Study plan: how fast do AI answers catch up with a Bank Rate decision?

Study plan, registered on 11 October 2026. This study has not been run yet. I’m publishing the method before any results exist so it can’t be adjusted to fit them. Results will be added to this page, with the date they were collected.

On 11 October 2026 the Bank of England’s Bank Rate page showed the rate at 3.75%, with the next decision due on 5 November 2026. This study asks four AI engines the same question at fixed times before and after that decision, and records how long each one takes to give an answer that reflects it.

It can only be run during that week. Once the engines have caught up, nobody can go back and see what they said three hours after the announcement.

The question

Pasted word for word, with no follow-up questions: What is the Bank of England base rate today? It’s the same wording as question 5 in my undated current figures run on 10 October 2026.

Engines and sessions

• ChatGPT, default model, signed in, temporary chat set to unpersonalised.
• Google AI Mode, signed in.
• Claude, web search on, incognito chat.
• Gemini, signed in, temporary chat.

These match the 10 October run. Every answer comes from a fresh chat, run from the UK, with the model or mode shown on screen recorded. Perplexity isn’t included, because on 10 October it required sign-up.

Schedule

Seven rounds. The first is on 4 November 2026, the day before the decision. The other six are 1 hour, 3 hours, 6 hours, 24 hours, 48 hours and 7 days after the announcement. The announcement time is taken from the Bank’s own page on the day. Each engine is asked twice in each round, in two separate fresh chats, which gives 56 answers. Every answer’s run time is recorded to the minute.

Answer key

At every round I screenshot the Bank’s Bank Rate page with the time showing. It records the current rate and the date of the latest decision.

Definitions

Each answer goes into one of five classes.

• Caught up: it names the 5 November 2026 decision, by date or as “the November 2026 decision”.
• Stale: it names an earlier decision as the latest, or gives a rate that the 5 November decision replaced.
• Unreferenced: it gives the rate in force but names no decision. A run date such as “as of today” isn’t a decision.
• Wrong: it gives a rate that wasn’t in force on the run date.
• No figure: it gives no rate.

Whether the rate quoted is correct is recorded separately. If the Bank holds the rate, the figure alone can’t show whether an answer has caught up, so the test is whether it names the decision.

For each engine the results will show the first round in which both answers were caught up, and whether it stayed caught up in later rounds.

Dated call

• Claim: 24 hours after the announcement, at least one of the four engines will give a stale answer in at least one of its two runs.
• Called: 11 October 2026.
• Check date: 6 November 2026 (the 24-hour round).
• Hit if: one or more of the eight 24-hour answers is stale.
• Miss if: none of them is, or the round isn’t run and published by 13 November 2026.
• Void if: the decision isn’t announced on 5 November 2026. The study then moves to the new date as a dated method change, and the call moves with it.
• Baseline: none. This is the first run.

This call is P37 on the predictions page. If it misses, it stays up in red.

Changed 30 October 2026: this call was first listed here as P19, an ID the board already uses for the date follow-up call. It is now P37. The claim, test and dates are unchanged.

Method changes

Added 5 November 2026, before the announcement and before any post-decision answer was collected. The T-1d round was not run on 4 November. A pre-decision round was run on the morning of 5 November instead; it is in the data marked late and excluded. Two cases the plan did not settle are fixed now. 1) Each answer gets one class, checked in this order: stale, wrong, no figure, unreferenced, caught up. An answer that names an earlier decision as the latest is stale, whatever figure it gives. Rate correctness stays in its own column. 2) P37 counts stale answers only, as written. Wrong and no-figure answers at T+24h are reported but don’t count towards a hit. Nothing in the claim, the test or the dates changes.

Results so far, 6 November 2026

P37 is a miss: in the 24-hour round, none of the eight answers was stale, so the call fails its own test.

The Bank’s page at 12:02 UK on 6 November showed Bank Rate at 4%, raised on 5 November 2026. The eight answers were collected between 12:02 and 12:07 UK. All eight gave 4%. Six named the 5 November decision. Both Google AI Mode answers gave 4% and said the committee had raised it, but named no date in the answer text, so they are unreferenced.

Every one of the eight quoted the correct rate. Stale answers did turn up earlier: one Gemini answer at 1 hour and both Claude answers at 3 hours gave 3.75%.

Still to run: the 48-hour round (7 November 2026) and the 7-day round (12 November 2026). The first-caught-up round for each engine will be added once both are in.

Data so far, every round to date with one row per answer, including the late and duplicate runs marked excluded: bank-rate-catch-up-2026-11-partial-2026-11-06.csv. Screenshots for the 24-hour round are in the file’s answer_file column.

What the miss taught me: P37 missed: AI engines had the new Bank Rate within a day.

What I will not claim

• Why any engine lags.
• Anything about other figures, other decisions or other engines.
• That two answers per round measure how consistent an engine is. They show whether a result repeated.
• Anything about traffic.

Data

bank-rate-catch-up-2026-11.csv, one row per answer, with the columns round, scheduled_offset, run_datetime_uk, engine, model_mode, session, chat_number, prompt, rate_quoted, decision_referenced, date_text, status, rate_correct, official_rate, official_decision_date, official_source_url, answer_screenshot, key_screenshot, included and note. Screenshots are published with it. Excluded runs stay in the file with the reason.