LAB · Part 1
Does a graphics card beat Claude Fable 5? We measured it.
Does a graphics card in my office beat Claude Fable 5?
That is not a rhetorical question. I did not know the answer beforehand. Claude Fable 5 is the strongest model you can rent today. On the other side stands a model anyone may download for free, running on a graphics card that fits into an ordinary computer. So I gave both the same work — the work that lands in our company every day.
What it was about: our bookkeeping, not exam questions
Most AI model comparisons run on test questions that can be memorised. I am interested in something else: what happens with real documents? So the test ran on our own live data — the way it arrives in the inbox and the scanner:
- Sorting mail: put 100 real e-mails into the right drawer (invoice, payment, authority, advertising …).
- Reading out fields: from 20 real business documents, extract amount, date, sender, reference — every field has to be exactly right.
- Questions to the archive: 10 questions we actually ask (“When did we last pay this?”), answered from the pile of documents.
- Valid format: does the model return the answer so that a program can process it without rework?
- Reading scans: transcribe scanned paper documents, character by character.
We checked every expected answer by hand before anything was counted. Every task ran three times; the majority answer was scored. The rules — which discipline counts how much — were fixed before the first result came in. That matters, otherwise you bend the weights afterwards.
The two contenders
The graphics card: a single card with 32 GB of memory in a machine on our premises. It runs Qwen 3.8 with 27 billion parameters — an open model (Apache licence), an 18 GB download, released on 14 August. It never leaves the building. No document goes outside.
The cloud: Claude Fable 5 by Anthropic via the official API, with the same instructions and the same documents. For comparison, Claude Opus 5 and Claude Haiku 4.5 ran as well.
Same tasks, same instructions, same documents. Only the model changes.
The result
| Discipline | Graphics card (Qwen 3.8) | Claude Fable 5 | Winner |
|---|---|---|---|
| Sorting mail (accuracy) | 92.0 % | 86.0 % | Graphics card |
| Reading out fields | 97.0 % | 95.6 % | Graphics card |
| Questions to the archive | 95 % | 100 % | Cloud |
| Valid format | 100 % | 99.2 % | Graphics card |
| Reading scans | 92.9 % | 85.3 % | Graphics card |
| Typical response time | 1.3 s | 5.3 s | Graphics card |
| Cost of the whole run | CHF 0 | $9.34 (about CHF 8) | Graphics card |
4 out of 5 quality disciplines go to the graphics card. On the archive questions the cloud was better — that was about phrasing, and there the large model has the edge.
Per 1,000 documents: the cloud costs about 18 dollars. The graphics card costs electricity. And it was not slower but faster: a typical answer came after 1.3 seconds instead of 5.3.
The caveats — all of them
I want you to trust this result. So here is what speaks against it:
- It is our case with our data. 100 mails, 20 documents, 10 questions, 3 scans. That is small. It is small because we checked every answer by hand — thousands of unchecked examples would have looked more impressive and proved nothing. Reading scans with only 3 documents in particular is a hint rather than proof.
- The graphics card ran in “thinking mode”. The model reflects briefly before answering. Without that mode it wins only 3 out of 5 disciplines (mail sorting drops to 84 %, archive questions to 90 %) — but answers in 0.4 seconds, and the whole run is done in 9 minutes. With thinking, the whole run took 46 minutes; the cloud needed 33. Per single answer the graphics card was faster; over the whole stack the cloud was.
- Claude Opus 5 was the best model of all at sorting mail (93 %). And Claude Haiku 4.5 had the highest throughput. The cloud did not lose — it simply is not needed for this work.
- The instructions were frozen. We did not coach any of the models. With fine-tuning of the prompts all of them would probably improve — the ratio would likely stay similar, but I did not measure that.
- You have to buy and run the graphics card. If you have 50 documents a month, the cloud is cheaper. From a few thousand documents a month the calculation tips.
What this means for your company
The tasks above are not special cases. Sorting mail, reading invoices, matching payments, filing documents — that is the daily office work that piles up in every company. And for that, an open model on a graphics card in your cabinet is enough today.
What you gain:
- Your documents stay in-house. No customer data, no salaries, no contracts go to a provider abroad.
- The price per document drops to zero. No bill that grows with every scan.
- You are not dependent on what a provider will cost or switch off next year.
The result here is no miracle, it is diligence: choose the right tasks, check answers by hand, only then decide. That is exactly what I do with you if you want to know whether this works for you too. The series continues — every new model goes through the same test. I am happy to walk you to this result.
On method: the weighting (quality 70, speed 20, footprint/cost 10) was fixed before the first run. Every task ran three times, the majority answer was counted, all expected answers were checked by hand. Speed was measured server-side, i.e. without the network share. The documents are real and stay private; only aggregates appear here.