Claude Sonnet 5 review
Claude Sonnet 5, called directly with a file and a list of fields, beat every dedicated extraction product we tested except one. It's the strongest case in our results for building instead of buying. It costs a few cents a page, it comes with none of the product around it, and its weak spots are exactly the ones you'd expect from a model with no document software in front of it.
This row is a general AI model with nothing wrapped around it. The file goes in, the field list goes in, the answer comes out. We include the raw models because that's what many teams build first, and the question they're really asking is how much a dedicated product adds on top. In Claude's case, for most of the field, the answer was less than you'd think. It came second overall, and only DocuPipe scored higher.
The surprise was long documents. Raw models are supposed to run out of room on catalogs and directories with hundreds of entries, and the other two we tested did exactly that, sometimes returning nothing at all. Claude didn't. It worked through the long registers and came back with most of the entries, which put it level with the best dedicated products and well ahead of several of them. There are no modes to pick and nothing to tune, and nothing came back empty. That's why it scores well on resilience even though it badly misread a couple of ordinary forms without any warning.
Where it loses is perception, not reasoning. Hebrew and Arabic layouts pulled its score well below its own average. So did handwriting and photographed scans. Those are the cases where the page has to be read correctly before any intelligence helps, and a model with no layout or scanning stage in front of it has nowhere to lean. This is exactly where a dedicated product earns its keep.
The price is the easy part. At $2 per million input tokens and $10 per million output, a page costs a few cents at most, there's no plan to buy, and you pay only for what you send. As a raw cost of extraction it undercuts every dedicated product here, and batch processing brings it lower still for anything that isn't time-sensitive.
The product is the hard part, because there isn't one. The API itself is well suited to the job, with native PDF input up to 600 pages per request, a Files API, structured outputs, citations back to the page, and official SDKs in seven languages. But there's no document dashboard, no upload flow, no webhooks, no job queue, no review screen, and no retention control beyond the platform's own policy. Every one of those is something a dedicated product includes and you'd be building. If you already run a model pipeline, the marginal cost is close to nothing. If you're starting from zero, the fair comparison is the price of the model plus the weeks it takes to wrap it.
So if you already have a model pipeline and your documents are printed, left-to-right, and of moderate length, this is the model to use, and you may not need to buy anything. If your documents arrive rotated, handwritten, or in a right-to-left language, or if you need the product around the model, the gap to a dedicated system is real.
Teams that already run a model pipeline, work with printed left-to-right documents of moderate length, want the lowest cost per page, and can live without document tooling.
Scanned, rotated, handwritten, or right-to-left inputs, and any team that needs webhooks, a dashboard, review, or retention controls out of the box.
How it scores, area by area
The accuracy score comes from the benchmark. Every other score is ours, with the reason written next to it. We use the same weights on every review and comparison.
| Area | Claude Sonnet 5 |
|---|---|
| Accuracy35% of total | 9.2Ranked 2nd of 8 systems on the benchmark. |
| Resilience15% of total | 7.0No settings to tune and nothing returned empty, but a few ordinary forms came back badly misread with no warning, and there is no layout stage to catch them. |
| Price15% of total | 9.0$2 and $10 per million tokens; a few cents per page at most. |
| Getting started10% of total | 3.0No document product. You build upload, review and results tooling yourself. |
| API and integration10% of total | 5.0Excellent general SDKs, native PDF input, a Files API, batch mode and structured outputs; no document jobs, webhooks or pipeline. |
| Data control7.5% of total | 5.0Platform policy only; zero retention for eligible organisations. Nothing document-specific. |
| Compliance7.5% of total | 6.0Platform certifications apply to the API; your pipeline is your responsibility. |
| Overallweighted | 7.2weighted score out of 10 |
How each area is scored
- Accuracy (35%). The benchmark result, on a 0 to 10 scale.
- Resilience (15%). Does it just work? Fewer modes to choose, no tuning to reach the ranked result, no runs that fail outright, and no documents that come back quietly empty. Informed by the benchmark, judged by us.
- Price (15%). Published cost per page for parse plus extraction, cheapest scoring highest. Unpublished pricing is marked down.
- Getting started (10%). How far you get without a sales call: a dashboard, a free tier with no card, help authoring schemas, a test environment.
- API and integration (10%). Webhooks and which tier they are on, SDKs, no-code connectors, whether the API is built around document jobs at all.
- Data control (7.5%). Deletion controls, scheduled retention, zero-retention availability and on which tier, choice of region.
- Compliance (7.5%). Certifications stated on public pages: SOC 2, ISO 27001, HIPAA with a BAA, GDPR, private deployment options.
By document type
Average accuracy on each group of benchmark documents. The evidence page has every document and score.
| Document type | Claude Sonnet 5 |
|---|---|
| Statements, filings, reports2 to 5 pages · 21 documents | 92.3% Good |
| Short forms, invoices, receiptsone page · 30 documents | 91.7% Good |
| Long registers, catalogs, directories6 pages and up · 21 documents | 91.3% Good |
| Non-English documents12 languages · 25 documents | 85.2% Good |
| Scans and photosimages, not PDFs · 7 documents | 80.1% Good |
Price, product, and policies
| Claude Sonnet 5 | |
|---|---|
| Pricing model | Per token through the Anthropic API. No document-specific product or plan. |
| Entry price | $2 per million input tokens, $10 per million output tokens. |
| Cost per page | Depends on page size and how much you ask it to return; typically a fraction of a cent to a few cents per page. |
| Free tier | No ongoing free tier for API use. |
| Pay as you go | Yes. Pure usage billing. |
| Dashboard | No document product. A console workbench for prompts; upload, review and results tooling is yours to build. |
| Webhooks | None for document jobs. You run the pipeline. |
| API and SDKs | Official SDKs in Python, TypeScript, Java, Go, Ruby, C# and PHP. PDF input up to 600 pages per request, a Files API, batch processing at reduced cost, structured outputs and citations. |
| Data retention | Governed by Anthropic's API data policies; zero-retention arrangements exist for eligible organizations. Nothing document-specific. |
| Compliance | Anthropic's platform certifications apply to the API; your pipeline around it is your responsibility. |
We checked these against the vendors' own pricing and documentation pages on 2026-09-11. Sources: Claude Sonnet 5: platform.claude.com. Prices change, so confirm before you budget.
Compare Claude Sonnet 5 with another system
Frequently asked
Is Claude good at document extraction?
Yes. Called directly with a file and a field list, Claude Sonnet 5 came second of the eight systems in our benchmark, ahead of every dedicated extraction product except DocuPipe. It handles long documents unusually well for a raw model. Its weak spots are handwriting, scanned images, and right-to-left layouts.
What does it cost to use Claude for document extraction?
It's billed per token, $2 per million input and $10 per million output, which comes to a few cents per page at most. Nothing document-specific is included. You build the pipeline, dashboard, and webhooks yourself.
Scores are macro-average field accuracy from the DocuBench 0.2.0, 72-document run. Every document, answer key, and raw extraction is open in the explorer; the scoring rules are on the methodology page, and the replication kit re-runs the scorer. A document a system could not process scores zero rather than being dropped.