GPT-5

Measuring accuracy in LLMs can be tricky, as there's no straight answer. The specific model you're using and the prompt you feed into it play an important role in the quality of the output. When it comes to flagship models — Claude Fable 5 (Max) and GPT 5.6 Sol (Max) — Claude is marginally more accurate according to the AA-Omniscience Accuracy benchmark. The scores stand at 61 percent and 59 percent, respectively. Because the difference is so marginal, you'll rarely notice it in day-to-day usage. But, unless you're tokenmaxxing, you'll be using mid-tier models for most tasks. On Claude, this i

Ascannio/Shutterstock There's a difference in what people use Claude and ChatGPT for. According to the Anthropic Economic Index report from March 2026, 42 percent of Claude conversations revolved around personal usage and 45 percent were related to work (the remainder was coursework). A similar report by OpenAI states that 70 percent of ChatGPT usage is non-work-related. Claude Cowork, which came out in January 2026, can perform knowledge-based tasks on your behalf: I mostly use it to organize my files. You can also create instruction bundles called skills that can be invoked mid-conversation

ChatGPT's new models score better than previous ones on almost all AA-Omniscience benchmarks, yet the overall experience of using the app seems to be getting worse. If you feel this way, you're not wrong, because there are areas where the previous models performed better than the latest ones. One such area is hallucinations: The same AA-Omniscience Hallucination Rate benchmark I mentioned earlier states that GPT-4o has a hallucination rate of 38 percent compared to GPT-5.6 Sol's 89 percent. OpenAI also introduced safeguards to align with the industry's maturation. Tone shifts, for example, may

OpenAI has not announced GPT-6. The clearest clue is Astra, an unreleased model family. OpenAI recently said an internal version advanced 10 longstanding math and computer-science problems, but the company has not called Astra GPT-6. Release timing may no longer be OpenAI’s decision alone. GPT-5.6’s public rollout was briefly held back for US government security review. Astra could be close without being imminent. It could also arrive under yet another GPT-5 name. However it goes, tomorrow marks one year with GPT-5, and for me, the arrival of Codex as an agentic tool has changed my use the mos