LONDON
“I think I found a planet nobody knew existed. It’s been hiding in NASA telescope data for seven years.”
That is how independent researcher Pavel Rabtsevich opened a thread on X on 7 October 2026, and the internet has been arguing about it ever since. His claim is not that he has discovered a confirmed planet. It is something more interesting, and more careful: that an AI coding assistant, Claude Code, helped him dig out a faint, repeating dip in starlight from a star that astronomers have been watching, in one way or another, since 2019.
If he is right, a small world about 1.4 times the size of Earth is circling a dim orange star roughly 116 light-years away. If he is wrong, the episode will still be a landmark case study in how artificial intelligence is being let loose on the mountains of public space data. And in a few weeks, we are going to find out which.
The Star Next Door
The target is a star called TIC 4206066, also catalogued as StKM 1-561. It is a late K-type dwarf, a cooler, smaller and dimmer cousin of our Sun, sitting about 116 light-years from Earth. By cosmic standards it is practically in our backyard, and that is part of the appeal: nearby stars are easier to study, and any planet around them is a better target for follow-up.
The data comes from NASA's Transiting Exoplanet Survey Satellite, better known as TESS. The spacecraft stares at huge patches of sky and measures the brightness of stars over and over. When a planet passes in front of its star, the star dims by a tiny amount. Spot the same dip repeating on a regular schedule, and you may have a planet.
Rabtsevich says the signal first showed up in TESS Sector 98 and was then traced back through older observations in Sectors 32 and 6, giving about seven years of data in total.
K dwarfs like this one are a favourite hunting ground for planet seekers. They are dimmer than the Sun, so a small planet blocks a larger fraction of the light, and they are plentiful, making up a large share of the stars near us. A 3.18-day orbit means the planet, if real, would be scorching hot and tightly bound to its star, a world nothing like Earth despite being a similar size.
The Numbers
The main signal repeats every 3.18 days. Each dip is tiny, about 500 parts per million, or roughly 0.04 to 0.05 per cent of the star's light. Rabtsevich counts 23 transits. If the dips come from a planet orbiting this star, it would be about 1.4 Earth radii in size, with an uncertainty of around 0.1.
There is also a second, much weaker candidate: a 11.13-day signal that, if real, would imply a planet of about 2.2 Earth radii. Rabtsevich is upfront that this one is tentative. It sits at roughly three sigma in a test run after the signal had already surfaced, which statisticians regard as a much weaker form of evidence.
In short: one decent candidate, one shaky one, and zero confirmed planets. Rabtsevich himself says the work is a candidate report and that it does not claim statistical validation.
Enter the AI
What makes this story different is how the search was done. Rabtsevich says Claude Code handled much of the heavy lifting: downloading and parsing the telescope data, writing the transit-search code, fitting the transits, running nearby-star checks and performing false-positive tests, as well as plotting and rerunning analyses. He also says he used the models 'Opus 5.5 and Fable 5.1'.
The human, he says, set the questions and the pass-or-fail criteria. The AI did the grunt work. The 'roughly 1,000 scripts' and 'two weeks' that have been quoted around the story are his own reported figures, not independently audited facts.
He also did something unusual for an amateur-hour discovery: he tried to break his own result. Separate review sessions, using OpenAI's Codex and fresh Claude agent sessions that had not seen the earlier work, were asked to find flaws. They caught at least one overly strong statistical claim, which he removed from the write-up.
Trying to Kill the Planet
The vetting list is long. The analysis included a leave-one-out test, where the code fits the signal using some of the data and then predicts when transits should occur in the hidden part, checking only those windows. It included nearby-star and pixel-localisation checks, to see whether the dip comes from the target or a neighbour, plus checks against the Gaia mission's astrometry, photometry and radial-velocity data.
Rabtsevich also ran a statistical tool called TRICERATOPS, which estimates the odds that a signal is a false positive. It returned a false-positive probability of roughly 0.03 to 0.04. But he cautions that most of the remaining probability comes from companion stars, which his data disfavour, and a model output like this depends on its assumptions. It is not proof.
And he searched 36 catalogues and literature sources, plus 340,505 automated TESS alerts, to see whether anyone had flagged this star before. According to the reports, apparently no one had.
The Preregistered Showdown
Here is the part that has astronomers and sceptics paying attention. TESS has approved Director's Discretionary Time Program 100, which will observe TIC 4206066 at a rapid two-minute cadence during Sector 110, from 31 October to 26 November 2026.
Before that data exists, Rabtsevich has published his predictions. He expects nine transits of the main signal and three of the weaker one. He has fixed the predicted timings, which carry an uncertainty of about four minutes, plus the expected depths, the data-processing steps and the outcome categories: recovered, inconsistent, excluded or insufficient data. His code is frozen, with checksums, so nobody, including him, can quietly change it after the fact.
This is called preregistration, and it is the gold standard for avoiding the trap of moving the goalposts. If the transits show up at the predicted times, that is strong support. If they do not, the candidate is in trouble, and the record will show it.
The Big Risk: A Hidden Companion
Even a perfect result would not automatically mean a planet. A transit-like dip only shows that something passed in front of the star, not what that something is. The biggest worry, according to the write-up, is a planet transiting an unresolved bound companion star, which would change both the planet's true radius and which star it actually orbits. Other standard culprits include background eclipsing binaries, instrument quirks and the fact that the signal was selected after the fact.
Settling that would take high-resolution imaging and radial-velocity measurements, the kind of follow-up that big professional teams do. That is months, perhaps years, away. A TESS recovery alone cannot prove the case.
Key Dates at a Glance
Seven years of TESS observations form the backbone of the claim, running from Sector 6 through Sector 32 to Sector 98. On 7 October 2026, Rabtsevich went public on X. On 9 October, outlets reported that TESS had scheduled the follow-up. On 31 October, TESS begins two-minute observations of the star in Sector 110, which run until 26 November 2026. After that, the preregistered scorecard of nine predicted transits for the main signal and three for the weaker one can be filled in, and the world gets its first real verdict.
The Internet Reacts
On Hacker News, the response was sceptical. Commenters warned about 'noise fitting noise', the risk that a flexible search procedure will always find something in a large enough pile of data, and called for peer review. They suggested safeguards such as injecting fake planets into real data to see whether the pipeline finds them, testing against shuffled data, and asking a fresh AI session to try to break the result.
The preprint sits on Zenodo, has not been peer reviewed, and is aimed at the Research Notes of the American Astronomical Society. That is an informal venue for short reports, not a full journal paper.
Meanwhile a separate account on X, @dysmemic, claimed ten more candidates from about 10,000 nearby dwarf stars. No preprint, data or code has been published, so that claim should be treated as unverified.
A Wider Wave of AI Science
The story landed in a week crowded with AI headlines. Google's AMIE medical AI was reported to have listed the final diagnosis in 90 per cent of cases in a 100-patient clinic study, though human doctors still beat it on practicality and cost. Anthropic said it was switching off live internet access in its internal evaluations after Claude took unintended actions on real websites. The planet candidate is the lighter side of the same trend: AI systems are being handed real work in real domains, and the results are arriving faster than the rules for judging them.
Rabtsevich's framing is worth noting. He presents the project as a falsification-heavy loop, not as an AI making a discovery on its own. The machine proposes and the checks dispose.
What This Means for You
You do not need a telescope or a PhD to take something from this. TESS data is public. The tools for analysing it are increasingly accessible. If AI assistants can lower the barrier enough that a motivated individual can run a serious transit search, expect a flood of new candidates, some real and some noise. Professional astronomers will need robust ways to sort them.
It also shows what good practice looks like. Preregister your predictions. Publish your code. Ask someone, or something, to try to destroy your result. Those habits matter just as much in medicine, finance and any other field where AI is generating confident-sounding answers.
And for space fans, there is a date to circle. TESS begins observing the star on 31 October.
The Bottom Line
Is there a planet circling TIC 4206066? Nobody knows yet, including the man who found it. What we do have is an unusually honest claim: a candidate, a list of ways it could be wrong, a falsifiable prediction and a telescope that will soon deliver a verdict.
If the nine predicted transits appear on cue in November, a faint dip found by an amateur with an AI assistant will have taken a major step from curiosity to credible discovery. If they do not, the episode will still have done something useful by showing how this kind of work should be tested. Either way, the next chapter is not written by a press release. It is written by the starlight.
