AmberPen vs LanguageTool
LanguageTool is a mature open-source checker built on thousands of hand-written rules plus statistical spell checking. Rules are fast and precise for the patterns they encode — but an error that no rule anticipates is invisible to it. AmberPen uses a language model that reads the whole sentence, so it catches errors that are only wrong in context.
Benchmark results
Both tools ran over the same corpora — details and method on the benchmarks page.
| Metric | AmberPen | LanguageTool |
|---|---|---|
| JFLEG fluency (GLEU, 754 sentences) | 0.561 | 0.307 |
| CWEB errors found (recall, 6,845 sentences) | 65.5% | 18.3% |
| CWEB precision | 37.4% | 37.8% |
| CWEB ERRANT F0.5 | 40.9 | 31.2 |
| Median latency | 217 ms | 287 ms |
The headline number is recall: on real web text, LanguageTool flagged fewer than 1 in 5 of the errors human annotators marked, while AmberPen found roughly 2 in 3 — at the same precision. On raw speed LanguageTool's rule engine is slightly ahead at the median and much tighter at the tail; if sub-300 ms p99 matters more to you than catching errors, it wins that trade.
What rule-based checking misses
Each sentence below was sent to a live LanguageTool server (July 2026, default en-US settings). LanguageTool reported no issues for any of them; AmberPen corrected all five.
| Input | AmberPen correction | Why rules miss it |
|---|---|---|
| He poured over the documents all night. | He pored over the documents all night. | “Poured” is a valid word spelled correctly — only the context makes it wrong. |
| She was very weary of strangers after the incident. | She was very wary of strangers. | Same: a real word, wrong meaning. No spelling rule fires. |
| Walking down the street, the trees looked beautiful. | Walking down the street, I thought the trees looked beautiful. | A dangling modifier — grammatical on the surface, absurd in meaning. |
| I have visited Paris last year. | I visited Paris last year. | Present perfect clashes with a past time expression — requires understanding tense across the sentence. |
| The manager, along with his employees, were praised. | The manager, along with his employees, was praised. | Agreement across an interrupting phrase; the nearest noun fools pattern matching. |
To be fair: LanguageTool does catch the classics its rules encode — “could of”, its/it's, their/there, less/fewer, and common homophone pairs like principle/principal. The failures above share a pattern: whenever correctness depends on the meaning of the sentence rather than the shape of the words, a rule has to have been written for that exact case, and mostly one hasn't.
When LanguageTool is the right choice
- You need a free, self-hosted checker and can accept the recall ceiling.
- You only care about spelling and the well-known mechanical rules.
- You need hard sub-500 ms p99 latency on every request.
If you want corrections closer to what a human editor would make — returned as structured edits with offsets, at a median of ~220 ms — that's what AmberPen is built for.
This page covers AmberPen and LanguageTool head to head. If you are still shortlisting, the grammar checker API comparison reviews seven options side by side, the true cost of self-hosting prices the free-and-open-source route, and how grammar correction is measured explains what the scores above do and do not prove.