GPT-5.5 review
GPT-5.5, called directly with a file and a list of fields, is as good as almost anything we tested on a typical short document, and it returns nothing at all on the longest ones. It's the priciest of the three direct models, and like the others it comes with no document product around it. Its middling rank is the average of those facts, and none of them describe it well on their own.
Like the other raw models in our test, this row is a general AI model with no document software around it, included because it's what many teams reach for first. On the document most businesses actually process, a form or invoice or statement of a page or two, GPT-5.5 was a very capable extractor. Its typical document came back nearly perfect, and on a few it matched or beat the products that finished well above it, DocuPipe included.
The problem is what happens past a certain length. On nine of the long catalogs, directories, and registers in our set it returned nothing usable, and on the rest of the long ones it returned a fraction. This isn't the model misreading the page. It's the model running out of room to answer and stopping, with everything after that point lost. A single call has no way to work through a long document in pieces, so the failure is structural and no setting fixes it. That's why it scores low on resilience. On your documents it'll happen the same way it happened on ours.
The fix is a pipeline that splits long documents and stitches the results back together, which is real engineering and exactly what a dedicated product is selling. Interestingly, Claude called the same way didn't collapse on the same documents, so the raw models aren't interchangeable here. Handwriting and degraded scans were a secondary weakness, and right-to-left layouts were middling rather than bad.
At $5 per million input tokens and $30 per million output it's the most expensive of the three direct models, though still only cents per page, and the batch API at half price is a real saving for anything that isn't time-sensitive. There's no free tier for API use and no plan to buy. You pay for usage.
The API is well equipped for documents as far as it goes: PDF, Office, and spreadsheet input up to 50 MB per request, a Files API, structured outputs, and official SDKs. What it lacks is everything a document product provides. No dashboard, no upload flow, no webhooks, no job queue, no review screen, and no retention controls beyond OpenAI's platform policy. You're building the pipeline.
So it's a strong reader with no strategy for length, at a price above its two direct rivals. It's fine to build on if every document is short and you're already standardized on it. Otherwise it's hard to justify for this particular job.
Teams already standardized on OpenAI, building their own pipeline around short, printed documents, who'll take on the job of handling anything long themselves.
Any corpus with long list-shaped documents, and any team that wants document tooling or a cheaper direct model.
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 | GPT-5.5 |
|---|---|
| Accuracy35% of total | 7.6Ranked 5th of 8 systems on the benchmark. |
| Resilience15% of total | 4.0Nine documents came back empty and there is no setting that fixes it; you have to build the chunking yourself. |
| Price15% of total | 8.0$5 and $30 per million tokens, the priciest direct model, still only cents per page; batch at half price. |
| Getting started10% of total | 3.0No document product. You build the tooling yourself. |
| API and integration10% of total | 5.0Strong general SDKs, PDF and Office input, a Files API, batch at half price and structured outputs; no document jobs or webhooks. |
| Data control7.5% of total | 5.0Platform policy only. Nothing document-specific. |
| Compliance7.5% of total | 6.0Platform certifications apply to the API; your pipeline is your responsibility. |
| Overallweighted | 6.1weighted 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 | GPT-5.5 |
|---|---|
| Short forms, invoices, receiptsone page · 30 documents | 88.7% Good |
| Scans and photosimages, not PDFs · 7 documents | 88.7% Good |
| Statements, filings, reports2 to 5 pages · 21 documents | 85.1% Good |
| Non-English documents12 languages · 25 documents | 69.4% Mixed |
| Long registers, catalogs, directories6 pages and up · 21 documents | 50.4% Weak |
Price, product, and policies
| GPT-5.5 | |
|---|---|
| Pricing model | Per token through the OpenAI API. No document-specific product or plan. |
| Entry price | $5 per million input tokens, $30 per million output tokens. |
| Cost per page | Depends on page size and output length; typically a cent to several cents per page, higher than the other two direct models. |
| Free tier | No ongoing free tier for API use. |
| Pay as you go | Yes. Pure usage billing; batch processing at half price. |
| Dashboard | No document product. A playground 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 and others. PDF, Office and spreadsheet input up to 50 MB per request, a Files API, a batch API at half price, structured outputs. |
| Data retention | Governed by OpenAI's API data policies. Nothing document-specific. |
| Compliance | OpenAI'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: GPT-5.5: developers.openai.com. Prices change, so confirm before you budget.
Compare GPT-5.5 with another system
Frequently asked
Is GPT-5.5 good for document extraction?
On short documents, yes. Called directly it was competitive with the best of the eight systems we tested on one-page forms and statements. On long documents it fails structurally, returning nothing on nine of the longest in our set because its answer runs out of room.
What does GPT-5.5 cost for document extraction?
It's billed per token, $5 per million input and $30 per million output, the highest of the three direct models we tested, though still cents per page. Batch processing is half price. Nothing document-specific is included.
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.