Gemini 3.5 Flash review
Gemini 3.5 Flash is the cheapest thing we tested and, on a single page, one of the best. On anything long it's the worst, and some runs failed outright. Whether it belongs in your pipeline comes down to two questions: how many pages are in your documents, and how much plumbing you're willing to build around a model that ships with none.
Let's be clear about what this row is. It isn't a document product. It's Google's fast general model, handed a file and a list of fields with nothing built around it. We test it alongside dedicated products like DocuPipe and Reducto because that's exactly what a lot of engineering teams build in a weekend, and the real question is whether they need to buy anything at all.
For short documents, we think they mostly don't. On one-page forms, invoices, receipts, and photographed or rotated scans, Gemini was second only to DocuPipe, and it returned more documents completely correct than any other system, DocuPipe included. That surprised us. A receipt photographed at an angle on a phone is supposed to be the hard case, and this model treated it as routine.
Then the long documents arrive, and the story flips. On registers, catalogs, and directories of six pages or more it posted the worst average of any system, and it was the only one where several runs simply failed and returned nothing. The reason is structural rather than a bug. A single model call has a fixed amount of room for its answer, so on a document with hundreds of entries it starts, fills that room, and stops. Everything past that point is lost, and nothing warns you. That's also why we score it low on resilience. The same call is excellent one day and empty the next, and the only thing that changed was the page count.
The price is where it earns its place. At $1.50 per million input tokens and $9 per million output, a typical page costs well under a cent. There's a free tier for development, no plan to buy, and no minimum. On short documents the accuracy per dollar is hard to argue with, and if your pipeline is one page per file we'd struggle to justify paying much more than this for accuracy alone.
What the price doesn't include is everything else. There's no dashboard, no upload flow, no webhooks, no notion of a document job, and no retention control beyond Google's platform policy, which differs between the free and paid tiers. You're buying a model, not a product. The upload, queueing, retry, review, and storage layers are yours to build and run, and so is the compliance story around them. Teams that already have that plumbing will find this a bargain. Teams that don't should price the engineering before comparing it to a product that ships all of it.
So for short, high-volume, image-heavy work where cost matters and you have your own pipeline, it's near the top. For mixed or long documents it's nowhere near a shortlist. Its middling overall rank undersells it on one and oversells it on the other. Decide on your documents, not on the average.
High-volume single-page work such as receipts, forms, declarations, and photographed documents, run by a team that already has a pipeline and wants the lowest possible cost per page.
Anything with many pages or hundreds of rows, any team without its own document plumbing, and any workflow that needs webhooks, retention controls, or a dashboard 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 | Gemini 3.5 Flash |
|---|---|
| Accuracy35% of total | 7.3Ranked 6th of 8 systems on the benchmark. |
| Resilience15% of total | 3.0Several runs failed outright and the long-document average was the lowest in the test; excellent one day and empty the next depending on page count. |
| Price15% of total | 10.0$1.50 and $9 per million tokens plus a free tier; well under a cent per page. |
| Getting started10% of total | 4.0No document product, though the free tier makes trying it trivial. |
| API and integration10% of total | 5.0Strong general SDKs, the largest PDF page limit of the direct models, a Files API and structured output; no document jobs or webhooks. |
| Data control7.5% of total | 4.0Platform policy only, and it differs between the free and paid tiers. |
| Compliance7.5% of total | 6.0Platform certifications apply to the API; your pipeline is your responsibility. |
| Overallweighted | 6.2weighted 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 | Gemini 3.5 Flash |
|---|---|
| Scans and photosimages, not PDFs · 7 documents | 95.5% Strong |
| Short forms, invoices, receiptsone page · 30 documents | 94.1% Good |
| Statements, filings, reports2 to 5 pages · 21 documents | 87.5% Good |
| Non-English documents12 languages · 25 documents | 68.9% Mixed |
| Long registers, catalogs, directories6 pages and up · 21 documents | 28.3% Weak |
Price, product, and policies
| Gemini 3.5 Flash | |
|---|---|
| Pricing model | Per token through the Gemini API. No document-specific product or plan. |
| Entry price | $1.50 per million input tokens, $9 per million output tokens on the paid tier. |
| Cost per page | Depends on page size and output length; typically well under a cent to a couple of cents per page, the cheapest of the direct models. |
| Free tier | Yes. A rate-limited free tier for development. |
| Pay as you go | Yes. Pure usage billing. |
| Dashboard | No document product. AI Studio for trying 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, JavaScript, Go and others. PDF input up to 1,000 pages or 50 MB, a Files API with 48-hour storage, structured output. |
| Data retention | Governed by Google's API data policies, which differ between the free and paid tiers. Nothing document-specific. |
| Compliance | Google'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: Gemini 3.5 Flash: ai.google.dev. Prices change, so confirm before you budget.
Compare Gemini 3.5 Flash with another system
Frequently asked
Is Gemini 3.5 Flash good enough for document extraction?
For single-page documents, yes. It was second-best of the eight systems in our benchmark on one-page forms and scanned images. For multi-page or list-heavy documents, no. It had the lowest average of any system on long documents and several runs failed outright.
What does Gemini 3.5 Flash cost for document extraction?
It's billed per token, $1.50 per million input and $9 per million output on the paid tier, which works out to well under a cent for a typical page. There's a free tier for development. Nothing document-specific is included. You build the pipeline 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.