Free AI Is Getting Worse on Purpose

Google is thinning free Gemini on Friday. It's not cost-cutting. It's conversion.

On Friday, Google will start limiting free Gemini users to Flash-Lite, keeping Flash for paid plans and Pro for AI Pro and up. Google’s own help page says the change “will start to take effect for users without an AI subscription on October 9th.” List prices are not going up — AI Plus is still $4.99 a month and AI Pro $19.99. The free default is getting worse on purpose.

That is not a glitch. It is the lever that moves the bill once intelligence got cheap enough that most households never had to pay.

Cheap AI built a free habit

At a given level of performance, the cost of AI has fallen about 47% a quarter since 2023 — roughly 13× a year — per Epoch AI (@EpochAIResearch). One example: early 2025, OpenAI’s o3 hit 75% on GPQA Diamond at about 30 cents a question; under 18 months later, GPT-5.6 Luna matched that score for $0.0004 — a 725× drop in what Epoch called “the price of thought.” That is cost-per-performance on an API benchmark, not the ChatGPT Plus sticker. Plus still sits near $20. The crash funded a better free tier that trains habit without a bill.

read image description

ALT

As of April, barely ~2% of U.S. households were paying for some AI service, a16z’s David George found in State of Markets II. Free got good. Paying stayed rare. AI is getting cheaper faster than people form the habit of paying for it — so platforms that need a household line item make free worse.

A $20 Plus month at Epoch’s $0.0004 frontier unit price buys about 50,000 GPQA Diamond-equivalent answers. That is benchmark arithmetic, not what a consumer app costs to run — but it shows how much value $20 now carries while free still gets a thinner model.

read image description

ALT

A better car gets you there in comfort. A better model gets you somewhere else.

Think of a car lease. A $1,000 lease buys a smoother ride, quieter cabin and better navigation than a $300 one. But both cars pull into the same driveway. The destination is fixed, so the car is a commodity: you pay for the ride, not the result. Intelligence works the other way. A better model doesn't just get you to the same answer more comfortably. It gets you to a different answer, and it gets there faster: deeper reasoning, fewer dead ends, work you can actually ship. That is why thinning free works. Google isn't raising the price of the same ride. It's moving the better destination behind a paywall. A worse free tier is not a price hike. It is a product fork.

Use is common. The bill is not.

Some 64% of U.S. adults use AI, and 25% use it daily, Menlo Ventures (@MenloVentures) found in its July survey of 5,067 adults. Households paying for some AI service were around 2% as of April by a16z’s national yardstick, with PNC’s retail card panel at 2.2% of its households in May. Survey payers — work seats, family plans, bundles — run far higher than the household card; that gap is the argument.

@OehlerStoma66 put the freemium bind in one line: “Biggest tool on earth, still running on freemium muscle memory.”

read image description

ALT

@a16z posted the missing piece in plain English on Oct. 5: “Most people aren’t looking to save time, they’re looking for ways to spend their time.” In a16z’s Top 100 Consumer AI Apps (7th ed.), 9 of 15 consumer internet categories still have zero AI products in the Top 100 — streaming, social, dating, gaming, travel, and more. People already pay for places they hang out. Free AI has mostly been a time-saver sitting next to those bills, not a destination with its own. Habit without a hangout is why the receipt hasn't caught up with the usage.

The pay line is not frozen. PNC’s card panel ran ~0.1% in early 2023 to 2.2% in May 2026 — roughly 22× from a tiny base. Paying is growing fast. It is still 2 in 100.

read image description

ALT

What actually converts

Among adults who say they pay for any AI product — Menlo’s survey payers, not the household card ~2% — daily use is nearly twice as common: 50% vs. 26%. Agent use is five times as common: 65% vs. 13%. Habit and always-on agents turn a free tool into standing usage. Standing usage is what becomes a monthly line.

That is also where the K shows up: a K-shaped split, where one group pulls ahead while the other falls behind. People who can afford the better model get the better reasoning — and the gap between the top and bottom of the K widens, because free and paid no longer lead to the same destination.

Search stayed free for a generation, so the easy forecast is ads forever. Maybe. Ads can skim the casual majority. Subscriptions and bundles monetize people who need depth, agents, and higher limits. Google thinning free Gemini is one fork of that fight. The winner is whoever turns daily habit into a monthly bill — or a bundle seat that behaves like one.

Free AI is getting worse on purpose because free was never the product. Conversion is.

Sources

• Google, “Changes to Gemini model access and limits,” Gemini Apps Help (Oct. 9 free-tier change): https://support.google.com/gemini/answer/17004136?hl=en
• 9to5Google, Abner Li, Oct. 3, 2026: https://9to5google.com/2026/10/03/gemini-model-limits-oct-26/
• Google Gemini subscriptions (US): https://gemini.google/subscriptions/
• Luke Emberson and David Roodman, “The plunging price of thought,” Epoch AI, Sept. 22, 2026: https://epoch.ai/publications/the-plunging-price-of-thought
• @EpochAIResearch, Sept. 22, 2026: https://x.com/EpochAIResearch/status/2102510281176023529
• David George, “State of Markets II,” a16z, Sept. 30, 2026: https://a16z.com/state-of-markets-ii/
• @a16z, Oct. 1, 2026 (“98% of US households aren’t paying for AI yet”): https://x.com/a16z/status/2105726958021885977
• @a16z, Oct. 5, 2026 (“Most people aren’t looking to save time…” + 9 of 15 categories): https://x.com/a16z/status/2107199186286604763
• a16z, Top 100 Consumer AI Apps, 7th ed. (Olivia Moore, Oct. 5, 2026): https://a16z.news/p/top-100-consumer-ai-apps-seventh · companion post: https://x.com/a16z/status/2107109225814233391
• PNC Research, Consumer Health Check, June 2026: https://www.pnc.com/content/dam/pnc-com/pdf/aboutpnc/EconomicReports/consumer_health_check/PNC_Research_Consumer_Health_Check_June_2026.pdf
• PNC Research, Consumer Health Check Chartbook, April 2026 (Fig. 13 Gen AI subscription share series): https://www.pnc.com/content/dam/pnc-com/pdf/aboutpnc/EconomicReports/consumer_health_check/CHARTBOOK_PNC_Consumer_Health_Check_April_2026.pdf
• Shawn Carolan, Amy Wu Martin and Sam Borja, “2026: The State of Consumer AI,” Menlo Ventures, Sept. 16, 2026: https://menlovc.com/perspective/2026-the-state-of-consumer-ai/
• @MenloVentures, Sept. 17, 2026: https://x.com/MenloVentures/status/2100672540431843579
• @OehlerStoma66, Oct. 4, 2026 ("Biggest tool on earth, still running on freemium muscle memory"): https://x.com/OehlerStoma66/status/2106723973438259289
• OpenAI, “What is ChatGPT Plus” ($20/month): https://help.openai.com/en/articles/6950777-what-is-chatgpt-plus