Claude came first in a battle of AI in a YouGov satisfaction survey with a net score of 56.2, ahead of Gemini at 46.5 and ChatGPT at 46.0. TechRadar covered the results, which came from YouGov's BrandIndex tracking of UK users. Alexa landed at 42.6, Copilot at 37.2, and Siri and Grok below that.
The more telling number is further down. YouGov also split current users from former ones, and Claude scored highest of any assistant among people who had stopped using it. Grok's former-user score was -4.7. ChatGPT sat 63.7 points higher with current users than with people who had left. It tells you what people say about a product once they have no reason to defend it, and Claude is the only one that comes out of that test well.
But now look at when the data was collected.
The survey mostly predates the Claude people are complaining about
YouGov ran this from March 1 to July 31, 2026. Sonnet 5 shipped on June 30. Opus 5 shipped on July 24.
So across a five-month window, Opus 5 existed for barely any of it. For the great majority of the period being scored, the Claude people were rating Opus 4.6, then 4.7, then 4.8. That is the era this 56.2 belongs to.
Which matters, because the complaints that have piled up since are specifically about the new models. Too long, too chatty, too eager to expand a small request into a project. Those are not vague vibes either, they are in the release notes as intended behavior. Anthropic's own Opus 5 page says default responses "run longer," the model "narrates its progress to the user more often," and it "verifies its own work without being told to."
A satisfaction score is a lagging indicator and this one was earned by a version of Claude that Anthropic has since changed on purpose.
What the old models had was restraint
"It was better in February 2026" is easy to dismiss as nostalgia.
Opus 4.6 answered the question you asked. It did not think unless you turned thinking on, so a small request stayed small. It did not narrate its plan, delegate to subagents or verify work you had not asked it to verify. If you wanted more, you asked for more.
That restraint is what produced the reputation the survey measured. Not raw capability, since 4.6 is comfortably behind Opus 5 on every benchmark that exists. People rated it well because working with it was low friction, and low friction is what a satisfaction score actually captures.
The lead is real and it is the easiest kind to lose
Beating Gemini by ten points and holding the best former-user score is a genuinely strong position, and it is worth noticing what Anthropic did to earn it, which was mostly to build something people did not resent using.
But satisfaction leads are fragile in a way capability leads are not. A benchmark gap takes a competitor months of training to close. A goodwill gap closes the moment your product starts being annoying, and it closes fastest among exactly the heavy users who generated the score, because they are the ones who feel every extra paragraph. Codex and the open models have already closed most of the capability gap, which means pleasantness is a larger share of what is left to compete on, not a smaller one.
The fix is not a smarter model. It is a default that has less friction and just easier to work with. Ship a model that answers the question, stays inside the scope it was given, and only escalates when it is asked to. Anthropic knows how to do that because they did it in February.
The next YouGov window will cover a period where Opus 5 is the default rather than a rounding error. That number is the one worth watching.
Sources
TechRadar: Claude wins out in major user satisfaction survey - The YouGov BrandIndex results covering March 1 to July 31, 2026 among UK users, with Claude at 56.2, Gemini at 46.5, ChatGPT at 46.0, Alexa at 42.6 and Copilot at 37.2, plus the current-versus-former user split showing Claude highest among former users, Grok at -4.7, and gaps of 63.7 points for ChatGPT, 62.5 for DeepSeek and 62.4 for Grok.
Anthropic: What's new in Claude Opus 5 - The July 24, 2026 behavior changes stating that default responses and written deliverables run longer, the model narrates progress more often, delegates to subagents more readily and verifies its own work without being told, plus thinking being on by default as a change from Opus 4.8.
