What Unlimited Free ChatGPT Actually Costs

OpenAI's announcement today leads with generosity: a sharper GPT-5.6 Sol for paying ChatGPT users, and — the headline — unlimited free text chats for everyone else, arriving next week. It's a real giveaway. It's also the cleanest market segmentation OpenAI has ever shipped, and buried in the fine print is the change API developers should care about most: the Sol running in ChatGPT is no longer the Sol you're calling.

The trade free users are actually making

Until this week, ChatGPT's free tier ran on GPT-5.5 Instant with a cap of roughly ten messages every five hours before bumping you to a lighter fallback. The new deal: GPT-5.6 Luna becomes the default for Free and Go users this week, text-chat limits disappear next week (subject to abuse guardrails), and a new Think button lets Luna spend more time reasoning on hard questions. Caps on uploads, image generation, and voice all stay.

For everyday use that's a straight upgrade. Luna is a newer model, OpenAI claims it makes 62% fewer factual errors than GPT-5.5 Instant on factual-detail prompts, and unlimited beats ten-per-five-hours by any measure.

What's gone is the ceiling. Free users used to fall back from a mid-tier model; now they're permanently on the smallest member of the family. The Think button is the tell — it buys Luna more thinking time rather than escalating you to Sol, and OpenAI's availability tables list Sol as paid-only at every effort level. Free users never touch the flagship, a point The Decoder made sharply in its coverage.

The economics explain the generosity. After OpenAI's July price cuts, Luna runs $0.20 per million input tokens and $1.20 output against Sol's $5 and $30 — a 25× gap, per Vellum's pricing breakdown. Unlimited is an easy promise when the marginal conversation costs a fraction of a cent. This is the warehouse-club sample table: all you can eat of the cheap thing, hard ceiling on the good thing, upgrade path clearly signposted.

ChatGPT's Sol is not your Sol

The paid-tier half of the announcement is a version of GPT-5.6 Sol tuned for everyday conversation — more direct answers, tighter formatting, detail that scales with the question. OpenAI says responses containing at least one factual error dropped 68% versus GPT-5.5 Instant in internal evals across finance, medicine, and law prompts.

Two things about that. First, the caveat: it's an internal eval against a superseded model, on undisclosed prompt sets, graded by OpenAI. Directionally plausible — fewer hallucinations has been the entire GPT-5.x pitch — but treat the number as marketing until independent benchmarks weigh in.

Second, and more consequential for anyone shipping on the API: the tuned Sol applies to ChatGPT only. OpenAI explicitly says Codex and ChatGPT Work keep the existing model, and the API's gpt-5.6-sol is untouched. The consumer-tune-versus-API split that started with gpt-5-chat-latest back in 2025 is now standing policy: ChatGPT is a product with its own fork, not a demo of the model behind your endpoint.

The practical rule follows: stop validating prompts in ChatGPT. If your eval intuition about "how GPT-5.6 behaves" comes from the consumer app, it's measuring a different model — one deliberately tuned to be terser and less markdown-happy than the API default. Pin your evals to model IDs, run them through the API, and expect the gap between the two surfaces to widen, because OpenAI just demonstrated it will retune the ChatGPT side on its own schedule.

The thinking slider is an API parameter growing a UI

The other paid change: the separate Instant and Thinking modes are gone, replaced by one Sol with a slider — five effort levels, per The Decoder — controlling how long it reasons. That's the API's reasoning-effort parameter wearing a consumer face, and it confirms where the product line is heading: one family, three sizes, one dial. The model picker OpenAI has been trying to kill since the GPT-5 router is finally dying, replaced by an effort knob users drag instead of model names they memorize.

For API developers the mirror is exact. You pick a size — sol, terra, or luna, with the bare gpt-5.6 alias routing to Sol — and an effort level, and that's the whole interface. If you're still maintaining prompt variants per model generation, the architecture is nudging you toward routing logic instead.

What to do with this

If you're building on Luna, today was quiet validation: OpenAI is about to serve it, unlimited, to the free slice of a billion weekly users. A vendor trusting its cheapest model at that volume says more about production readiness than any benchmark. The caveat is sharp, though — Vellum's testing shows Luna at 41.3% on MRCR long-context recall against Sol's 91.5%. High-volume classification and summarization, yes. Document analysis or big-codebase reasoning, pick a bigger model.

If your product is a thin chat UI over the API, the ground just moved. Free ChatGPT now offers unlimited text chat on a current-generation model, so a wrapper charging $10 a month for roughly that experience has about a week left of looking reasonable. The survivable position, same as it's been since GPT-4o went free in 2024, is owning workflow, data, or distribution that ChatGPT can't reach — because the price of plain chat just hit zero at higher quality than ever.

The verdict: this is good for free users and honestly constructed — nothing was quietly taken away; the ceiling was just made permanent. But read it as the pricing document it is. OpenAI has finished converting its lineup into classic good-better-best segmentation, ChatGPT into the funnel for it, and the ChatGPT-versus-API model split into policy. The free lunch is real. It's also the menu.