Almost nobody's work is being deleted. It is being renamed, split in two, or quietly stopped at the entrance. Here is where it went, and what the new version is called.
The most useful thing to understand about this moment is also the least dramatic: the overwhelming majority of jobs touched by AI are not being abolished. They are becoming a next version of themselves. The recruiter becomes the person who can tell a real machine learning engineer from a fluent one. The sales engineer becomes the person who demonstrates a system that is sometimes wrong. The lawyer becomes the person who audits a model's reasoning. The pattern repeats across almost every function, which means the useful question is not whether your job survives but what it is called now.
The second thing is harder to hear. The damage is real and it is concentrated almost entirely at the entrance. Employment for software developers aged 22 to 25 has fallen about a fifth since 2024, and the mechanism is not redundancy. Nobody was fired. Companies simply stopped opening the door, which is precisely why there was never a headline about it and why applying feels like shouting into an empty room. Against that, engineering hiring overall is down far less than the rest of tech. The profession is not shrinking. Its entrance is crowded.
The third is the one that pays. Some jobs are hiring hard while almost nobody is qualified for them, and they sit in unglamorous places: licensed electricians, liquid cooling technicians, commissioning agents, grid interconnection specialists, AI auditors. Several want a trade or a certificate rather than a degree. They are wide open for a dull structural reason, which is that the training takes four years and the demand arrived last year.
Search the fashionable title and you will find nothing, while the work is being hired for hard under a different name.
Six of them. The same move repeats across nearly every function.
Somebody in your office pasted an invoice into ChatGPT this week. Was that fine? It depends on the account they used, and almost nothing else.
These companies are not charities and running this costs real money. So somebody pays. If it is not your credit card, it is your documents. On a free account, everything you send is used to train their AI: what you type, the files you upload, the photos, the voice notes. Paying twenty dollars a month changes almost nothing, because the vendors put the free plan and the paid personal plan in the same line of their own terms. It buys you features. It does not buy you a different contract.
A company account does. There, not using your documents is written into the agreement you sign rather than offered as a setting somebody has to remember. Of sixteen tiers checked against the vendors' own policies, every personal one that could be verified trains on your data, and not one company account does. So the fix is not finding a safer vendor. It is moving your staff off personal logins, and it costs about what they are already paying out of pocket.
The more you pay, the less they use it. That one sentence explains every row in the tables.
The question every organisation is stuck on, answered without the usual deflection about how you already use Gmail.
No, you cannot hand it anything blindly. No, it is not all dangerous either. The same model is safe or unsafe depending on the door you walk it through: on your organisation's own account your text is not used to train anybody's model, it is held on your terms and it is auditable, and on a personal login none of that is true.
Which makes the decision smaller than it feels. What kind of information is this, which account am I using, and what is the thing allowed to do once it has it.
Nine decisions determine whether an agent works in production, and most teams get stuck on the second one because they collapse state and memory into the same idea. They are not the same idea. One is what the system knows right now, the other is what it is allowed to carry between runs, and confusing them is why agents behave well in a demo and badly on a Tuesday.
Every model, chip, platform and lab we track, with specs, prices, licences and the people who built them. This is the material the writing above is drawn from, and every figure carries its source and its date.
01
How the field fits together
1ModelsThe frontier and the open weights.Opus 5 · GPT-5.6 · Kimi K3
2MediaImage, video, voice and music.Kling 3.0 · GPT Image 2
3ToolingAgents, routers and frameworks.Claude Code · LangGraph
4SiliconChips, clouds and capital.B300 · MI400 · RunPod
Each layer depends on the one beneath it. Silicon decides what can be trained, tooling decides what can be shipped, and the models everyone argues about sit on top.
Output tokens, per million, and a 20× spread across the frontier. Capped at two models per lab; only 14 labs publish per-token prices at all, which is itself worth knowing.
Every LLM you can call or download today — frontier, mid-tier and open weights. · 50 entries · 8 older versions hidden · ranked by capability and adoption · verified 20 Sept 2026
Turning pages into structured text. Now dominated by vision language models rather than the classical OCR engines. · 13 entries · ranked by document parsing benchmarks · verified 2 Sept 2026
Finding and boxing things in images and video. The licence matters as much as the accuracy here. · 11 entries · ranked by COCO mAP, and licence realism · verified 13 Aug 2026
Where embeddings live, and the layer every retrieval system is built on. · 12 entries · ranked by adoption, then measured latency · verified 13 Aug 2026
Retrieval, augmentation and generation, plus the graph approaches that connect facts rather than merely matching them. · 14 entries · ranked by fitness for retrieval work, then adoption · verified 13 Aug 2026
