GrindBenchThe daily grind, benchmarked.

Can AI do your work?

Three questions, in order: can any model do this kind of work yet, which one, and what would it cost. We only cover the kinds of work we've tested, and a person still has to check every draft.

1. Can it?

The task: Redact a record under the named ATIA / Privacy Act exemptions and cite each one.

Not yet. The best model for it, Claude Opus 5.5, got only 33% of its 55 drafts right.

If it drafted 20 of these: 7 right, 13 not. Nobody can tell which without reading all 20.

The typical model managed 7%. For now this work needs a person doing it, with AI at most a first pass.

A price is the AI's bill for one draft. The real cost is the person checking and fixing it, which this benchmark doesn't measure.

2. Which model, and what's the bill?

Pick a job to start from, set how many of each task you do in a week, and how good the drafts need to be. The bill is the model's own charge at list price, for the drafting only.

Where you work
How many of these you do in a week

Only the kinds of work we test, not your whole job. These are rough starting guesses: change them to match your week.

How good does it need to be?

22 models clear your bar. Drafting this much with the cheapest would cost about 1.4¢ a month in AI fees.

For 4 tasks a week, with every task type at least 50% acceptable. That's the AI's bill for drafting only: a person still checks every draft, and that time isn't counted here. Cheapest first:

  1. BGPT-6 LunaOpenAI · at least 69% acceptable on your work1.4¢a month · 0.12¢ per good result
  2. ANemotron 3 SuperNVIDIA · at least 81% acceptable on your work3¢a month · 0.21¢ per good result
  3. AMercury 3Inception · at least 86% acceptable on your work3.8¢a month · 0.25¢ per good result
  4. BDeepSeek V4.1 FlashDeepSeek · at least 78% acceptable on your work4.7¢a month · 0.35¢ per good result
  5. BGLM-5.3 FlashZ.ai (Zhipu) · at least 78% acceptable on your work6.3¢a month · 0.47¢ per good result
  6. AQwen3.8 FlashAlibaba (Qwen) · at least 83% acceptable on your work6.9¢a month · 0.48¢ per good result
  7. AUltramarine 1 FlashPaper Instruments · at least 83% acceptable on your work7.3¢a month · 0.5¢ per good result
  8. CQwen3.6 35B A3BAlibaba (Qwen) · at least 58% acceptable on your work8.9¢a month · 0.89¢ per good result
  9. BMiniMax M3MiniMax · at least 68% acceptable on your work9.1¢a month · 0.77¢ per good result
  10. CCommand A+Cohere · at least 53% acceptable on your work9.5¢a month · 1¢ per good result
  11. ANemotron 3 UltraNVIDIA · at least 83% acceptable on your work$0.15a month · 1.1¢ per good result
  12. CDeepSeek V4 ProDeepSeek · at least 51% acceptable on your work$0.17a month · 1.9¢ per good result
  13. BGPT-6.1 SolOpenAI · at least 80% acceptable on your work$0.19a month · 1.4¢ per good result
  14. BMuse Spark 1.3Meta · at least 78% acceptable on your work$0.27a month · 2¢ per good result
  15. AGemini 3.8 FlashGoogle · at least 81% acceptable on your work$0.32a month · 2.3¢ per good result
  16. BMuse Spark OpenMeta · at least 80% acceptable on your work$0.34a month · 2.4¢ per good result
  17. AClaude Sonnet 5.5Anthropic · at least 90% acceptable on your work$0.63a month · 4¢ per good result
  18. AKimi K3.1 ProMoonshot AI · at least 90% acceptable on your work$0.79a month · 5¢ per good result
  19. AGLM-5.3Z.ai (Zhipu) · at least 88% acceptable on your work$0.80a month · 5.2¢ per good result
  20. AMuse WonderMeta · at least 81% acceptable on your work$0.89a month · 6.3¢ per good result
  21. AClaude Opus 5.5Anthropic · at least 97% acceptable on your work$1.13a month · 6.8¢ per good result
  22. AKimi K3Moonshot AI · at least 88% acceptable on your work$1.45a month · 9.5¢ per good result
Models that miss your 50% bar (6)
  1. CNemotron 3.5 LightningNVIDIA · at least 42% acceptable on your work3.3¢a month · 0.45¢ per good result
  2. DMistral Large 4Mistral AI · at least 39% acceptable on your work6.4¢a month · 0.95¢ per good result
  3. FNemotron 3 Nano OmniNVIDIA · at least 15% acceptable on your work1.7¢a month · 0.63¢ per good result
  4. FGemma 4 31BGoogle · at least 14% acceptable on your work0.83¢a month · 0.35¢ per good result
  5. FGemma 4 26B A4BGoogle · at least 2% acceptable on your work0.65¢a month · 2.2¢ per good result
  6. FMistral Small 4Mistral AI · at least 0% acceptable on your work1.5¢a month · no good results

Costs are the AI's bill only: the tokens each model used per task in our tests, at OpenRouter list price. They don't include the time to check its work, and you should always check it. A month is 4.33 weeks.

Next

How we test

Where the tasks come from, how every answer is marked, and how we check the marking.