Can AI read foreign receipts accurately?

Less accurately than US receipts. In our October 2026 test every model scored lower on 24 Indonesian receipts. The best, Gemini 4 Argon, scored 82.0.

Which AI is best for non-US receipts?

Gemini 4 Argon scored 82.0 on our Indonesian receipts, ahead of Grok 4.7 at 77.4. Gemini 4 Argon was also inconsistent, so check its work.

Did you test European or Canadian receipts?

No. Our non-US set was 24 Indonesian receipts. Other countries may score differently.

Which AI should I avoid for receipts?

Grok 4.3 scored 52.7 on Indonesian receipts and invented totals on unreadable receipts.

RealRun / Guides

Can AI read foreign receipts?

Short answer: Yes, but not as well as US receipts. Every one of the 8 models scored lower on our 24 Indonesian receipts. Gemini 4 Argon did best on them at 82.0 (94.5 on US receipts). GPT-6.1 Sol, the top model overall, scored 97.7 on US receipts but 76.3 on Indonesian ones. Grok 4.3 fell to 52.7. If you buy from overseas vendors or travel for work, have a person check those receipts.

Scores: US vs Indonesian receipts

Model	Indonesian receipts	US receipts	Overall	Severe errors	Cost
Gemini 4 Argon	82.0	94.5	93.6	1 of 3 passes	medium
Grok 4.7	77.4	84.6	83.1	both passes	high
Claude Opus	77.0	91.8	89.7	both passes	subscription
GPT-6.1 Sol	76.3	97.7	95.8	none	medium
Claude Haiku	75.6	95.2	92.8	1 of 3 passes	subscription
Gemini 3.5 Flash-Lite	73.5	88.5	85.4	both passes	low
GPT-6 Luna	72.3	96.0	91.4	none	low
Grok 4.3	52.7	76.8	69.2	every pass	medium

Our non-US set was 24 Indonesian receipts, part of the 79-document receipts test. We did not test European, Canadian or other non-US receipts, so don't assume the same scores for them. The test also had 15 degraded photos.

How the testing works: we wrote our own synthetic receipts, bills and invoices with planted traps: math that doesn't add up, fake invoices, duplicates and unreadable documents. A separate held-back set of 20 documents, written blind and kept out of all prompt and scoring tuning, only ran at the end. Every model got the same 79 documents (59 main + 20 fresh held-back documents; 24 non-US, 15 degraded photos), 1,501 graded answers, passes: 2+ per model. Score: field, category and flag accuracy, needs-human recall, minus 5 per severe error, mean of passes. All data is synthetic, never real accounts.

Results updated 15 Oct 2026. Tests run October 2026. Full results: Receipts.

What we found

Every model scored lower on the 24 Indonesian receipts (e.g. GPT-6.1 Sol 97.7 on US vs 76.3 on Indonesian).
Gemini 4 Argon: best on Indonesian receipts, but inconsistent. Its passes ranged from 84.1 to 98.8 overall.
Grok 4.3: Invented totals on unreadable receipts, severe errors in every pass, and wrong duplicate pairs.
No model accepted a fake invoice.

For you

The way this goes wrong is quiet. A receipt from an overseas supplier or a work trip gets read with a wrong field, and nothing looks off until the books don't match. Every model we tested slipped on foreign receipts, so send those to a person for a quick check, even if you trust the AI on your US ones.

For your clients

If the bookkeeper or business you build for has overseas vendors, don't sell them the overall score. Show them the foreign receipt column, where the best model scored 82.0, and build a human review step for non-US receipts into the setup.

Want the exact prompt and setup?

Get the playbook for this job, or all five for $99 (save $46).

Receipts AI Playbook, $29 or get the free weekly AI test report

FAQ

Can AI read foreign receipts accurately?: Less accurately than US receipts. In our October 2026 test every model scored lower on 24 Indonesian receipts. The best, Gemini 4 Argon, scored 82.0.
Which AI is best for non-US receipts?: Gemini 4 Argon scored 82.0 on our Indonesian receipts, ahead of Grok 4.7 at 77.4. Gemini 4 Argon was also inconsistent, so check its work.
Did you test European or Canadian receipts?: No. Our non-US set was 24 Indonesian receipts. Other countries may score differently.
Which AI should I avoid for receipts?: Grok 4.3 scored 52.7 on Indonesian receipts and invented totals on unreadable receipts.

More from RealRun

Scores reflect our tests on the dates shown, not a guarantee. Claude models ran on a Claude subscription, so they have no metered API cost. cost_tier is relative within a task (low/medium/high) from our run spend; 'subscription' = not metered. Tested October 2026 by RealRun (Base Ledger IP LLC).