Case Study7 min read

Which AI Model Grades Writing Level Best? We Tested 6 Across 2 Languages

AI can score a learner CEFR writing level, but which model does it best, and how accurately? We built one flow to detect the language, route it, and grade it, then benchmarked 6 models in English and Hebrew. Here is the scoreboard.

By Danny Lev, Founder & CEO

AI can grade a learner's writing level, but which model does it best, and for how much? We built one flow to detect the language, route it, and score the CEFR level, then benchmarked 6 models across English and Hebrew. Here is the scoreboard:

Model English level Hebrew level $ / item
claude-haiku-4-5
best English, best value
96.7%
92.3%
$0.0012
claude-sonnet-4-6
best all-rounder, tops Hebrew
96.3%
94.4%
$0.003
baseline (mixed)
the flow's default models
95.9%
91.2%
$0.010
gemini-3.5-flash
95.1%
92.7%
$0.014
gpt-5.4-mini
cheapest
94.9%
92.1%
$0.0007
gpt-5-mini
94.6%
91.2%
$0.0012

Two scores per model on the same 120 items, on a 0 to 1 scale where an exact CEFR band is 1.0 and one band off is 0.8. Detection (which language) is left off the table because every model scored 99 to 100 percent once we graded it by dominant script.

Bottom line: ship claude-haiku-4-5 for value, or claude-sonnet-4-6 if Hebrew is your priority. Detection is solved, so use the cheapest model for it. We ship a mixed pipeline.

4 things we learned

1

The best model is per task, not per project.
Sonnet tops Hebrew (94.4%), Haiku tops English (96.7%). The right answer is a mix, not one model everywhere.

2

Reframing beats a bigger model.
Asking a model to pick English, Hebrew, or neither stalled at 87%. Asking which script dominates pushed detection to 99 to 100 percent, on every model.

3

Cheap keeps up.
gpt-5.4-mini lands within about two points of the best scorer, at a fraction of the price. Once the task is framed well, small models compete.

4

More reasoning did not pay.
We tested high-reasoning variants expecting a bump. They came back equal or slightly worse, at two to three times the cost and latency.

How the flow works

One flow does the whole job. The first agent reads the writing and picks the language by its dominant script. A condition node routes the text to the matching examiner, an English CEFR scorer or a Hebrew one, and each returns a level with a short piece of evidence from the text.

The fine print

  • Level score is 1 minus the distance on a 0 to 1 scale, so an exact CEFR band is 1.0 and one band off is 0.8.
  • Reference levels were graded once by Claude Opus 4.8 with high reasoning, so every model is measured against the same yardstick.
  • Detection is scored by dominant script, which is why it is near perfect and left off the table.

Try it on your own task

Build the same flow in Evaligo: detect the language, route it, score the level. Point it at a few examples, pick the models you are curious about, and get this scoreboard for your own task in minutes. The whole benchmark above cost about eight dollars.

#case study#evaluation#model comparison#cefr#multilingual
DL

Danny Lev

Founder & CEO at Evaligo

Founder of Evaligo. Building AI automation tools that help teams ship faster. Previously led engineering at enterprise AI companies.

10+ years in AI/ML engineeringBuilt systems processing millions of AI requests

Ready to Build This?

Start building AI workflows with Evaligo's visual builder. No coding required.

✓ No credit card✓ Free tier available✓ Deploy in minutes

Need Help With Your Use Case?

Every business is different. Tell us about your specific requirements and we'll help you build the perfect workflow.

Get Help Setting This Up

Free consultation • We'll review your use case • Personalized recommendations