What did Google release in Document AI on September 21, 2026?
On September 21, 2026, Google released two Document AI processor versions in Preview: a custom extractor powered by Gemini 3.1 Flash Lite that supports fine-tuning in the US and EU, and a Gemini layout parser on the same model.
The source is the Document AI release notes, entry dated September 21, 2026, which we read on October 30, 2026. It is still the newest entry on that page. The entry reads, in part:
"Custom extractor model pretrained-foundation-model-v3.1-lite-2026-07-15 powered by Gemini 3.1 Flash Lite LLM is available in Preview. This processor version supports fine-tuning and is serving multi-region locations in the US and EU."
The second item says the "Gemini layout parser model pretrained-layout-parser-v3.1-lite-2026-08-11 is in Preview." The custom extractor is the Document AI tool for pulling named fields out of document types that have no prebuilt parser, which makes this release relevant to anyone extracting invoice, contract or form data. It is the kind of service we choose between when we build AI document processing for a client.
| Item | Model version | Status | Release note date |
|---|---|---|---|
| Custom extractor, Gemini 3.1 Flash Lite | pretrained-foundation-model-v3.1-lite-2026-07-15 | Preview | September 21, 2026 |
| Gemini layout parser, Gemini 3.1 Flash Lite | pretrained-layout-parser-v3.1-lite-2026-08-11 | Preview | September 21, 2026 |
What is different about the Gemini 3.1 Flash Lite custom extractor?
It is the only Gemini 3 generation custom extractor version that Google lists with both processing and fine-tuning in the US and EU. The earlier Gemini 3 versions use a global endpoint that Google says is not compliant with data residency standards.
Custom extractor model versions
Google's custom extractor documentation lists every model version. The newest ones compare like this, as listed on October 30, 2026:
| Version | Model | Channel | Processing in US/EU | Fine-tuning in US/EU | Released |
|---|---|---|---|---|---|
| v1.5 | Gemini 2.5 Flash | Stable | Yes | US, EU | May 5, 2025 |
| v1.5 Pro | Gemini 2.5 Pro | Stable | Yes | No | June 20, 2025 |
| v1.6 Pro | Gemini 3 Pro | Release candidate | Yes, via a global endpoint | No | December 1, 2025 |
| v1.6 | Gemini 3 Flash | Release candidate | Yes, via a global endpoint | No | January 13, 2026 |
| v3.5 | Gemini 3.5 Flash | Release candidate | Yes | No | May 26, 2026 |
| v3.1-lite | Gemini 3.1 Flash Lite | Release candidate | Yes, with multi-region endpoints | Yes | September 21, 2026 |
Source: Custom extractor with generative AI, checked October 30, 2026. For the two v1.6 versions Google warns that requests "in US and EU endpoints might route to anywhere globally". For v3.1-lite it says customer-managed encryption keys (CMEK) are "not fully supported in all regions" and that support "requires being allowlisted", so a business that needs its own encryption keys has to contact its Google Cloud account team first.
Fine-tuning matters because Google's guide rates it as the highest-accuracy way to train a generative custom extractor, ahead of zero-shot and few-shot, with 10 to 50 or more training documents recommended. Until this release, fine-tuning in the US and EU was only listed for the Gemini 2.5 Flash based v1.5.
Does the new extractor change the price?
No. Google prices the custom extractor at $30 per 1,000 pages for the first million pages a month and $20 above that, and says the cost is the same whether you use a generative AI processor version or a custom model.
From the Document AI pricing page, read October 30, 2026: the custom extractor is $30 per 1,000 pages up to 1 million pages and $20 above, with savings plan prices of $27 (1 year) and $24 (3 years). Note 1 on the page says "The cost to call the custom extractor is the same whether you use a generative AI-powered processor version or custom model." Layout Parser, including initial chunking, is listed at $10 per 1,000 pages. The release note itself says nothing about price.
The same model is also sold directly. On the Gemini API pricing page, Gemini 3.1 Flash-Lite costs $0.25 per million input tokens and $1.50 per million output tokens. In our 2026 document AI cost-per-page dataset, that works out to about $1.24 per 1,000 one-page invoices on our stated assumptions, against $30 through the custom extractor. The gap is what you pay Document AI for: a labelling and evaluation workbench, fine-tuning, the processor API and Google Cloud's Document AI data terms.
What is still missing while it is in Preview?
Preview means no SLA and, unless Google says otherwise, use in test environments only. Google's documentation also lists confidence scores only for custom extractor versions 1.4, 1.5 and 1.5 Pro, so check whether v3.1-lite returns them before you build review rules on them.
- No production commitment yet. Google's launch stage definition says Preview offerings come with "no SLAs or technical support commitments" and are "intended for use in test environments only" unless Google states otherwise. It puts the average Preview stage at "about six months" (Google Cloud launch stages).
- Confidence scores. The model version page says "Versions 1.4, 1.5 and 1.5 Pro support confidence scores." The v3.1-lite version is not named in that sentence. Confidence scores are how a workflow decides which fields a person must check, so this is the first thing we would test.
- Encryption keys. CMEK needs allowlisting, as above.
One related feature landed earlier in the year. On June 4, 2026, Google added document validation and correction to the custom extractor in Preview, using Common Expression Language rules. That covers checks such as line items adding up to the total, which we otherwise build in the workflow.
What this means for your business: should you move your extractor?
If you run Document AI custom extractors and need US or EU data residency, test v3.1-lite on a copy now and keep the stable version in production until it leaves Preview. For a new choice, the price comparison is unchanged.
Switch now or test first?
How we would act on it, by situation:
| Your situation | What we suggest |
|---|---|
| You use the Document AI custom extractor in production on v1.5 | Stay on v1.5 in production. Create a test processor version on v3.1-lite, fine-tune it on the same labelled set, and compare field accuracy on 20 to 50 held-back documents. |
| You avoided the Gemini 3 versions because of the global endpoint | v3.1-lite is the first Gemini 3 generation custom extractor listed with US and EU processing and fine-tuning. Check the CMEK allowlist if you need your own keys. |
| Your review rules depend on per-field confidence | Confirm v3.1-lite returns confidence scores before switching; the documentation names only 1.4, 1.5 and 1.5 Pro. |
| You are choosing an invoice extraction service for the first time | Price per page is unchanged at $30 per 1,000. Compare it with the $10 prebuilt invoice models and direct AI model calls in our cost dataset before deciding. |
| You use Document AI layout parsing for search or retrieval | Test the Gemini 3.1 Flash Lite layout parser against your current version on your own tables and figures; the stable default is still v1.0. |
Whichever version you use, keep the checks in your workflow: totals that add up, a known supplier, no duplicates, and a review queue for anything uncertain. Our OCR vs AI document processing FAQ explains where a custom extractor fits, and our guides to Xero invoice processing and QuickBooks AP automation show where the extracted fields go next. If you want a second opinion on your setup, our strategy call and audit are free.
More questions about the Document AI update
Is the Gemini 3.1 Flash Lite custom extractor generally available?
No. Google's release note of September 21, 2026 lists it as Preview, and the model version table lists it as a release candidate. Google's launch stage definition says Preview offerings have no SLA and are intended for test environments unless Google states otherwise. The stable custom extractor versions are v1.5 and v1.5 Pro, both based on Gemini 2.5.
Does the new custom extractor cost more than the old one?
No. The Document AI pricing page lists the custom extractor at $30 per 1,000 pages up to 1 million pages a month and $20 above, and says the cost is the same whether you use a generative AI processor version or a custom model. The release note itself does not mention price.
Can I fine-tune the Gemini 3.1 Flash Lite extractor in the EU?
Google's model version table marks v3.1-lite with fine-tuning in the US and EU, and the release note says it serves multi-region locations in the US and EU. Customer-managed encryption keys are not fully supported in all regions and need allowlisting through your Google Cloud account team.
What is the Gemini layout parser for?
The layout parser reads a document's structure, such as headings, tables and figures, and breaks it into chunks for search and retrieval systems. Google's page says Gemini versions add table parsing and descriptions of figures and tables. The new pretrained-layout-parser-v3.1-lite-2026-08-11 is in Preview; the stable default version is still pretrained-layout-parser-v1.0-2024-06-03.
Should I call Gemini directly instead of using Document AI?
It depends on what you need around the model. Calling Gemini 3.1 Flash-Lite directly costs about $1.24 per 1,000 one-page invoices on our assumptions, against $30 through the custom extractor. Document AI adds a labelling workbench, fine-tuning, evaluation and its processor API. For a small, well-defined job, a direct call with a fixed schema and validation in an n8n workflow is often enough.
Sources and further reading
- Google Cloud: Document AI release notes (September 21, 2026 entry), checked October 30, 2026
- Google Cloud: Custom extractor with generative AI (model versions, confidence scores), checked October 30, 2026
- Google Cloud: Custom extractor overview, checked October 30, 2026
- Google Cloud: Layout parser (processor versions), checked October 30, 2026
- Google Cloud: Document AI pricing, checked October 30, 2026
- Google Cloud: product launch stages, checked October 30, 2026
- Google: Gemini API pricing, checked October 30, 2026
