Recommended Free Tools
OpenAI says its October 2026 GPT-6 Sol and GPT-6 Luna models made factuality gains over GPT-5.6 on nearly all of the company’s selected evaluations. That is progress, not proof that ChatGPT has stopped making things up: the tests were deliberately difficult and do not measure how often errors occur in ordinary conversations.
What OpenAI says changed in GPT-6
In its October 7, 2026 system card, OpenAI reported significant factuality improvement for GPT-6 Sol and GPT-6 Luna over GPT-5.6 Sol and GPT-5.6 Luna on nearly all of its evaluation metrics. The stated exception was claim-level factual errors on factuality-heavy prompts. The result is a company-reported comparison on selected tests, not an independently established error rate for everyday ChatGPT use.
As an Amazon Associate I earn from qualifying purchases.
The system card said GPT-6 Sol and Luna launched globally across free and paid ChatGPT plans, replacing GPT-5.6 Sol and Luna in ChatGPT. It distinguishes those ChatGPT versions from model versions then used in Codex and ChatGPT Work, so the claim should not be generalized to every OpenAI product.
What the factuality tests measure
OpenAI used three kinds of prompts: factuality-heavy examples intended to represent factual ChatGPT conversations, de-identified conversations users had previously flagged for factual errors, and difficult prompts involving medicine, law, and finance. An LLM-based grader with web access assessed factual errors.
#1 Best Overall
The card reports both errors at the claim level and the share of responses containing at least one error. Those measures answer different questions: a response can contain several incorrect claims, while the response-level measure counts whether it contains any error. The published caveat matters as much as the comparison: OpenAI says these evaluations were designed to expose failure modes and do not estimate errors’ overall prevalence in production or the average user experience.
Why ChatGPT can still make things up
OpenAI defines hallucinations as plausible but false statements generated by language models. In its September 5, 2025 explainer, the company argues that accuracy-focused tests can create an incentive to guess. If a benchmark marks an uncertain answer wrong but gives credit to a lucky guess, a model may be rewarded for answering rather than admitting uncertainty.
Rank #2
That is one explanation, not a complete account of every error. OpenAI notes that arbitrary or infrequent facts can be difficult to infer from patterns in training data, that information needed to answer some questions may be unavailable, and that ambiguous prompts can make a reliable answer impossible. Training and evaluation that handle uncertainty better may help, but they cannot guarantee that a model will never produce false claims.
A September 2025 paper by Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang makes a related argument: training and evaluation procedures can reward guessing over acknowledging uncertainty, and hallucinations have statistical causes in the modern training pipeline. This is the authors’ analysis; it does not establish evaluation incentives as the sole cause.
Why abstaining can be better than guessing
“I don’t know” is sometimes the more useful answer. OpenAI’s September 5, 2025 SimpleQA example illustrates the trade-off for two older models; it is not a GPT-6 comparison:
| Model and source date | Abstention | Accuracy | Error |
|---|---|---|---|
| gpt-5-thinking-mini, OpenAI, September 5, 2025 | 52% | 22% | 26% |
| o4-mini, OpenAI, September 5, 2025 | 1% | 24% | 75% |
In this example, o4-mini answered more often and had slightly higher accuracy, but also a much higher error figure. The comparison shows why accuracy alone can obscure the cost of confident mistakes. It does not establish how often current ChatGPT models abstain or err in real-world conversations.
Rank #4
What the results do—and do not—tell ChatGPT users
- They support a qualified progress claim: OpenAI reports improvement over GPT-5.6 on nearly all of its selected factuality measures for GPT-6 Sol and Luna.
- They do not show that hallucinations are solved: the card identifies an exception and says its challenging tests do not represent typical production prevalence.
- They are not an everyday error-rate estimate: the published evaluation design is targeted at difficult prompts and prior user-flagged errors, not a representative sample of all ordinary conversations.
For context, OpenAI’s earlier GPT-5 System Card reported that GPT-5 main had a hallucination rate 26% smaller than GPT-4o, while GPT-5 thinking was 65% smaller than OpenAI o3 on the described production-representative evaluation. Those are earlier comparisons, not GPT-6 results or a measure of overall production prevalence. OpenAI said its web-enabled LLM grader achieved 75% agreement with human judgments when validating factuality assessments.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsThe practical reading is straightforward: OpenAI’s published evidence indicates progress on difficult factuality evaluations, while leaving the everyday frequency of errors unquantified. Treat consequential claims—especially medical, legal, and financial ones—as claims to verify rather than as facts guaranteed by a fluent answer.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
