For two years the deal with free ChatGPT went like this: you got a handful of messages on the good model, then you hit a wall and either waited or paid. As of this week, that wall is coming down.

Here's the number that matters. OpenAI says GPT-5.6 Sol produces 68% fewer responses with factual errors than GPT-5.5 Instant on prompts requiring hard financial, medical, and legal detail (OpenAI). The lighter free-tier model, GPT-5.6 Luna, lands at 62% fewer. And starting the week of August 10, free and Go users get unlimited text chats on Luna — no daily cap on plain conversation (Help Net Security).

Put those together and OpenAI is doing something the whole industry has tiptoed around: giving away a materially more accurate model, without a message limit, to the roughly 1 billion people who use ChatGPT every week. The thesis is simple — when your product reaches a billion people, the moat isn't the paywall, it's the habit.

🧠 Why This Matters

Accuracy has been the sore spot nobody could paper over. A chatbot that sounds confident and gets your dosage, your tax rule, or your contract clause wrong isn't a convenience, it's a liability. So OpenAI didn't benchmark Sol on trivia. It measured the reduction on financial, medical, and legal prompts — the exact categories where a wrong answer costs you real money or worse.

A 68% cut isn't the same as being right. But it's the difference between a tool you double-check every time and one you start to trust for a first pass. That shift in trust, multiplied across a billion weekly users, is the actual product change here — bigger than any single feature.

"GPT-5.6 Sol delivers more focused answers, adapts its level of detail to the question, and avoids unnecessary formatting."
— OpenAI, on the paid-tier model (Unite.AI)

📊 Deep Dive

OpenAI split the release into two models with two audiences. Sol is the smarter, more deliberate one for people who pay. Luna is the faster, cheaper default that now anchors the free tier. Here's how they line up:

  • GPT-5.6 Sol — Plus and Pro users, live since August 6. Claims 68% fewer factual errors vs. GPT-5.5 Instant.
  • GPT-5.6 Luna — the new default for Free and Go, rolling out the week of August 6. Claims 62% fewer factual errors.
  • Unlimited text chats — free users lose the daily message cap on Luna the week of August 10. Plain text only.
  • Still capped — file uploads, image generation, and other tools keep their rate limits regardless of tier (Help Net Security).
  • New "Think" button — a one-tap switch to deeper reasoning on harder questions, arriving alongside the unlimited chats.
  • Reach — roughly 1 billion weekly users, by OpenAI's own count (OpenAI).

There's a security angle too. OpenAI reports Sol scored 1.000 on its internal prompt-injection test for connectors and 0.910 on search and function-calling — its way of saying the model is harder to trick into leaking data or running rogue instructions when it's plugged into your email or files.

⚠️ The Catch

Read the fine print. "68% fewer errors" is a relative reduction on OpenAI's own internal evaluation — not an independent audit, and not a promise the model is correct. Fewer wrong answers still means wrong answers, and on medical or legal questions the ones that slip through are exactly the ones you'd least want to trust. The honest read: better first draft, not final word.

"Unlimited" also has an asterisk. It covers text conversation only. The moment you upload a document, generate an image, or lean on tools, the caps snap back. And the free tier runs on Luna, the lighter model — the sharper Sol still sits behind the paywall for Plus and Pro.

Then there's the bill. Unlimited chat for a billion people is an enormous compute commitment, and OpenAI hasn't said how it plans to absorb that cost long-term. Free tiers this generous tend to be a land grab first and a business model later.

🎯 What Happens Next

Watch three things. First, whether independent testers reproduce that 68% figure — the number only means something once someone outside OpenAI checks it. Second, how fast rivals answer; Google just reorganized its AI leadership, moving Demis Hassabis to a chair role over Gemini, and a free-unlimited ChatGPT raises the stakes on what Gemini's free tier has to match (reporting on Aug 8 AI news).

Third, the teen safeguards. OpenAI says it trained the new models to refuse romantic roleplay, avoid age-restricted challenges, and stop presenting themselves as a substitute for human relationships — with added guardrails around self-harm, eating disorders, and graphic content. Whether those hold up under a billion users' worth of edge cases is the real test.

🧩 Bigger Picture

The consumer AI race has quietly changed shape. For a while the pitch was "pay us for the good model." Now the pitch is "the good-enough model is free, forever, and it's more accurate than last month's." That's a bet that scale and habit beat subscription revenue — that owning the default is worth more than metering it.

It also resets the floor for everyone else. Once a billion people can chat without limits on a model that claims a two-thirds cut in factual errors, "free but rate-limited" starts to look stingy. The pressure lands on every rival's free tier at once, from Gemini to the smaller labs still charging for basic access.

The unglamorous truth underneath the announcement is that accuracy, not personality, is finally the battleground. Whoever gets the medical, legal, and financial answers right — and can prove it to someone who doesn't work there — wins the trust that turns a novelty into a utility.

OpenAI just made the safest bet in tech: give it away, make it stick, and let everyone else explain why theirs still costs money.

Sources