AI guessed the exchange rate, so every price came out high.
What the Reel says
AI guessed the exchange rate, so every price came out high.
Say you're quoting a client in euros.
In my test, it couldn't look anything up.
So it warned once, then made one up.
€1,104
for $1,200 a month
€1,104 → €1,073.44
Europe's central bank figure
€30.56
too high, every month
Paste today's rate in and it's correct.
Sources and assumptions
- OpenAI API docs: GPT-5.5 model page: GPT-5.5 has a 'Dec 01, 2025 knowledge cutoff'; web_search is listed as a supported tool (checked 2026-11-03)
- OpenAI API docs: Web search guide: Web search is turned on by 'configuring it in the tools array in an API request'; it 'allows models to access up-to-date information from the internet'. Without it the model answers from its training data only (checked 2026-11-03)
- European Central Bank: euro reference exchange rate, US dollar: ECB reference rate on 3 November 2026: 1 EUR = 1.1179 USD (2 Nov 1.1216; 30 Oct 1.1172). Published on business days only (checked 2026-11-03)
Assumptions:
- Tested only on GPT-5.5 through the OpenAI API (Chat Completions, no tools, so no web search). The ChatGPT app can search the web and may look up a live rate. That was not tested, so the post must not claim what the app does.
- Claude was not tested: the Claude test runs failed to start from this research session and returned no output (they're in tests.json with output null). Name only GPT-5.5 (OpenAI API).
- The price list (Juniper Studio) is fictional.
- The 'real' rate is the ECB reference rate for 3 Nov 2026 (1 EUR = 1.1179 USD, so 1 USD ≈ 0.8945 EUR). Banks and payment apps charge their own rates and fees, so the euros a client actually pays can differ.
- The guessed €0.92 rate is wrong only for today. The post must not say when, or whether, €0.92 was ever the real rate.
- The fx-before output was run twice (one scratch probe, one recorded run). Both used the same €0.92 example rate and the same euro figures. Only the recorded run is in tests.json.
- Rounding: each price divided by 1.1179 and rounded to the cent. The model's fx-after figures match this hand calculation on all five lines.


