September 23, 2026 · Daily Radar
From seventy percent to fifty percent. In six months. According to Comscore's Q2 2026 AI report, published on September 22, that's how far ChatGPT's share of conversations among AI assistants fell between January and June [A]. Over the same period, Gemini rose from 17% to 30%, and Claude from 2% to 11%. The whole pie is growing while ChatGPT's slice shrinks; in other words, it isn't losing users, but it's no longer the only address.
The same day, a few hours apart, Anthropic announced its new flagship, Claude Opus 5.5, and cut the API price by 20% [A]. The two stories are unrelated, but put side by side, I think they tell companies one thing: at the end of 2026, "We use ChatGPT" isn't a strategy, it's a habit. In a company that hasn't decided which model to use for which job, this picture gets a little more expensive every month.
Quick answer: According to Comscore data, ChatGPT's share of conversations in the assistant market fell from 70% to 50% in the first half of 2026; Gemini and Claude grew fast, and on the same day Claude Opus 5.5 launched at a lower price. This matters to every company that has connected AI to its business processes or plans to — especially automations built around a single vendor. The task isn't to switch models; it's to list which job runs on which model and start making those jobs vendor-independent.
There are two separate announcements, both dated September 22.
Comscore Q2 2026 AI Intelligence Report. Comscore measured AI assistant usage from its own opt-in panel (between 500,000 and 1 million participants, according to Digiday) [A][B]. The conversation share figures are above. In addition, on desktop in June 2026, 168 million conversations were counted on ChatGPT and 89 million on Gemini; on Claude, the number of conversations rose from 1.1 million in October 2025 to 22.3 million in June 2026 [B]. The rate at which AI Overviews appear in Google searches also rose from 25.8% in July 2025 to 39.4% [A].
Let me add a note: this data is US-focused. There's no separate breakdown for Turkey, and if there were, it would probably look different. I'm not carrying the percentages over one-to-one; I'm carrying over the direction.
Claude Opus 5.5. According to Anthropic's own announcement, the model costs $4 per million input tokens and $20 per million output tokens on the API; that's a 20% cut compared with Opus 5 [A]. The price of cache reads dropped 60%, to $0.20. The company says total cost falls by 40% on "typical workloads." The model is also available through AWS, Google Cloud and Microsoft Azure, and the five-hour usage limits for Pro, Max and Team subscribers were raised [A]. Sonnet 5.5 and Haiku 5.5 are said to be coming "in the coming weeks," with no date.
Because most companies adopted AI not as a tool but as a brand. The marketing team got used to ChatGPT, someone built an automation, that automation was connected directly to the OpenAI API, and that's where things stayed. Nobody asked "which model is best for this job," because a year ago the answer was already obvious.
It isn't obvious anymore.
A few numbers show why this is more than a technical detail:
In my view, the really interesting part isn't the measurement. In the words of Comscore executive Smriti Sharma, assistants are now "a layer of consumer behavior" [B]. If it's a layer, it's only natural that everything you build on top of it is affected by changes in that layer.
| Setup tied to a single vendor | Vendor-independent setup | |
|---|---|---|
| When a price cut arrives | Taking advantage of it requires code changes | It's tried with a settings change |
| When the vendor changes limits or policy | Work stops or slows down | You switch to a backup model |
| When a new model comes out | The "should we switch?" debate drags on for weeks | It's compared in a day with the same test set |
| Setup cost | Low | Higher at first |
| Who it's for | A team in the trial phase | A company that has connected AI to customer processes |
Don't skip the last row. Not every company needs an independent setup; if you're still in the trial phase, this table is more than you need, and that's fine.
I explained where to start with AI without getting locked into a single vendor in the AI transformation guide. I also covered the search visibility side of the same Comscore data in a separate post.
Write every job that AI touches in your company on a single list: the job name, which tool or model, who uses it, and whether it depends on an API key. It's a half-hour task. At the end of the list, you'll have the answer to "if this vendor doubles its price tomorrow, which job stops?" If the answer is "none," you're in good shape. If not, you now know where to start.
| News | Why people are talking about it | Who it affects |
|---|---|---|
| Claude Opus 5.5 launched, API price cut by 20% (September 22) | Cache cost in long agent jobs fell 60%, but there are four breaking API changes | Teams building agents and automations → AI Automations |
| Live video ads on Instagram open on September 29 | Meta's new Creator Marketing Hub lets brands boost a creator's live stream as a Partnership Ad | Brands working with creators → Influencer Marketing |
| New AI Max reports in Google Ads, AI Brief in more languages (September 23) | Google also published a guide to comparing AI Max with Dynamic Search Ads through experiments | Anyone managing search ads → Google Ads |
| Claim that the external link rate in AI Overviews has risen above 25% | Peec AI data; Search Engine Roundtable was also cautious, and the methodology wasn't disclosed | Content publishers → GEO Agency |
| "Dive Deeper" test in Google Discover | It steers users to an AI summary instead of the publisher's article | Sites getting Discover traffic → SEO Management |
| Sponsored answer rate in hotel queries from 6% to 24% (Comscore) | The first serious data where in-assistant advertising has started to be measured | Brands watching assistant advertising → ChatGPT Ads |
I'm deliberately keeping the AI Overviews external link claim in the table, but I'm not giving it weight. It's single-source, the method is unclear, and the editor who covered the story openly disclosed having an investment in the source [C]. Until it's verified, I'm only watching it.
No, at least not based on this data. Comscore's figure is market share; it doesn't tell you which model is better for your work. The right approach is to try your three most-used tasks on two different models and compare the results side by side.
Partly. The list price dropped 20%, that's certain [A]. The 40% figure assumes running at a lower "effort" setting [C]; if your work keeps its quality at that setting, it holds, and if not, it doesn't. Without measuring on your own workload, it's hard to say anything.
Unknown. The data Comscore published is US-focused and has no breakdown for Turkey. I'm not going to guess; separating AI-referred visits by assistant in your own site's analytics is the most reliable local data you have right now.
If you'd like to clarify together where AI stands in your company and which job should run on which model, take a look at our Enterprise AI Consulting page. For those who want to see the big picture first:
When you train employees on AI, giving everyone the same seminar doesn't work. We cover three role-based levels, the modules, a time plan, data rules and how to measure whether the training worked.
AI projects often either never get started or get stuck in the pilot. Here is a 90-day plan split into three phases, discovery, pilot and rollout, with the deliverables you should have in hand at the end of each one.
October 4, 2026: starting October 9, Google is limiting the Gemini app on free personal accounts to the Flash-Lite model only, and Pro is also going away on AI Plus.
We use cookies to improve the site and measure advertising performance. Analytics and advertising cookies run only if you accept them. Details are in our Privacy Policy and Data Protection Notice.
