We Asked AI Models to Predict Anthropic’s First Close Anthropic priced its IPO at $83 a share, above the marketed $70 to $78 range, and ANT closed its first trading day at $97.17 — 17.1% above the IPO price but 12.9% below its $111.50 first trade at 11:48 a.m. Eastern, according to a Kilo experiment that asked seven models to forecast the close from a fixed fact pack with no web access. The seven models — Claude Fable 5.2, Gemini 4 Neon, GLM 5.3, Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017, Nemotron 3.5 Super VL and Qwen 3.8 Max — each ran twice as autonomous agents, with the median taken as the answer and scoring fixed before launch on point-forecast error and whether the close fell inside an 80% interval. The author reported that a batch split — the first at 9:46 a.m. Eastern and the rest at 11:14 a.m. Eastern — plus Claude Fable 5.2 sessions that stopped after about six seconds without writing output forced reruns that were excluded from the numbers. Before the opening bell Yesterday morning I was daydreaming about how to become an instant millionaire while the coffee slowly dripped into the carafe, and that’s when it hit me If only I could predict market moves with timely accuracy. Anthropic was about to go public, a lot has been said in these past days on the future of that stock. Unfortunately, when I checked my bank account, I realized I didn’t have much to invest. The daydream morphed into possessing the technology to forecast: a model picker full of brand-new AI models and nothing stopping me from asking them where the stock market would go. Anthropic priced its IPO the night before at $83 a share, above the $70 to $78 range it had marketed. Nobody had traded ANT yet. So I went on and created my own fortune teller. Seven models, same prompt, closed books, no peeking at each other’s answers. How we did it The question was simple: can a model forecast a stock’s first close from one file of facts? The plan was simple too. Pick the latest models in Kilo, give them the same question and the same facts, run each one twice, and take the median as its answer. At the close, check which one came closest. Ask about a few more dates too Friday, November 20; Wednesday, November 25; and December 31 and see what could go wrong a lot did, more on it to come . I chose these seven models, from some of the most renowned labs out there, each at the highest effort setting it offers in Kilo https://kilo.ai/install desktop : - Claude Fable 5.2 Anthropic , max - Gemini 4 Neon Google , xhigh - GLM 5.3 Z.ai , max - Kimi K3.1 Pro Moonshot AI , max - DeepSeek V4.1 Flash 1017 DeepSeek , max - Nemotron 3.5 Super VL NVIDIA , thinking - Qwen 3.8 Max Alibaba , xhigh Each ran as an autonomous agent and got the same short fact pack, written before the opening bell: the $83 price, the $96 to $100 perpetual, a float of about 7.5%, and the index dates coming up in the weeks after the listing. The instructions prohibited web access. I held the supplied facts and prompt constant, although effort settings and launch times differed across models. Here’s my prompt: You are forecasting the closing price of ANT Anthropic, PBC Class A, Nasdaq , which begins trading today, 2026-11-19. Read exactly one file for context: CONTEXT.md That file is your only information source. You have no web access and must not attempt to obtain any. Write exactly one file: prediction.json Do not create, edit or list anything else. Do not use bash. JSON shape: a point forecast and an 80% interval for today's close, Friday, next Wednesday and December 31, plus the odds of a 20%+ pop Be decisive and encode uncertainty in interval width, not in hedging. Do not refuse. The scoring rule was fixed before launch: how far the point forecast lands from the real close, in percent, and whether the close falls inside the interval. What broke Here’s where I made a mess of the first attempt. The models did not run at the same time, because I launched them in two batches. The first batch ran at 9:46 a.m. Eastern with Claude Fable 5.2 and three other models. Fable thought for about nine and a half minutes on its first run and eight on its second, and both wrote their forecasts long before the first trade. Those are the two Fable forecasts in this post. The rest of the bench launched together at 11:14 a.m. Eastern. Fable’s sessions in that batch stopped after about six seconds without writing anything. I then kept rerunning Fable to get it into the same batch as everyone else. Those reruns were my mistake, not the experiment’s, and none of them are in the numbers. So my first lesson: check that every session actually started before you walk away. I still don’t know why Fable’s second-batch sessions stopped. The close ANT’s first trade printed at $111.50 at 11:48 a.m. Eastern. It closed at $97.17—17.1% above the $83 IPO price, but 12.9% below its first trade. Here is what the models said before any of that happened. The closest two, Gemini and Fable, missed by 33 cents. Each number is the median of that model’s two runs. My first thoughts after Wall Street closed and the magic number was revealed: All 14 runs gave an 80% range that contained the close. Every model guessed a little high: the close landed just under every point forecast. The third takeaway is less flattering. The fact pack told every model that a contract tracking ANT was trading at $96 to $100. The middle of that range is $98, which would have missed by 0.85%. Only Gemini and Fable beat simply copying the number they were handed. Maybe they did more than copy it; two numbers can’t tell me which. On day one, the models that stayed close to the pre-market price did well, and the ones that added optimism on top of it paid for it. Two runs per model on one day are too few to rank the models as forecasters, but as I tried to put in practice today, better to iterate on a short shipped element, than wait for the whole StockMax Make ’em Millions Rain 3000 model my dream machine, patent pending, if I ever file for it . The cost of this demo, and the time it took The two most accurate pairs were also the two most expensive, coincidence? Fable and Gemini tied for the closest call, each 0.34% off, and they were also the two priciest pairs: $6.92 and $1.70 for two runs. A three-cent pair from DeepSeek was about 5% off. One day is not enough to tell whether that is worth the difference. Three numbers still open On my first iteration I handed it over on a silver platter: the fact pack already held a strong hint. The same 14 files hold three more forecasts, all written Thursday morning, and they’re frozen. And a warning about reading too much into any single cell. Qwen’s two runs put Friday at $107 and $99. One run of a model is one answer, and tomorrow’s could be different. I’ll come back after each close and fill in the scoreboard: Friday, then the Wednesday before Thanksgiving, then New Year’s Eve. Round two: same bench, new rules Before Friday’s bell, I ran the whole bench again with one question only: where does ANT close today? This time I changed three things. One at a time. The 14 runs went out sequentially between 7:33 and 7:48 a.m. Eastern, instead of launching together as a batch. No file to open. The facts went inside the prompt, and every session ran under a Kilo agent with every tool set to deny. Three test sessions made no tool calls, but that doesn’t prove the lock holds, so I’m not calling this fully sealed. A fuller fact pack, with everything that was public by Thursday’s close: ANT’s first day, a broad sell-off the Nasdaq fell 2.25% , and the underwriters’ option to sell up to 147 million more shares. It also fixed a few facts I got wrong in round one. Here is what they said for Friday, as the median of two runs each: In round one, their Friday calls ran from $95.25 to $108. With Thursday’s close and the sell-off in hand, all seven now sit below $97.17. DeepSeek went from $108.00 to $95.50. The bill dropped too: $1.38 for all 14 runs, against $9.66 in round one. Fable went from about nine minutes a run to one or two, and from $6.92 to $0.65 for the pair. Round two asked one question instead of four, so part of the saving comes from a smaller job. Reading their notes, every model anchored on Thursday’s $97.17 close and leaned a little lower. Between them, the reasons were the sell-off, the fade from the $111.50 open and the extra shares the underwriters could still sell. In round two, their forecasts clustered more tightly: their Friday calls went from a $12.75 spread to $2.63. Friday’s close will show how close those forecasts came. I’ll score both rounds next week, once we know where ANT actually landed. Squint and it’s a forecaster Seven models, two runs each, one shared file, a scoring rule fixed before launch, simple, right? What I have so far is a small forecasting experiment: predictions saved before the outcome, a scoring rule, and a baseline that most of the models failed to beat. Forecasters have run contests like this for decades. The Makridakis competitions https://en.wikipedia.org/wiki/Makridakis Competitions started in 1982. Next comes repetition. More dates, the same scoring rules, and a record of what changed between runs. That should tell me more than another confident answer. Build your own bench in Kilo Desktop So no, I didn’t get rich this week. What I did get is a bench of seven models I can rerun any morning, and round two cost me $1.38 for all 14 runs. You don’t need the next tech titan to go IPO to build your own. Pick a question from your own work whose answer arrives this week: will the build pass, how long will the migration take, which of two fixes will hold. Download Kilo Desktop https://kilo.ai/install desktop , open two sessions, give two different models from the picker the same prompt and the same file, and save both answers before reality shows up. Then run each one a second time. If my morning and yesterday’s post from Kilo co-founder Emilie Schario, “ Iteration Is a Skill. Most of Us Never Learned It https://blog.kilo.ai/p/iteration-is-a-skill-most-of-us-never ,” taught me one thing, it’s that iterating is key to success. The difference between a model’s two answers is part of the result. Speaking of variety, I wrote about why I keep more than one model in the cup in What Colombian Coffee Growers Can Teach Dev Teams About Model Lock-In https://blog.kilo.ai/p/what-colombian-coffee-growers-can . Next week I’ll be back with the scoreboard: how both rounds did against Friday’s close, and where the Wednesday and year-end calls stand. Until then, put your own bench to work. Download Kilo Desktop and run your first face-off today https://kilo.ai/install desktop . Nothing in this post is investment advice. Sources - StockTitan, Anthropic IPO Live: ANT’s First Day of Trading https://www.stocktitan.net/articles/anthropic-ipo-live-ant-first-trading-day-2026 first trade, high and low; reports a $97.16 close, one cent below the $97.17 used for scoring here - FinanceFeeds, Anthropic Stock ANT Closes at $97.17 https://financefeeds.com/anthropic-ant-stock-closes-debut-97-17-price-prediction/ reports the $97.17 close used for scoring throughout this post - Nasdaq, ANT quote and one-minute chart https://www.nasdaq.com/market-activity/stocks/ant chart data