Six Invoices. Qwen Matched Opus. Does Your Workflow Need a Premium Model?

An invoice arrives. Your application needs its number, currency, amount, due date and bank account.

Before sending another routine extraction request to a premium model, it is worth asking: could Qwen do this job?

We put that question to a small, practical test through unoblox.

Same invoices. Same requirements.

We gave Qwen3.8-27B, Claude Opus 4.8 and GPT-5.6-Luna six text-based invoice examples. Each model had to return exactly five fields in valid JSON, preserve the facts and leave missing information as null.

The examples covered missing details, misleading instructions inside the document, zero amounts, tax totals, negative credit amounts and bank-account numbers with leading zeros.

Qwen and Opus got all six exactly right. Luna got five.

Model	Exact invoice extractions
Qwen3.8-27B	6/6
Claude Opus 4.8	6/6
GPT-5.6-Luna	5/6

Luna missed an invoice ID that was present in one document. We checked answers against expected values written before the run, without retries or output repairs.

The trade-off was speed

Qwen took a median of 8.25 seconds, compared with 1.77 seconds for Opus and 1.86 seconds for Luna.

For an invoice processed in the background, that extra time may be acceptable. For a user waiting on screen, it may matter more.

This is where the model decision becomes practical: what accuracy, response time and cost does your workflow actually require?

What did it cost in rupees?

Model	Cost for six attempts	Per correct extraction	Projected per 1,000 correct extractions
Qwen3.8-27B	₹0.042	₹0.007	₹7.00
Claude Opus 4.8	₹1.484	₹0.2473	₹247.29
GPT-5.6-Luna	₹0.06646	₹0.01329	₹13.29

Prices checked September 25, 2026. Update, 6 October 2026: Qwen3.8-27B is now free on unoblox (₹0 per token), so this workload costs nothing on Qwen today; current Opus 4.8 and GPT-5.6-Luna rates are on their model pages. Opus and Luna figures are token-price estimates, not settled wallet charges; they exclude tax, cache adjustments, per-request rounding and other workflow expenses. Luna's cost per correct result includes all six attempts divided by five successes. Projections assume the same token use and observed success proportion.

Qwen's cost for this workload is ₹0.007 per extraction, confirmed by unoblox. That is ₹0.042 for these six extractions, or ₹7 for 1,000 at the same per-extraction cost. For comparison, the catalogue token-price projections are ₹247.29 for Opus and ₹13.29 for Luna per 1,000 correct extractions.

Find out before your volume grows

Six synthetic examples cannot establish production reliability or prove that one model replaces another. But Qwen matching Opus on these examples is a reason to evaluate it on your own invoices.

Start with a representative sample. Include incomplete documents, unusual amounts and the formats your customers actually send. Check every required field, then compare cost per accepted extraction—including failed attempts and retries.

Our test used text only. Scanned invoices and OCR need their own evaluation.

Your invoice workflow may have room for a less expensive model. The way to find out is to give Qwen the same job and check the result.

Try Qwen3.8-27B through unoblox and run the comparison on your documents.

Test scope: six distinct synthetic text invoices, one attempt per model, identical prompts and an 8,192-token output ceiling. Models used their default reasoning behavior. Results describe this small test, not a general model ranking.