Thirty months after the publication of Co-Intelligence, its author is now releasing Co-Existence, a second book dedicated to artificial intelligence. In a blog post illustrated with graphs, he traces the progress made since then: according to him, 180 models now surpass GPT-4 as it was at launch. He also explains that he entrusted the creation of these graphs to an AI, which he says illustrates the changing way of using these tools.
The GPT-4 model, developed by OpenAI, was from February 2023 to early 2024 the best model in the world, according to the author. Since then, twenty different models have held the top spot at one time or another, all from Anthropic, Google, or OpenAI. Among the 180 models that now surpass GPT-4, many are open-weight, meaning anyone can use or modify them freely. Nineteen can run on a recent smartphone.
To measure the progression of capabilities, the author relies on the METR Long Horizon Benchmark, a test that evaluates the amount of human work an AI can accomplish with at least a 50% chance of success from a single instruction. According to this graph, the volume of work achievable doubled eight times between GPT-4’s launch and the end of this measurement. In 2024, an AI could perform a four-minute task. By the time the author submitted his manuscript in April 2026, it could accomplish 17 hours of human work. METR subsequently indicated it could no longer measure this figure accurately with its current methods. Other studies suggest it could reach 17 weeks of work, or even more.
At the same time, prices have fallen. The Google Proof Question and Answer Exam, a test of questions that master’s students struggle to solve, cost 30 cents in AI usage per correct answer. It now costs only four hundredths of a cent to achieve the same result. Progress remains uneven, however: the author judges that AI remains mediocre at writing long texts, while mathematical problem-solving has gone, over the same period, from very poor to superhuman.
A New Way of Working
These gains are changing, according to the author, the way of working with AI. In Co-Intelligence, he described two approaches: ‘centaur’ work, where human and AI tasks are divided, and ‘cyborg’ work, where the two are intertwined. With the arrival of agents, which chain tasks on their own, he adds a third figure, the ‘conjurer’. This one calls upon agents when needed, entrusts them with work, and evaluates their results. The graphs in the blog post were created in this way, using the Opus 5.5 model in Claude Projects, which works with a main agent tasked with launching other agents. The author sees this as the most powerful way to use AI, but also a source of new risks, since one loses sight of what the AI is actually doing to achieve these results.
To illustrate this point, the author also revisited his usual test, the prompt ‘Otter on an Airplane Using Wifi’, which he has used for several years. When given to Opus 5.5, the result, a video created using Google and ElevenLabs services, was produced from the script, images, music, and technical aspects, without any help from him. The author acknowledges that the video has a few flaws. His book, he specifies, focuses on trends rather than the latest models, and on the question of what this changes for humans. Co-Existence is available online.
