OpenAI Refines GPT-5.6 Sol for Paying Users, Demotes Free Tier to Luna

What changed in GPT-5.6 Sol

Starting August 6, ChatGPT Plus ($20/mo, approximately €18.40 in most EU markets after VAT) and ChatGPT Pro ($200/mo, approximately €184) users got an updated version of GPT-5.6 Sol inside the Chat experience. The update brings three main improvements, according to OpenAI's announcement.

First, more focused answers. OpenAI says the new Sol trims unnecessary detail and excess formatting. For a simple question, you get a direct answer with just enough context. For complex tasks — multi-step planning, research, writing — the response is fuller but still keeps the key recommendation visible. The company included a side-by-side example of a weather-and-biking query where GPT-5.6 Sol gives a two-sentence verdict ("yes, you'll stay dry, but bring a wind layer"), versus GPT-5.5 Instant's multi-paragraph forecast that buried the answer.

Second, better factual accuracy. In an internal evaluation across finance, medicine, and law prompts, responses containing at least one factual error dropped by roughly 62 percent for GPT-5.6 Luna and 68 percent for GPT-5.6 Sol compared with GPT-5.5 Instant. These are OpenAI's own numbers — no independent verification yet — but the magnitude is substantial if it holds in real-world use.

Third, a reasoning slider. Paying users can now choose how much "thought" ChatGPT puts into a response. Lower settings handle everyday questions; higher settings are meant for research, coding, and multi-step planning. The slider spans five levels and aims to make quick answers and deeper reasoning feel like one consistent model rather than two with different personalities. Notably, this does not apply to ChatGPT Work or Codex — the Sol version powering those agentic products remains unchanged.

What free users lose — and gain

The bigger structural change is what happens on the free tier. According to release notes, GPT-5.6 Luna will become the default — and only — model for ChatGPT Free and ChatGPT Go users later this week. Unlimited text chats arrive next week, along with a new "Think" button that lets Luna spend more time reasoning through harder questions. File upload, image generation, and other tool limits stay in place.

This is a clear tier stratification. Luna is OpenAI's smaller, cheaper model. It runs faster and costs the company less to serve — but smaller models are also measurably more error-prone. The Decoder flagged a relevant precedent: an earlier Instant model produced a completely wrong data analysis compared with its larger reasoning counterpart. The Think button only gives Luna more inference time — it does not switch you to Sol's reasoning capacity.

For free users, the practical takeaway is this: straightforward questions, fact-checks, translations, and simple summaries will work fine. But anything requiring careful multi-step reasoning — contract analysis, debugging complex code, legal or medical cross-referencing — now runs on a model that OpenAI's own data says is less factually reliable than the one paying users get.

Pricing and availability in Europe

All updates are available to users in the European Union and European Economic Area. ChatGPT remains accessible across the EU without a VPN requirement, and all paid plans are bookable with European payment methods.

The current pricing breakdown for European users:

For context: a European developer or startup that needs reliable factual responses for client work effectively needs at least the Plus tier. At €18.40/month, it's priced similarly to a mid-range SaaS tool — and the jump from Go (€5.50) to Plus (€18.40) now buys you not just more capacity, but a fundamentally more accurate model.

From AI Arena to production: why model tiering matters

We operate AI Arena, our own benchmarking rig, and run production AI pipelines on our own servers — so we have a practical read on what "cheaper model" means operationally. When you downgrade from a large reasoning model to a smaller distilled variant, you typically see 15–30% drops on benchmarks like MMLU-Pro and GPQA, but the bigger gap shows up in long-chain reasoning tasks where one wrong intermediate step cascades into a completely wrong final answer. That's the risk with the Luna-only free tier.

OpenAI's internal numbers bear this out: Luna's 62% error reduction is real progress over GPT-5.5 Instant, but Sol's 68% is better — and the absolute error rate matters more than the relative improvement. If GPT-5.5 Instant hallucinated in, say, 8% of factual prompts, Luna drops that to roughly 3%, and Sol to about 2.6%. The gap looks small on paper but compounds when you run hundreds of queries per day.

The reasoning slider is also worth a skeptical look. OpenAI staffer Vaibhav Srivastav previously outlined that higher levels burn more tokens and take more time — the recommendation is to start low and scale up only when needed. In practice, this shifts the optimization problem onto the user: you're now expected to guess the right reasoning depth for each task. OpenAI's own history with the model switcher, which most users ignored, suggests that interface complexity rarely translates to better real-world outcomes.

What this means for the competitive landscape

This update lands in a market where Anthropic's Claude Opus 5 is pushing prompt-to-game generation, Google's Gemini models are integrated across Workspace, and open-weight models from Mistral and Meta keep improving. OpenAI's move to stratify free versus paid model quality mirrors what Google did earlier with Gemini — but the magnitude here is sharper: free users don't just get rate-limited, they get a structurally different, less capable model.

For European companies weighing AI providers, the calculation now includes a new variable: if you're building internal tools that rely on ChatGPT's free or Go tier, you're now running on Luna, not Sol. That's a downgrade worth testing against your actual workloads before committing to any architecture that depends on it.