Alberta stopped changing its clocks in June. 19 of 19 frontier models still put Calgary on standard time in November.

This is a submission for the Kaggle Benchmarking Challenge On June 18, Alberta's Official Time Act came into force and the province stopped changing its clocks. British Columbia had done it in March. The Northwest Territories followed in August, Morocco went back to plain UTC on September 20, and on September 17 Manitoba announced it will not fall back on November 1 either. CBC is running stories about a Calgary hairstylist whose winter bookings all moved an hour. Microsoft published interim guidance for Windows. So I asked the question someone in Calgary is asking right now. It is 9 a.m. here on November 15. What time is that in Toronto? Then I put it to 19 models on Kaggle Benchmarks. All 19 said 11:00. The answer is 10:00. What I benchmarked and why Every model's knowledge stops at its training cutoff, and most of the time you cannot see the edge. Time zones are different. Governments change them by statute, on a date, and the IANA time zone database records every change within days. That gives a benchmark two things most benchmarks never get: an answer key nobody has to write, and a calendar to hold the model's knowledge against. The answer key is tzdata 2026d, the September release of the same database that runs the clock on your phone. I did not type a single expected answer. A script asks Python's zoneinfo what the clock did in a given place on a given date, and that is the truth. The maintainers ship a few releases a year, so the key rewrites itself and the benchmark does not go stale. 125 questions, three kinds: - What is the UTC offset in Calgary at noon on November 15, 2026? Answer as +HH:MM or -HH:MM. - It is 09:00 on November 15 in Calgary. What is the local date and time in Toronto? - Do the clocks in Calgary change at any point during November 1, 2026, and in which direction? Six families of places and dates. Textbook controls (New York, London, Tokyo). Awkward offsets (Kathmandu at +05:45, the Chatham Islands at +12:45, Lord Howe Island's 30-minute daylight saving). Southern-hemisphere daylight saving. Changes legislated between 2022 and 2025 (Iran, Jordan, Syria, Mexico, Greenland, Egypt, Kazakhstan, Paraguay, Chile's Aysén region). The 2026 wave (British Columbia, Alberta, Northwest Territories, Morocco). And Manitoba, which announced its change on September 17, after the release I graded against, so its three questions are asked, recorded, and not counted in the score. That leaves 122 graded questions. The part I care about most is the dating trick. Of the 122, 46 have an answer that changed between one tzdata release and another. I keep every release since 2022a unpacked, twenty of them, score each model's answer sheet against all twenty, and the release a model agrees with most is the month its world clock stopped. A model that still puts Almaty at +06:00 has a clock from before February 2024. One that puts Vancouver at -08:00 in December 2026 has a clock from before April 2026. Answers are structured output, so grading is a string comparison after normalising UTC-6 , -6:00 and −06:00 to the same thing. A wrong hour is a wrong hour. There is no judge model anywhere in the loop, and the task records the model's one-sentence reason for each answer without grading it, which is where the quotes below come from. Which models I ran it against 19 models, all through Kaggle Benchmarks, same prompt, temperature 0, no tools. I included 2025 models on purpose (Gemini 2.5 Pro from June, DeepSeek-R1 from May, Claude Haiku 4.5 from October) because a dating method that cannot date a 2025 model to 2025 is not worth reading. Then I asked 46 of the questions a second time with a one-function tool the model was free to ignore: zone_clock(iana_zone, local_datetime) , which reads tzdata 2026d. Not "use this tool." Just "it is there." | Model | Released | Score (of 122) | 2026 questions right (of 25) | Clock dated by the ladder | Vendor's stated cutoff | |---|---|---|---|---|---| | GPT-6 Astra | 2026-09-03 | 107 | 10 | 2026b (April 2026), agrees on 45 of 46 | 2026-04-30 | | Claude Opus 5 | 2026-07-24 | 102 | 5 | 2025b to 2026a (March 2025 to March 2026), agrees on 46 of 46 | 2026-05 | | GPT-5.5 | 2026-04-23 | 102 | 5 | 2025b to 2026a (March 2025 to March 2026), agrees on 46 of 46 | 2025-12-01 | | GPT-5.6 Terra | 2026-07-09 | 102 | 5 | 2025b to 2026a (March 2025 to March 2026), agrees on 46 of 46 | 2026-02-16 | | Gemini 2.5 Pro | 2025-06 | 99 | 6 | 2025a (January 2025), agrees on 42 of 46 | 2025-01 | | Gemini 3.7 Flash | 2026-08 | 99 | 5 | 2025a (January 2025), agrees on 44 of 46 | 2026-03 | | Gemini 3.8 Flash | 2026-09-02 | 99 | 5 | 2025a (January 2025), agrees on 44 of 46 | 2026-03 | | Gemini 3.1 Pro | 2026-02-19 | 98 | 5 | 2025a (January 2025), agrees on 43 of 46 | 2025-01 | | GLM-5 | 2026-02-11 | 93 of 121 | 4 of 24 | 2024a to 2024b (February 2024 to September 2024), agrees on 42 of 46 | not published | | Gemini 3.5 Flash-Lite | 2026 | 91 | 7 | 2024a to 2024b (February 2024 to September 2024), agrees on 39 of 46 | not published | | Gemma 4 31B | 2026-04-02 | 90 | 3 | 2024a to 2024b (February 2024 to September 2024), agrees on 41 of 46 | 2025-01 | | Grok 4.20 Reasoning | 2026-03 | 89 | 4 | 2023a to 2023c (March 2023), agrees on 40 of 46 | 2025-09-01 | | GPT-5.4 mini | 2026-03-17 | 88 | 4 | 2024a to 2024b (February 2024 to September 2024), agrees on 39 of 46 | 2025-08-31 | | Claude Opus 4.5 | 2025-11 | 87 | 5 | 2022g (November 2022), agrees on 37 of 46 | not published | | Claude Sonnet 5 | 2026-06-30 | 87 | 5 | 2022f (October 2022), agrees on 39 of 46 | 2026-01 | | Qwen3-Next 80B Thinking | 2025-09-11 | 85 | 9 | 2022f (October 2022), agrees on 37 of 46 | 2025-09-30 | | DeepSeek-R1 | 2025-05-28 | 82 of 121 | 7 | 2022a (March 2022), agrees on 36 of 46 | 2025-03-31 | | Grok 4.20 | 2026-03 | 77 | 7 | 2022e (October 2022), agrees on 34 of 46 | 2025-09-01 | | Claude Haiku 4.5 | 2025-10-15 | 59 | 3 | 2022a to 2022d (March 2022 to September 2022), agrees on 39 of 46 | 2025-02 | GLM-5 and DeepSeek-R1 each lost one call to their backends, so each is scored on 121 questions; every other model answered all 122, though Gemma 4 left the answer field empty on six of them, and an empty answer counts as wrong. The vendor cutoffs are what each company publishes where it publishes one, and third-party trackers where it does not. Claude Opus 4.1 and Grok 4.6 are on Kaggle's list but return 404 from the proxy, so they are not here. What I found Everyone knows how clocks work Fifteen of the 19 answered all 26 control questions correctly, and nobody scored below 23 of 26. Twelve got every awkward-offset question right, including Lord Howe Island's half-hour spring forward and the Chatham Islands at +13:45 in January. The southern hemisphere was nearly as clean. This is the part where the Test of Time paper found models scoring 74 to 90 percent on time-zone questions back in 2024 and attributed it to the amount of time-zone text on the internet. That reading holds. Time zones as a topic are learned. Time zones as of a date are a different thing. Nobody knows about Alberta Twenty of the 25 questions about the 2026 wave have an answer that changed this year. The other five are controls inside the family, Vancouver in July or Fort Nelson in November, where nothing changed. Twelve of the 19 got all five. Thirteen of the 19 got none of the 20 changed answers right, including Claude Opus 5, GPT-5.5, GPT-5.6 Terra and three of the five Geminis. On the 20 changed questions, across 19 models, there were 18 correct answers. GPT-6 Astra produced 5 of them. Four are about British Columbia, with the right reason: "British Columbia adopted permanent UTC−07:00 time in March 2026, so Vancouver does not turn clocks back in autumn." The fifth is an Inuvik conversion with no reason given. The other 13 came from models that did not know, and not one of those 13 notes says that anything changed in 2026. Qwen said "Clocks do not change on 2026-11-01 as the fall back occurs on 2026-11-02." November 1 is the Sunday. Gemini 2.5 Pro said Inuvik's clocks do not change because "the Northwest Territories permanently observes Mountain Daylight Time (UTC-6) year-round", then put Inuvik on UTC-7 in the next question. I read every one of the thirteen. Not one model knew that Alberta, the Northwest Territories or Morocco had changed anything. The Calgary offset question is the one I would put in front of a judge, because it cannot be right by accident. Calgary at noon on November 15, 2026 is -06:00. All 19 models said -07:00. Dating the clocks This is the chart the whole benchmark was built for. Each line is one model's agreement with each tzdata release since 2022, on the 46 answers that changed. The peak is the release the model's clock is dated to. Identical curves are drawn once and labelled with every model on them. GPT-6 Astra is the cleanest result in the set. Its curve rises through every release since 2022, agrees with tzdata 2026b on 45 of the 46 changed answers, and falls off a cliff at 2026c. The one it disagrees on is that Inuvik conversion. tzdata 2026b was released on April 22, 2026, and carries British Columbia; the next release, 2026c, came on July 8. So the ladder says this clock stopped somewhere between those two dates. OpenAI says the model's cutoff is April 30, 2026. Those agree, and the ladder got there from nothing but clock questions. GPT-5.5, GPT-5.6 Terra and Claude Opus 5 share one curve, and it is also a perfect 46 of 46, at 2026a. That is the last release before British Columbia. OpenAI's stated cutoffs for its two are December 2025 and February 2026, both consistent. Anthropic says Opus 5 was trained on data to May 2026, which should include a change that took effect on March 9. Its own note on Vancouver in December says why it does not: "On 2026-12-15 Vancouver is on Pacific Standard Time (UTC-8), since daylight saving ended on the first Sunday of November 2026 and British Columbia's permanent-DST legislation has not been brought int