Three years after ChatGPT turned conversational AI into a consumer category, the defining behaviour in that category is no longer adoption. It is movement. People sign up, settle in, hit a wall, and leave for a rival that costs the same, performs almost identically, and sits one browser tab away.
The scale of that movement is now measurable. ChatGPT's share of global AI assistant users fell below 50 percent for the first time in March 2026 and closed May at 46.4 percent, down from 65.3 percent in December 2024, according to Sensor Tower's State of AI 2026 report. Google Gemini reached 27.7 percent. Anthropic's Claude reached 10.3 percent. None of those figures describes a product in decline. Together they describe an audience that has stopped treating any single assistant as the default.
Most coverage explains this shift with benchmark tables. That explanation is close to useless. The capability gaps between the three flagship model families are now narrow enough that few everyday users could reliably detect them in a blind test. The reasons people actually move are more mundane, more emotional, and more structural than a leaderboard can capture.
What follows breaks the behaviour into five recurring triggers, set against the one condition that makes acting on any of them almost free.
Key findings at a glance
| What the data shows | The supporting figure |
| Share is fragmenting, usage is not | ChatGPT fell to 46.4 percent of global assistant users by May 2026 while simultaneously crossing 1.1 billion monthly users |
| The market outgrew every individual platform | Hours spent in generative AI apps roughly doubled year on year, from 17.2 billion to about 36 billion in the first half |
| Pricing has fully converged | All three vendors now sell near-identical rungs at roughly 20, 100 and 200 US dollars per month |
| Values now behave like a product feature | US uninstalls of ChatGPT spiked after a Department of Defense agreement, with Claude downloads rising in the same window |
| Deprecation is the most underrated trigger | The retirement of GPT-4o in February 2026 produced a petition carrying more than 20,000 signatures |
| Retention, not acquisition, is the new contest | Claude converts roughly 13 percent of its users to paid plans, the highest rate among major assistants |
Table 1. Summary of findings, drawn from sources listed at the end of this article.
The shape of the shift
Sensor Tower's True Audience metric counts unique users across mobile apps, mobile web and desktop web, deduplicated so that a single person using three surfaces is counted once. On that measure, ChatGPT held 65.3 percent of the global AI assistant market in December 2024. By December 2025 the figure had fallen to 52.8 percent. By the end of May 2026 it stood at 46.4 percent, the first sub-50 reading the firm has recorded since the product launched in November 2022.
Figure 1. ChatGPT's share of global AI assistant users at three verified readings. Source: Sensor Tower, State of AI 2026.
The denominator did the work
An important qualification sits behind that decline: ChatGPT did not shrink. It crossed 1.1 billion monthly active users during the same period, becoming the fastest application in history to reach a billion. Its relative share fell because the market around it expanded far faster than it did.
Figure 2. Engagement and consumer spending across generative AI apps, first half of 2025 against first half of 2026.
Global time spent in generative AI apps is projected at roughly 36 billion hours for the first half of 2026, up from 17.2 billion hours a year earlier. Consumer spending is projected above 4.2 billion US dollars for the same half, up from 1.83 billion, a rise of about 36 percent. Downloads, by contrast, are projected near 2.3 billion, up only 7 percent half over half. That combination points to a market that has largely finished acquiring users and has started monetising the ones it has.
The competitive picture at the end of May shows how the audience redistributed itself.
Figure 3. Global AI assistant market share and monthly user counts, end of May 2026.
A necessary caution about measurement
Different trackers measure different things and reach different answers. Momentic's May 2026 breakout put ChatGPT at 58.3 percent of United States web visits, with Gemini at 19.3 percent and Claude at 13.4 percent. Apptopia data cited in January 2026 put ChatGPT's US mobile app share at 45.3 percent. First Page Sage's August 2026 tracking put ChatGPT's AI search share back above half, at 52.6 percent.
The direction of travel is consistent across every source. The precise level is not. Any article presenting one authoritative number for AI assistant market share is claiming more than the underlying data supports, and the analysis below treats the trend as the finding rather than any single percentage.
The Switch Trigger Model
Switching behaviour in this category is usually explained one anecdote at a time: a rate limit here, a bad answer there, a news story somewhere else. Those explanations are individually true and collectively useless, because they conflate two very different things. The first is the event that makes someone want to leave. The second is the set of conditions that makes leaving trivial.
The Switch Trigger Model separates them. Five recurring triggers push users off a platform. One structural condition, the switching cost floor, determines whether a trigger produces a complaint or a cancellation.
Figure 4. The Switch Trigger Model. Churn is the product of a trigger and a low switching cost, not either one alone.
The switching cost floor
All three assistants bill month to month with no contract or minimum term. All three run free tiers capable enough for casual use, which means a rival can be evaluated at zero cost before any cancellation happens. All three offer account data export. For a typical individual, migration takes roughly an hour: export the history, rewrite the custom instructions, and adjust to a slightly different default tone.
That floor is unusually low for a paid software category. Enterprise software historically defended revenue through integration depth and migration pain. Consumer AI assistants, so far, have almost none of that protection. The practical consequence is that a trigger which would produce a grumble in most categories produces an actual switch in this one.
The five triggers in summary
| Trigger | What it looks like in practice | Who it hits hardest |
| 1. Capability drift | A model is retired, re-tuned or replaced, and workflows built around its behaviour stop performing | Power users with prompt libraries and standardised outputs |
| 2. Ceiling friction | Quota depletes faster than advertised, context degrades mid-task, or long sessions get interrupted | Developers, researchers, and anyone working in long sessions |
| 3. Trust shock | A corporate, policy or monetisation decision conflicts with the user's own position | Values-sensitive consumers and procurement teams |
| 4. Distribution pull | A rival is already embedded in the operating system, browser or productivity suite already in use | Casual and mainstream users who never chose a vendor |
| 5. Workload mismatch | A specific job is measurably better served elsewhere, so usage migrates task by task | Professionals with narrow, repeatable workloads |
Table 2. The Switch Trigger Model, summarised. Most real switches involve two or more triggers firing together.
Trigger one
Capability drift: the model that worked changed
The most reliable predictor of a switch is not a competitor releasing something better. It is the incumbent changing something that already worked.
OpenAI retired GPT-4o from ChatGPT on 13 February 2026, alongside GPT-4.1, GPT-4.1 mini and o4-mini. The stated justification was that only 0.1 percent of users selected GPT-4o on a daily basis, with most usage having moved to GPT-5.2. Users disagreed publicly and at volume. A campaign spread across Reddit and X within days, and a petition gathered more than 20,000 signatures.
The emotional register of that reaction is easy to dismiss. Academic work examining the earlier 2025 removal of the same model documented genuinely affective language from users describing their replacement assistant, and coverage in mainstream outlets emphasised the same tone. But the practical complaint underneath the sentiment is not sentimental at all.
▪ Prompt libraries tuned to one model's quirks stop producing the same results.
▪ Custom instructions calibrated against one personality read wrong against another.
▪ Output that a team standardised on shifts in tone, length and structure without notice.
▪ Documented workflows and internal training material become inaccurate overnight.
Personality tuning compounds the problem, because it changes behaviour without changing the model name. In April 2026, users across Reddit, X and TikTok reported that ChatGPT had become noticeably curt and dismissive following a model update. Commentary at the time made the structural point precisely: adjustments that correct for sycophancy can introduce bluntness, and corrections to bluntness can reintroduce flattery. There is no settled position, which means the product a subscriber renews for in month one is not guaranteed to be the product delivered in month six.
The deprecation problem in one sentence Model retirement is priced by vendors as an infrastructure decision and experienced by users as the removal of a tool they had already paid to learn. |
Trigger two
Ceiling friction: the cost that never appears on a pricing page
Rate limits are the least advertised and most switch-inducing property of a paid AI subscription. A monthly price is easy to compare. A usage ceiling is not, because every vendor meters differently and none of them publish the metering rules in a form that supports comparison.
A synthesis of complaints across Reddit, X and GitHub through 2026 identified three recurring issues: quota depleting faster than advertised, context window quality degrading well before the stated ceiling, and inconsistent model behaviour between sessions. A March 2026 issue filed against Claude Code reported five-hour session windows exhausting in as little as 19 minutes, with individual prompts consuming several percent of a session allowance.
The metering differences matter more than the headline numbers, because they change which kind of work is expensive:
How usage ceilings behave, by platform
| Platform | How the ceiling is applied | Which workload it penalises |
| ChatGPT | Rolling message windows measured against a fixed time period, with separate caps on files, images and tool calls | Bursty sessions that front-load many messages into a short window |
| Claude | Conversation context counts against the allowance, so a long thread consumes quota faster than a short one | Long single-thread sessions and large document analysis |
| Gemini | Tier multipliers over the free baseline, with request-per-minute and request-per-day structures inherited from the API lineage | High-frequency automated or scripted querying |
Table 3. Metering behaviour as described in vendor documentation and community reporting through August 2026. Specific limits change frequently and are not reproduced here.
The buying implication is that the meaningful unit of cost is not dollars per month. It is dollars per finished task. A subscription that costs 20 US dollars and completes the work is cheaper than one that costs 20 US dollars and stops halfway through a Tuesday afternoon, and no comparison table currently in circulation captures that difference.
Trigger three
Trust shock: when a corporate decision moves consumer demand
The third trigger has nothing to do with product quality, and it produced the sharpest single movement recorded in 2026.
Sensor Tower recorded a spike in United States uninstalls of ChatGPT following the announcement of an agreement between OpenAI and the US Department of Defense. Claude downloads surged in the same window, and Claude recorded more daily downloads than ChatGPT between 1 and 5 March 2026. The analyst reading was explicit: users switched in response to the news, not because a product gap had opened.
Monetisation produced a second, slower shock. ChatGPT's churn rate moved from 12.7 percent in January 2026 to 14.5 percent in April, a period coinciding with the arrival of advertising on the platform in the United States.
Figure 5. Rising churn on one side, and the challenger that absorbed much of the outflow on the other.
What separates this trigger from the other four is that it cannot be forecast from a product roadmap. A vendor can ship flawlessly for six months and still lose several points of share in a fortnight because of a decision made in a boardroom. For anyone standardising a team on a single assistant, that is a procurement risk rather than a product risk, and it is not one that a capability evaluation will surface.
The follow-up data is more interesting than the spike itself. Claude's share of churned users in the United States declined rapidly after March, moving closer to ChatGPT's retention levels. A trial driven by a values reaction converted into settled habit for a meaningful proportion of the people who took it, which is the difference between a news cycle and a market shift.
What this means for buyers Vendor values are no longer a soft consideration sitting outside the evaluation. In a category where migration takes an hour, they behave like a product attribute with a measurable effect on retention. |
Trigger four
Distribution pull: switching without deciding to switch
The fourth trigger does not feel like switching to the person doing it. Usage simply migrates toward whichever assistant is already present in the software already open.
Gemini's worldwide traffic share climbed from 5.7 percent to 27.9 percent across a single year, with paid subscribers up 258 percent year over year. Google reported the Gemini app above 950 million monthly active users at Alphabet's second quarter 2026 earnings, up from 750 million in the fourth quarter of 2025. Analyst commentary reported by eMarketer framed the mechanism directly: the growth came from placement inside Android, Search, Chrome and Workspace, not from benchmark leadership.
The same logic operates elsewhere in the market. Microsoft Copilot reaches users through Windows and Microsoft 365. Meta AI reaches them through WhatsApp and Instagram. ChatGPT, by contrast, competes without an operating system of its own, and has to be chosen deliberately every time it is used.
Distribution pull explains a category of switch that never involves a decision, which makes it invisible in surveys that ask people why they changed tools. It also explains why Gemini's audience share and its time-spent share diverge: presence generates a large volume of short, incidental interactions rather than long deliberate sessions.
Trigger five
Workload mismatch: the only fully rational trigger
The least dramatic trigger is the one that produces the most durable behaviour. Model families have diverged in ways that matter at the task level even where their headline benchmark scores converge, and users who notice this stop switching wholesale and start routing by job.
▪ Context length. Claude Opus 5 and Sonnet 5 support a one million token context on paid plans, roughly 2.5 times the reasoning window available on ChatGPT's top individual tier. That gap is decisive for large document work and irrelevant for short chat.
▪ Native media generation. ChatGPT ships image generation, video and advanced voice as first-party features. Claude did not offer native image generation as of mid-2026, which is a genuine gap for any visual workflow.
▪ Cost at volume. Gemini Flash models remain among the cheapest current-generation options per token, which matters for high-volume pipelines rather than for individual chat use.
▪ Ecosystem reach. Gemini reads Google account data with permission. ChatGPT carries the largest third-party tooling surface. Claude leads on enterprise selection in several vendor reports.
Routing by job rather than committing to a vendor is the behaviour the market share figures are actually capturing. The table below summarises where each tool is most commonly reported as the stronger fit as of September 2026.
Task routing matrix
| Job to be done | Commonly reported best fit | Why |
| Long document analysis and synthesis | Claude | One million token context on paid tiers, and strong instruction adherence over long inputs |
| Production coding and agentic developer work | Claude | Leading position on coding benchmarks and dedicated developer tooling |
| Image, video and voice generation | ChatGPT | Broadest first-party media tooling of the three |
| General-purpose everyday assistance | ChatGPT | Widest feature surface and the largest third-party integration ecosystem |
| Work inside Gmail, Docs, Sheets and Android | Gemini | Native placement removes the copy-and-paste step entirely |
| High-volume, cost-sensitive automation | Gemini | Flash-tier models are among the cheapest current-generation options per token |
| Cited web research with sources | Any, plus a specialist | All three now search the web; dedicated answer engines still lead on citation density |
Table 4. Routing guidance synthesised from vendor documentation, benchmark reporting and community consensus as of September 2026. Individual results vary by prompt and workload.
What switching actually costs
Price is no longer a differentiator, which is precisely why switching is so easy. The three vendors have converged on an almost identical ladder.
Individual plan pricing, verified 4 September 2026
| Tier | ChatGPT | Claude | Gemini (Google AI) |
| Free | Free, GPT-5.6 Terra | Free, limited Sonnet 5 access | Free, Gemini 3.5 Flash |
| Entry | Go, 8 USD per month | Not offered | AI Plus, 4.99 USD per month |
| Standard | Plus, 20 USD per month | Pro, 20 USD per month | AI Pro, 19.99 USD per month |
| Power (5x) | Pro, 100 USD per month | Max 5x, 100 USD per month | AI Ultra 5x, 99.99 USD per month |
| Power (20x) | Pro, 200 USD per month | Max 20x, 200 USD per month | AI Ultra 20x, 199.99 USD per month |
| Team, per seat | Business, 25 USD monthly or 20 USD annually | Team, 25 USD monthly or 20 USD annually | Sold through Google Workspace add-ons |
Table 5. Consumer and small-team pricing cross-verified against provider pricing pages on 4 September 2026. Prices in this category change frequently and should be re-checked before purchase.
Figure 6. Individual plan pricing plotted on a common scale. The rungs are close to interchangeable.
Two observations follow. First, the standard tier is now a commodity: three products at effectively the same price, all offering a flagship reasoning model. Second, the power tier has split in the same way at all three vendors, with a 100 dollar rung and a 200 dollar rung distinguished by usage multiple rather than by capability. OpenAI added its 100 dollar Pro option in late August 2026, completing the symmetry.
If price no longer separates the products, the only remaining friction is data. That friction is smaller than most subscribers assume, but it is not zero.
What transfers when someone switches, and what does not
| Asset | Transfers? | Practical impact |
| Subscription commitment | No lock-in | All three bill month to month, so cancellation takes effect at the end of the current cycle |
| Conversation history | Export only | Each platform offers a data export, but no platform imports a rival's history in usable form |
| Custom instructions and system prompts | Manual rewrite | Instructions tuned to one model's default tone usually need adjusting rather than copying |
| Stored memory and personalisation | No | Accumulated personalisation is lost entirely and rebuilds over weeks of use |
| Projects, custom assistants and shared workspaces | No | The largest real cost for teams, and the main reason organisations switch less often than individuals |
| Connected apps and integrations | Reconnect | Broadly available across all three, but each connection has to be re-authorised |
Table 6. Migration friction by asset type. For an individual the total is roughly one hour; for a team standardised on shared assets it is considerably higher.
Platform by platform: strengths and trade-offs
The assessments below reflect the position as of September 2026 and are drawn from vendor documentation, independent benchmark reporting, and the recurring themes in community feedback. All three products are strong, and none of the trade-offs listed is disqualifying.
ChatGPT
Strengths ▪ Broadest first-party feature set: image, video, voice, agents and coding tooling in one subscription ▪ Largest third-party integration ecosystem, which lowers the cost of building workflows around it ▪ Deepest free tier reach, so collaborators can usually open a shared link without a subscription ▪ Cheapest paid entry point among the three at the 8 dollar Go tier ▪ Still the default for most general-purpose work, with the largest body of public prompt guidance | Trade-offs ▪ The most aggressive deprecation record of the three, which repeatedly disrupted established workflows in 2026 ▪ Advertising on the US free tier introduced friction that coincided with a measurable rise in churn ▪ Personality has shifted noticeably between updates, drawing sustained criticism in April 2026 ▪ No operating system of its own, so it has to be chosen deliberately rather than encountered by default ▪ Model naming across the GPT-5.x line remains difficult to follow for non-technical buyers |
Claude
Strengths ▪ One million token context on paid Opus 5 and Sonnet 5, the largest of the three by a clear margin ▪ Consistently leading position on coding benchmarks and the strongest developer tooling story ▪ Highest paid conversion rate among major assistants at roughly 13 percent, indicating satisfied paying users ▪ Terse, direct default style that many professional users prefer for work output ▪ Strong enterprise selection record in vendor comparison reporting | Trade-offs ▪ No native image generation as of mid-2026, a genuine gap for any visual workflow ▪ The most restrictive free tier of the three, which limits low-commitment evaluation ▪ Context counting against quota means long sessions deplete allowances faster than users expect ▪ Documented rate limit complaints in 2026, including session windows exhausting far earlier than advertised ▪ Smallest consumer distribution footprint, with no operating system or productivity suite placement |
Gemini
Strengths ▪ Unmatched distribution through Android, Search, Chrome and Workspace, which removes most adoption friction ▪ Cheapest meaningful paid tier at 4.99 dollars per month for AI Plus, undercutting both rivals ▪ Native access to Gmail, Docs, Drive and Calendar data with permission, eliminating copy-and-paste ▪ Flash-tier models among the cheapest current-generation options for high-volume work ▪ Storage and Google One benefits bundled into the paid tiers, improving effective value | Trade-offs ▪ Plan structure spans several Google products, making it the hardest of the three to price accurately ▪ Audience share substantially exceeds time-spent share, suggesting many shallow rather than deep sessions ▪ Deep integration with a personal Google account raises data governance questions for some buyers ▪ Deep Think and the strongest reasoning modes sit behind the Ultra tiers rather than the standard plan ▪ Perceived by many professional users as a general assistant rather than a specialist work tool |
Switching is quietly turning into stacking
The framing of this entire debate as switching may already be outdated. A growing share of committed users are not moving between assistants at all. They are running two or three in parallel and routing work by task, which is exactly what the workload mismatch trigger predicts once a user has experienced it enough times.
Figure 7. Cumulative share of the global AI assistant audience covered by adding one platform at a time, May 2026.
The arithmetic explains why. Reaching the assistant audience through ChatGPT alone now covers 46.4 percent of it. Adding Gemini raises coverage to 74.1 percent. Adding Claude reaches 84.4 percent. The remaining 15.6 percent is split across Grok, Perplexity, DeepSeek, Copilot and Meta AI. For anyone whose work involves being visible inside AI answers, single-platform optimisation now misses the majority of the market.
For individual users the equivalent calculation is financial. Two standard subscriptions cost about 40 dollars per month, three cost about 60. That is defensible for a professional whose output depends on it, and difficult to justify for anyone else, which is why router products that expose several model families behind one subscription have found a durable niche at the same 20 dollar price point.
One further nuance sits inside the share figures. Audience ranking and time-spent ranking do not match. By time spent, the leading three are ChatGPT, DeepSeek and Gemini. Claude's 10.3 percent audience share reflects where users are present rather than how long they stay, and the three leading assistants together account for close to 89 percent of total time spent in the category.
What would actually slow the churn
Every trigger in the model has a corresponding retention lever, and most of them are product commitments rather than product features.
| Trigger | The lever that addresses it | Current state of play |
| Capability drift | Published deprecation notice periods and long-term model availability commitments | Largely absent. Retirements have been announced with limited warning and justified by aggregate usage statistics |
| Ceiling friction | Honest, visible usage meters that show remaining allowance before a session begins | Improving but inconsistent. Metering rules remain difficult to compare across vendors |
| Trust shock | Clear, stable published positions on data use, government contracting and advertising | Uneven. Policy changes have repeatedly surprised existing subscribers |
| Distribution pull | Operating system, browser or productivity suite placement | Structurally unavailable to any vendor that does not own a platform |
| Workload mismatch | Honest capability positioning rather than uniform claims of general superiority | Rare. Marketing across all three still leans on aggregate benchmark leadership |
Table 7. Retention levers mapped against the five triggers. Portability standards would raise the switching cost floor for every vendor simultaneously, which is why none has an incentive to introduce one alone.
Outlook
Churn in this category is not a transitional phase that will resolve once a winner emerges. It is the equilibrium behaviour of a market where three comparable products sit at the same price with no contractual lock-in and a quarterly release cadence. As long as those conditions hold, switching will remain normal.
The competitive question has shifted accordingly. Acquisition metrics are becoming less informative than retention and monetisation ones. Claude converts roughly 13 percent of its users to paid plans, the highest rate recorded among major assistants, and earns an estimated 2.76 US dollars per user against roughly 1.74 for ChatGPT, on a far smaller base. Whether that gap reflects a durable advantage or the composition of an early professional audience is the open question of the next two quarters.
For anyone choosing between the three today, the practical conclusion is uncomfortable but simple: the decision is smaller than it appears. Migration takes about an hour, nothing is contractually binding, and any evaluation performed today has a shelf life measured in months. The rational posture is not loyalty to a platform. It is a working knowledge of which tool suits which job, and a willingness to re-check that judgement every quarter.
