# How Elon Musk Had Predicted That GPT-6 Would Be “Indistinguishable” From The Smartest Humans All The Way Back In 2020 > Source: > Published: 2026-07-23 18:25:19+00:00 Elon Musk hadn’t only founded OpenAI, but he also had his pulse on the pace of AI long before AI became mainstream. In August 2020, months after OpenAI had quietly released GPT-3 to a small group of testers, Musk posted a short note on Twitter that reads almost eerily today. He wrote that the jump in capability from the original GPT to GPT-3 was impressive, and that if the same rate of improvement held, GPT-5 or GPT-6 could end up indistinguishable from the smartest humans. He was careful to frame it as his own opinion rather than an endorsement of where OpenAI was headed, and he added a line that has aged in its own interesting way — that he’d left OpenAI two to three years earlier and was now a neutral outsider looking in. Musk had left OpenAI’s board in February 2018, reportedly over disagreements about the company’s direction and a potential conflict of interest with Tesla’s own AI ambitions. By the time he posted his GPT-3 prediction, [Sam Altman and Greg Brockman](https://officechai.com/stories/tense-emails-reveal-how-elon-musk-sam-altman-ilya-sutskever-and-greg-brockman-had-negotiated-over-openais-structure/) were running an organization that looked very different from the one Musk had co-founded and bankrolled in 2015. OpenAI had converted part of itself into a capped-profit entity the year before, taken a major investment from Microsoft, and was just weeks away from opening up API access to GPT-3 to outside developers. GPT-3 itself was the reason the tweet landed the way it did. The model had 175 billion parameters at a time when that number sounded almost absurd, and it was showing an ability to write short essays, answer questions and generate some code from little more than a prompt and a few examples. Researchers and journalists spent that summer swapping GPT-3 outputs that ranged from unnervingly coherent to unintentionally funny, and the model became the first real glimpse for a lot of people outside AI labs of what a large language model could actually do. That’s the backdrop against which Musk’s comment about GPT-5 or GPT-6 was written — he wasn’t speculating in a vacuum, he was extrapolating from a curve that had just made a very visible jump. What stands out looking back at the thread is how normal the exchange between Musk and Brockman feels underneath it. Brockman replied to the tweet with a simple thank you, and it picked up its own engagement separate from Musk’s original post. Musk had left OpenAI on terms that were reportedly tense behind closed doors, and the two men would go on to become adversaries in public in the years since, with Musk eventually [suing OpenAI](https://officechai.com/stories/elon-musk-is-a-bully-likes-to-pick-fights-openai-ceo-sam-altman/) over its shift away from a non-profit structure and building a rival lab in xAI. None of that rivalry was visible yet in August 2020. The reply reads like two people who still had some goodwill left over from having built something together, even if Musk was no longer part of it. The prediction itself is worth sitting with because of how specific it turned out to be. Musk didn’t say AI would keep getting better in some vague sense — he named GPT-5 or GPT-6 as the point where the gap to the smartest humans could close. OpenAI didn’t even use that exact naming convention at the time; the jump from GPT-3 to GPT-4 alone would take another two and a half years. But the underlying model family did keep climbing roughly the way he described, and now that OpenAI is several versions past GPT-5 with its GPT-5.6 line, the comparison to top human performance isn’t really a hypothetical anymore. The clearest evidence for that came out of the University of Pennsylvania, where a statistics professor used GPT-5.6 Sol Pro to disprove a long-standing assumption about the Benjamini-Hochberg method, a widely used tool for controlling false positives in statistical testing. The model worked through the problem in about 90 minutes. Its predecessor, GPT-5.5, had been given the same problem and failed to produce a valid solution even after twenty hours of compute. Around the same time, OpenAI researcher Ethan Knight [shared a proof](https://officechai.com/ai/openai-says-gpt-5-6-sol-produced-a-proof-of-the-50-year-old-cycle-double-cover-conjecture-using-64-subagents-in-1-hour/) generated by GPT-5.6 Sol Ultra of the Cycle Double Cover Conjecture, a graph theory problem that had sat unsolved since the 1970s, produced using 64 parallel subagents in under an hour. The proof still needs to go through peer review since it’s written in natural language rather than a formally verified system, but the fact that a publicly available model produced something worth reviewing at all tells you where the field has moved. None of this means Musk’s tweet was some kind of secret technical roadmap. It was an opinion posted on a Friday afternoon by someone who had spent years around the people building these systems and had a rough intuition for where the curve was pointing. But intuition backed by proximity to the work counts for something, and six years on, a model that closes long-standing open problems in statistics and graph theory in the time it takes to have lunch is a fairly direct answer to the question he was asking in that thread.