.

The moment companies could build products on free model weights without legal risk. Thousands did.

You could finally hand a model an entire contract or codebase instead of feeding it in pieces and hoping it remembered.

Proved that well-cleaned web text beats smaller hand-picked datasets, and that serious models could come from outside the US and China.

Google's first genuinely competitive answer to GPT-4, and the relaunch of Bard.

GitHub Copilot had made code generation useful and closed at the same time, and its training on public repositories was already in dispute. This was the open counter-proposal: same capability, with the provenance of the training data open to inspection.

A different bet on how to keep AI well-behaved: write down the values explicitly instead of hoping they emerge from crowd-worker preferences.

The capability jump that triggered the corporate AI rush — and the moment frontier labs stopped telling anyone how their models work.

The leak accidentally created the entire local-AI movement. Within months people were running capable models on laptops.

100 million users in two months, the fastest adoption of any consumer product to that point. The interface, not the model, is what changed the world.

Accurate transcription went from a paid service to a free download, more or less overnight.

Image generation stopped being something you rented from a company and became something you owned. Both the creative explosion and the flood of misuse start here.

A deliberate push back against AI that only works well in English and only exists inside private companies.

Explained why some image generators follow complicated instructions and others ignore half of them.

Unusual honesty about how messy training a huge model actually is — most labs publish only the wins.

The practical recipe for giving a chatbot eyes, and why you can paste a screenshot into a chat today.

The point where AI images went from a curiosity to something people actually used, and where the argument with working artists began in earnest.

First strong hint that prompting a model to show its working makes it genuinely more accurate — the seed of today's reasoning models.

Redirected the industry's spending overnight. Bigger stopped being automatically better, and smaller-but-better-fed models made AI dramatically cheaper to run.

Kept the open ecosystem within reach of the commercial frontier.

The step that turned a text-continuation engine into an assistant that does what you ask. Being helpful turned out to be a training choice, not a size problem.

The model at the centre of the 2022 story about a Google engineer who believed it was sentient. It was not.

Showed scale helps enormously with knowledge and comprehension, but barely at all with logical reasoning — a gap that took another three years to close.

Made it obvious that frontier AI was becoming an infrastructure business only a handful of organisations could afford.

Became GitHub Copilot — the first AI product millions of professionals used in their daily work.

For about a year this was the best model an individual could download and run themselves.

The start of the open-source counterweight to closed labs — a theme running the length of this timeline.

The missing link between language and pictures. Typing a prompt and getting a matching image only works because something taught the machine which words go with which visuals — this is that something.

One architecture for everything. This is why a single model can now handle text, images, audio and video together instead of needing a separate specialist per sense.

This is the engine inside essentially every AI image and video tool you have used — DALL·E 2, Stable Diffusion, Midjourney, Sora, Veo. The pictures come from noise.

The birth of prompting. Instead of hiring engineers to retrain a model, you just describe what you want — which is why anyone can use AI today.

The idea that one general model can replace dozens of specialised ones, made explicit.

A serious technical rival to BERT that showed the field was still exploring how models should read.

The first serious public argument about whether an AI model was too dangerous to release — a debate that has never stopped.

Quietly went into Google Search in 2019 and improved results for millions of queries. Great at understanding text, but it cannot write — it was never built to be prompted.

The blueprint. Everything OpenAI shipped afterwards is this same idea with more data and more compute.

Context-awareness arrived, which is the difference between keyword matching and actual reading comprehension.

Established the train-once, adapt-many pattern that makes today's models economically viable.

The single most important paper on this page. Every major model after 2018 — GPT, BERT, Claude, Gemini, Llama, all of them — is a Transformer. The parallelism it unlocked is what made training on the entire internet financially possible.

Why voice assistants stopped sounding like a chopped-up recording and started sounding like a person.

Depth stopped being the limit on how capable a network could get — a prerequisite for everything larger that followed.

Machine translation stopped being obviously robotic. This is roughly when Google Translate became usable.

The original engine behind convincing fake faces and, eventually, the deepfake problem.

The first time machines had a usable notion of meaning rather than just matching characters. It is why search engines started understanding synonyms.

This is the moment computers got genuinely good at seeing. Almost every AI advance since traces back to the realisation that gaming graphics cards could train neural networks.

The architecture behind roughly two decades of sequence AI — Google Translate, Siri, autocomplete. It is also what the Transformer replaced: reading strictly one step at a time cannot be parallelised, which capped how large these models could ever get.