Unpacking the AI Model Market — The Revolution Is Over. The Settlement Has Begun.

November 2022 to March 2026

ChatGPT launched in November 2022. More than three years have passed. Line up what happened in between and a pattern emerges.

Major models only. Minor updates and open-source models (Llama, etc.) excluded.

Early on, OpenAI ran alone. GPT-3.5 to GPT-4 in about four months, no competition in between. Anthropic entered in the second half of 2023. Google arrived in force in early 2024. By mid-2024, three or four players were on the field.

Through this phase, the game was who could release something better first. Launch intervals were irregular — one company shipped, the others followed months later.

From the second half of 2025, the rhythm changes. Version numbers start moving in decimal increments. Claude 4 → 4.5 → 4.6. GPT-5 → 5.2 → 5.3 → 5.4. Gemini 2.5 → 3.0 → 3.1. These look less like new product launches and more like firmware updates.

Launch timing started syncing up. In early 2026: Claude 4.6 in February, GPT-5.4 in March, Gemini 3.1 in March — nearly simultaneous. Early on, one company fired and others scrambled to respond. Now they ship in parallel.

Public reactions shifted too. When GPT-4 dropped in 2023, the news cycle ran for days. When GPT-5.4 dropped in 2026, it barely registered. The technology didn't get worse. New models just stopped being events.

Prices Are Converging

It's not just timing. Prices are landing in the same place.

Consumer subscriptions:

Effectively $20 across the board. This isn't one company setting a price and others matching. The market collectively found this number — the monthly amount people are willing to pay for AI, the same way Netflix and Spotify found theirs.

API pricing still has some spread, but the direction is the same:

March 2026. Official documentation from each company.

Claude Opus is the most expensive — but one year earlier, Opus 4.1 was priced at $15 input and $75 output. That's a 66% drop. OpenAI has fallen sharply from GPT-4o pricing too. Overall, API prices are declining roughly 80% year-over-year. And the gap between the most and least expensive models is shrinking.

Benchmarks Are Within a Few Points

Performance as of March 2026:

Intelligence Index: Artificial Analysis. GPQA Diamond: company disclosures.

The gap between the top three: 4 points and 1.5 percentage points. When GPT-4 launched in 2023, the lead over competitors was far larger. Now the rankings shift depending on which benchmark you pick — Claude edges ahead on coding, Gemini on reasoning, GPT on general tasks.

The competitive framing shifted from "we're broadly better" to "we're better at this." General competition became segmented. From a user's perspective, the question moved from "which AI is better?" to "which one fits what I need?" Any of them is a reasonable choice.

Users Are Splitting

This is measurable. According to Apptopia, daily active user share for US AI chatbot apps:

Apptopia, February 2026.

ChatGPT fell from 69% to 45% in a year. But this isn't ChatGPT shrinking — the overall AI chatbot market grew 152% in the same period. A bigger pie split across more players instead of concentrated in one.

More telling: according to Apptopia, 1 in 5 AI app users now uses two or more AI apps. The single-app era is giving way to pick-what-fits. Apptopia's Tom Grant compared it to streaming — a few major players coexisting in their niches.

This lines up with the supply side exactly. When what's being sold is similar, buyers don't converge on one and stay there. They pick by purpose.

"But Agents Are a New Revolution"

Fair objection. 2025–26 brought agents — AI that operates autonomously on tasks. Claude Code, GPT-5.1 Codex Max, Gemini Deep Think. Qualitatively different from conversational AI. Doesn't that mean a new market opening, not a settled one?

Agents are genuinely new functional territory. But look at where they run. Claude Code ships inside the Claude subscription. Codex Max runs on top of the OpenAI API. No new pricing structure was created — same $20 subscription, same API pricing, with additional capabilities layered on top.

This mirrors the smartphone pattern. The iPhone was a revolution in the phone market. The App Store, a few years later, wasn't a "new market." It was capabilities expanding on top of existing device infrastructure. Agents work the same way — not a reset of the infrastructure, but a functional layer added to it.

Agents could reshape market structure over the long run. But by current data, they're a feature expansion within the settlement, not the start of a new revolution.

"Wasn't DeepSeek a Price Disruption?"

In January 2025, DeepSeek R1 entered at $0.55 per million tokens — over 90% cheaper than existing models. Does that disqualify the "settled market" reading?

What matters is what happened next. Existing players moved fast. OpenAI cut API prices roughly 80% over the following year. Claude Opus dropped 66%. The disruption happened, but the market re-converged to a new equilibrium quickly.

In an early-stage market, a price war like that runs for a long time — attrition with no clear winner. Here, the market rebalanced in months. Fast recovery to equilibrium is itself a characteristic of settled markets.

What the Settlement Looks Like

The last two rows matter most. Supply-side indicators — launch pace, prices, performance — show market structure changing. But the settlement's real meaning is that the question changed.

This pattern repeats across technology markets: the Browser Wars (Netscape vs. IE → Chrome/Firefox/Safari coexistence), cloud (AWS dominance → AWS/Azure/GCP oligopoly), smartphones (iPhone revolution → iOS/Android duopoly).

What the Data Does and Doesn't Say

What the data says: The AI model market is settling into a 3–4 player oligopoly. Consumer subscriptions have converged at $20. API prices are falling ~80% annually with shrinking gaps. Top model benchmarks differ by a few percentage points. Launch patterns have shifted from major upgrades to decimal updates with synchronized timing.

What the data doesn't say: Whether this settlement is permanent. Agents could reshape market structure. Open-source models could disrupt the oligopoly.

The Real Meaning of Settlement

Everything above is supply-side data — the selling side. The real meaning of settlement sits on the demand side.

Three years ago, when ChatGPT launched, reactions were extreme. "This changes everything" on one side, "isn't this dangerous?" on the other. Hallucinations, copyright, job displacement — every launch produced simultaneous excitement and fear. People went through all of it firsthand. They tried AI, were disappointed, tried again, and developed a sense — through experience — of where it works and where it doesn't.

Think about airplanes. When a plane crashes, the fatality rate is far higher than car accidents. Everyone knows this. People fly anyway — not because the risk disappeared, but because the tradeoff between time saved and risk accepted was judged to be worth it. Planes became a settled market not when they became safe, but when the risk-efficiency calculation was made.

AI is following the same arc. Hallucinations haven't disappeared. Bias hasn't been solved. But after three-plus years of direct use, failure, and more use, people developed a felt sense of how big the risk actually is. And the judgment appears to have tilted toward: the efficiency gained by accepting the risk is substantial.

A settled market isn't when prices and performance converge. It's when the risk-efficiency calculation concludes and the question changes — from "is this okay?" to "what can I do with this?" The data shows that transition happening on both sides.