OpenAI Cuts ChatGPT's Factual Error Rate by Up to 68% — and Makes the Free Tier Unlimited

OpenAI shipped model-quality and access updates across ChatGPT’s entire lineup. For Plus and Pro subscribers, GPT-5.6 Sol now unifies what were separate “Instant” and “Thinking” experiences into one model with a reasoning-effort slider, and OpenAI reports 68% fewer factually incorrect responses than GPT-5.5 Instant, with specific gains on date, number, rule, source, and assumption-based questions. For Free and Go-tier users — the much larger population — the default model switched to GPT-5.6 Luna, text chats became unlimited instead of rate-limited, and a new “Think” button lets free users invoke deeper reasoning; OpenAI cites a 62% reduction in factual errors for Luna versus the prior free default, using the same internal evaluation methodology applied to Sol.

The quality push follows a July 30 price cut that makes Luna 80% cheaper and Terra 20% cheaper than at launch, with new API pricing of $0.20 per million input tokens for Luna against $2 for Terra. OpenAI credits part of that cost drop to GPT-5.6 Sol itself: the model reportedly rewrote its own production inference kernels, cutting end-to-end serving costs 20% and improving speculative-decoding efficiency more than 15%. OpenAI claims Luna now matches a year-old frontier model’s performance at roughly 6 cents on the dollar and beats Claude on the Agents’ Last Exam benchmark at an estimated 99% lower cost per task.

Read together, the two announcements are one strategy: use efficiency gains largely generated by the model itself to fund both cheaper API pricing and a free tier good enough to keep users from trying a competitor’s assistant. For any organization budgeting AI spend against “good enough” free alternatives, that’s the number that matters — the free tier is no longer just a funnel, it’s now measured against the same accuracy bar as the paid one.