The market: models, agents, tools and prices

The book tells you who the labs that matter are and what each of them has chosen to be. That part does not age, and it lives on paper. Here is the other half: which model is on top today, what the tools cost, and what has changed since last time.

Who is out there, and with what

Next to every model is the date it was released: on its own that tells you how fresh the row is.

Two rows of this table contradict something the book takes as settled, and it is right to say so. Meta is no longer just the champion of open weights: the Muse family is born closed, you use it inside the company's own apps and through the programming interface, and what Meta publishes openly is a minor model. And Apple holds to its choice of local computation for the small models, but the large one it built together with Google and runs on somebody else's hardware.

What it costs, for people who work

Individual subscriptions, price per month for one user in Italy.

The SpaceXAI figures are the only converted ones: that supplier publishes in dollars only ($30 and $100). Converted at the European Central Bank reference rate of 13 August 2026, 1 euro = 1.1534 dollars, before VAT and before any list-price uplift applied outside the United States. All the other prices are the ones the supplier quotes in euro.

The leaderboards: how to read them

When a new model comes out, the first thing that circulates is where it sits in the ranking. It is worth knowing how that number comes about.

The most quoted ranking is the Arena, the former LMArena. It is not a test: it is a tournament. Somebody writes a question, receives two answers from two models whose names they do not know, and picks the one they prefer. Out of millions of blind comparisons comes an Elo-type score, the same system used to rank chess players.

That explains what the ranking actually measures: the average preference of the people voting, on questions of any kind. It is good information, and it is also the only kind gathered at scale by human beings rather than by an automated test. But there are three things it does not tell you.

It does not say	how the model copes with your work: the people voting ask generic questions, you do not.
It does not say	whether it is right: the people voting pick the answer they prefer, and the one that reads better is not always the correct one. It is the mechanism section 7.2 calls sycophancy.
It does not say	anything about cost: in the ranking a model that costs twenty times another sits one row above it.

This is why the Arena publishes separate rankings by type of task, and the top places are often different. As of 18 September 2026: at the top of the conversation ranking is Claude Fable 5, of the agents ranking Claude Fable 5.1, of the web development ranking GPT-6 Astra. Five weeks earlier the agents and the web development rankings were both Claude Opus 5's: two top places out of three have changed row, and one has changed supplier. Look at the ranking for the task that concerns you, not the general one.

Current ranking on arena.ai →

What has changed

18 September 2026 · revision

The archive of open weights has an owner. On 3 September Nvidia announced the purchase of Hugging Face for 12.93 billion dollars. Hugging Face is the place the open weights in this table are downloaded from: Kimi K3, Muse Glimmer, Qwen and V4.1-Flash all sit there. Nvidia writes that the platform stays open to the whole ecosystem and that nobody will be obliged to use its hardware. It remains a declaration: the open-weights column above now depends also on a promise made by the company that sells the cards.

Four releases in three days, and the top place in web development changes hands. Between 1 and 3 September out come Claude Fable 5.1, Gemini 3.8 Flash, Muse Spark 1.3 and GPT-6 Astra. On web development the Arena now puts GPT-6 Astra ahead of everyone, where in August Opus 5 was. Google, meanwhile, updates the Flash line for the third time in three months and leaves the Pro where it was in February: seven months.

DeepSeek retires its flagship by replacing it with a smaller model. V4.1-Flash comes out on 10 September with downloadable weights under an MIT licence and costs about a third of V4-Pro. Since 14 September, calls to the programming interface addressed to V4-Pro get the smaller model's answer, and the smaller model's rate, until a V4.1-Pro comes out. The large model stays downloadable; it stops being the one that answers.

14 August 2026 · first publication

Three movements matter more than the others, and none of them is about a score.

Meta has stopped being open only. Since April it has published the Muse family with closed weights, inside its own apps and through the programming interface. On 10 August it opened a single model of the family, Muse Glimmer, distilled from the flagship and published under an Apache 2.0 licence. Anyone who built on the idea that a downloadable model always comes out of Meta has to redo the sums.

Google's top of the range has not moved since February. Ever cheaper Flash models keep coming out - 3.7 on 13 August, at half the price of 3.6 three weeks earlier - while the Pro line stays at 3.1. The price is falling much faster than the capability is rising, and it is a trend that goes beyond Google.

Open models have caught up with closed ones in size. Kimi K3, from Moonshot AI, came out in July with 2.8 trillion parameters and downloadable weights: it is the first open model of this class. The company itself says it is behind the closed flagships, which makes the claim more credible, not less.

How to read a model you do not know

These are the four questions from chapter 3, in the order in which it makes sense to ask them.

1. What it can do on your task. Keep three tasks of your own to which you know the right answer, and have every new model do them again. In half an hour you know what no ranking will tell you.
2. How much context it really holds. The declared window is the capacity, not the grip: accuracy drops well before the limit.
3. What it costs per token, in and out, and how much less it costs with the cache. It is the number that decides whether an idea stands up when multiplied by the real volume.
4. Which licence it comes out under. It decides whether you can take the model onto your own servers, and therefore whether somebody can switch it off on you.

Sources

• Official price lists of OpenAI, Anthropic, Google, Microsoft, Mistral and SpaceXAI, pages consulted on 14 August 2026 from an Italian connection.
• Exchange rate: European Central Bank, euro reference rates, value of 13 August 2026.
• Release announcements from the respective suppliers for every model in the table, consulted on 14 August 2026. For Muse Glimmer: research.meta.ai, 10 August 2026, and the model card at huggingface.co/meta-models/Muse-Glimmer-30B.
• Arena, public rankings, consulted on 14 August 2026 and again on 18 September 2026.
• September 2026 releases: Anthropic, Claude Fable 5.1 and Mythos 5.1, 1 September; Google, Gemini 3.8 Flash, 2 September; Meta, Muse Spark 1.3, research.meta.ai, 2 September; OpenAI, GPT-6 Astra, 3 September; DeepSeek, V4.1-Flash and the model card on Hugging Face, 10 September; Alibaba, Qwen3.8-27B, 14 August. Consulted on 18 September 2026. GPT-6 Astra was rolled out in stages between 3 and 9 September, and the sources date the release differently: the announcement is what counts here.
• Nvidia, announcement of the acquisition of Hugging Face, blogs.nvidia.com, 3 September 2026. The figure, 12,930,300,000 dollars, is the one written in the announcement.
• Advertising on ChatGPT in Europe: international press of 21 August 2026 for the start date, the 31 markets and the plans concerned. Consulted on 18 September 2026.
• Where a date is missing, the supplier does not publish a specific release date for that family.