Choosing models while the big labs fight

For years the large AI labs competed mostly on benchmarks. In 2026 the rivalry moved into the open: into Super Bowl advertising, into a dispute with the Pentagon, into hiring, and into prices that fall every few weeks. OpenAI, Anthropic and Google lead, and Meta is pushing in from the side.

For a company that uses these models, the interesting question is not who wins. It is how to keep working well whichever lab is ahead next quarter. Here is what happened this year, and what we take from it.

Ads and principles

Earlier this year OpenAI introduced ads in ChatGPT to pay for its free and cheaper plans. In August it took them to India, and this month it adds visual ads next to the images ChatGPT generates, in the US first. OpenAI says the ads are labelled and do not influence the answers.

Anthropic took the opposite position and made a campaign of it. On 4 February it announced its first Super Bowl advertising, made by the agency Mother under the line A Time and a Place, with a promise that Claude will remain free of ads. It wrote that it funds Claude through business contracts and paid plans, and that it would explain openly if that ever changed.

Both are business models as much as principles. One lab pays for free users with advertising, the other with enterprise contracts. A company should know which one it buys from, because the incentive shapes the product.

The Pentagon split

The sharpest break came in February. On 26 February Anthropic’s chief executive, Dario Amodei, published a statement: the Department of War wanted its AI suppliers to accept any lawful use, and Anthropic refused two, mass surveillance of Americans and fully autonomous weapons. The next day Defense Secretary Pete Hegseth directed that Anthropic be designated a supply chain risk, and President Trump told federal agencies to stop using its products, with six months to phase them out.

Within hours OpenAI announced its own agreement with the Pentagon and said it kept the same red lines. Users noticed. On 28 February Claude rose to number one in the free app rankings of Apple’s US App Store, ahead of ChatGPT, TechCrunch reported. On 7 March OpenAI’s head of hardware, Caitlin Kalinowski, resigned and wrote that the announcement had been rushed before its guardrails were defined.

In March a federal judge in California blocked the designation while the case continued. In April Google gave the Pentagon access to its AI on classified networks, reportedly for all lawful uses, after about 950 of its own employees had asked it to follow Anthropic’s line. In August the same judge ruled the designation unlawful retaliation. On 25 September the appeals court in Washington upheld a parallel designation under a different law, by two votes to one.

People, prices and share

The labs also compete for the same few people. In May Andrej Karpathy, a cofounder of OpenAI, joined Anthropic’s pretraining team. In June John Jumper, who shared the 2024 Nobel Prize in chemistry for AlphaFold, left Google DeepMind for Anthropic, in the same week that Noam Shazeer left DeepMind for OpenAI.

Prices fall with almost every release. On 22 September Anthropic’s Opus 5.5 cut the output price from 25 to 20 dollars per million tokens. A week later OpenAI released a model that it says comes close to its flagship at a fifth of the standard price. On 30 September Google released Gemini 4 Argon, which it says beats both rivals on a range of benchmarks, so far for selected security partners only.

Who leads depends on what you count. Among businesses that buy models through an interface, Menlo Ventures estimated in December 2025 that Anthropic took 40 percent of spending, OpenAI 27 percent and Google 21 percent. Among consumers, Sensor Tower put ChatGPT at 46.4 percent of the AI assistant market at the end of May, below half for the first time, with Gemini at 27.7 percent and Claude at 10.3 percent. The Gemini app passed one billion monthly users in August, two months after ChatGPT, and in September Meta’s new assistant Muse reached the top of the US App Store.

Even revenue is hard to compare: Anthropic counts sales made through its cloud partners and OpenAI does not, TechCrunch noted this month.

When a supplier changes the rules

In a race this close, suppliers change their terms quickly, and customers find out late. In April Anthropic stopped Claude subscriptions from covering third party tools such as the agent OpenClaw, so heavy users had to pay separately. And Google is turning Gemini’s Gems, the custom assistants many people built, into a new format called skills from 17 November.

None of this is a scandal. It is what suppliers do in a race. But a company that has built its work into one of them inherits every one of these decisions.

Stay free to switch

Our view is simple. Use the best model for each job, and arrange things so that the model is a setting, not a foundation.

Keep the knowledge outside the model. Your prices, rules, processes, voice and skills belong in a brain your company owns, written down in files you control and connected to whichever assistant is best this year.

Test before you trust. Run new models on your own tasks, not only on someone else’s benchmark, and switch where it pays.

Decide where data must stay. For confidential work, a private model in your own cloud or on your own servers takes the question of whose terms apply off the table.

Choose suppliers with open eyes. Ads or no ads, which uses a lab permits, how it treats the tools you depend on: these are questions of values and risk, not only of benchmarks. The machines can compare the scores. The choice should be made by a person who can answer for it.

This is how we built Twin, a company’s AI brain: its knowledge, rules, skills and integrations, delivered as source code. It works in Claude, ChatGPT and Codex, or privately on Eigna, a model of a company’s own, built by our partner Eigna Labs, trained on its work and running in its own cloud or on its own servers. If a client changes assistant, the brain goes with it. We run crue the same way, on our own Twin, in Claude and ChatGPT and on our own hardware with our own model.

The labs will keep fighting, and that is good for prices. Make sure the fight never decides how your company works.