Ten months into 2026, 29 of our calculators still said 2025. Here is what re-checking them found

The starting point: 29 pages a year behind

In August we audited 102 calculators and fixed 29 of the 51 salary models. By October the catalog had grown to 247 calculators and decision tools, and 27 of them were still labelled for the 2025 tax year. Two more were labelled for fiscal years that had already ended: the UK's 2025-26 year closed in April 2026 and India's FY2025-26 in March. That made 29 pages whose own title told a reader they were out of date, including the German "Brutto Netto Rechner", the catalog's flagship page.

The first drafts of these models were written from an AI model's knowledge of each country's rules and then checked against official sources. A model's knowledge stops at a date, and tax law does not, so the second check is the one that decides whether a page can say "2026". For each of the 29 we opened the statute, the tax authority's published table or its own calculator, compared every constant, and recomputed every worked example through the engine rather than by hand.

What came out of it

• 18 calculators moved to their 2026 tax year: 10 needed figures changed and 8 needed only the label.
• 11 stayed on 2025 with a visible note on the page saying why: 7 illustrative models that never claimed a statutory schedule, and 4 where the 2026 figure could not be pinned to a primary source.
• 4 errors were found that had nothing to do with the year: a wrong legal date in page copy, a credit modelled as a flat maximum, a reduction the page described but the model never applied, and two wrong rates in a vehicle-tax table.

The changes were mostly the boring kind, which is the point

Germany's 2026 income-tax tariff raised the basic allowance (Grundfreibetrag) to 12,348 EUR and moved the zone edges to 17,799, 69,878 and 277,825; the social-insurance ceilings rose to 101,400 and 69,750. On a 48,000 EUR salary the model now computes taxable income of 36,294 and income tax of 6,060.24. Italy cut its second IRPEF rate from 35% to 33%. Spain raised the employee social-security rate to 6.50% and the contribution ceiling to 5,101.20 a month. Turkey moved its wage bands to 190,000, 400,000, 1.5 million and 5.3 million lira.

None of these is surprising, and every one of them makes a 2025 calculator wrong by a predictable amount for every user. Indexed thresholds move every year by design. A page that is not refreshed does not become approximately right; it becomes exactly wrong by the indexation.

The other eight needed nothing but the label. Czechia's figures were already the 2026 ones, Poland, Romania and Russia had not changed their rates, the UK froze its thresholds for 2026-27, and India's Finance Bill 2026 made no change to the new-regime rates. Saying "2025" on those pages was a false negative, and it cost the same trust as a false positive.

Four errors that every test had passed

Each of these models has test cases the engine runs on every build, and every one of them was passing. That is expected: a test written from the same understanding as the model checks the arithmetic, not the understanding. The errors only showed up when a person put the law next to the spec.

• Greece, vehicle tax: the table for cars registered from November 2010 to 2020 charged 2.92 EUR per gram of CO2 for 181-200 g/km and 3.20 for 201-250 g/km. The tax authority's table says 2.78 and 3.05. Every car in those two bands was overcharged; a 190 g/km car now comes out at 528.20 EUR.
• Greece, net salary: the page said tax was charged "after the credit", but the model never subtracted one. 2026 brought a new six-band scale (9% to 44%) and a 777 EUR tax reduction that tapers by 20 EUR per 1,000 EUR above 12,000. Both are now in the model.
• Netherlands, net salary: the labour credit (arbeidskorting) was a flat maximum. The real credit builds up across four bands to a peak of 5,685 EUR at 45,592 EUR and then tapers, so the flat version overstated the credit for everyone earning below the peak.
• Germany, vehicle tax: the rates were right, but the page said the stepped CO2 scale applies to cars first registered from July 2009. It applies from 2021. The 2009-2020 regime is now named as out of scope instead of silently mixed in.

Eleven pages that now say what they do not know

Seven of the eleven are illustrative vehicle-tax models: Australia, South Africa, Israel, Switzerland, Finland, Lithuania and Georgia, where the charge is set by state, province, canton or municipality, or the schedule is not published in a form we could reach. Labelling them "2026" would claim a currency they never had.

The other four are real models we could not refresh honestly. Belgium reformed its tax-free sum in July 2026, but the finance ministry's pages sit behind a captcha and the only 2026 figures we found were secondary. Moscow amended its vehicle-tax law on 1 November 2025, but the official document could not be parsed. Ireland's rates were reported unchanged, but every official table returned 403 or timed out. Romania restructured its 2026 car tax by Euro emission standard and engine size, with a municipal surcharge on top, which needs a new input rather than a refresh.

Each of these pages keeps its 2025 label and says why next to it. Under-claiming is the safe direction: a reader who sees "not yet confirmed for 2026" can go and check, while one who sees a confident "2026" has no reason to.

Making staleness a build failure

The 29 stale pages were not a research problem; they were a process problem. Nothing failed when January came. So the catalog now has a single constant for the current tax year and a test that fails for any calculator whose year is behind it unless that calculator carries an explicit, published reason.

Two details matter. The rule compares the year a tax period starts in, so 2026-27 and FY2026-27 count as current and 2025-26 does not. And the current year is a constant someone bumps on purpose, not today's date: a check against the clock would break every build on 1 January, and a check people learn to override is no check at all. Bumping it turns every unrefreshed page into a failing test, which is exactly the to-do list the next refresh needs.

Every calculator publishes its formula, constants, sources with the date each was checked, and the test cases it passes. That is what made this re-check a day of work rather than a rebuild, and it is what lets anyone repeat it.