A model that returns the right invoice total on one PDF may still fail on an image-only statement, a rotated table or a repeated page header. A useful API comparison starts with document parsing and traceable extraction rather than broad model rankings.
Comparison scope (verified 2026-10-25): An English-language operations group extracts invoice number, issuer, dates, totals and line items from scanned and native PDFs. No exact model benchmark or regional API entitlement is asserted for 2026-10-25.
Evidence at a glance
| Product, approach or evidence source | What the cited document establishes | What it does not establish |
|---|---|---|
| OpenAI Responses file inputs | PDF inputs on vision-capable models use extracted text and page imagery | No independently tested ranking, cross-vendor parity, universal plan access or current price is inferred. |
| Gemini API PDF document processing | Google documents native PDF vision with document understanding and extraction workflows | No independently tested ranking, cross-vendor parity, universal plan access or current price is inferred. |
| Anthropic tool flow | A model-issued tool call requires application-side handling and returned results | No independently tested ranking, cross-vendor parity, universal plan access or current price is inferred. |
These are documented capabilities or provider statements, not results from a hands-on comparative test. Unknown competitive measurements remain unknown.
Not all document capabilities are equivalent
OpenAI documents file inputs to the Responses API, including PDF processing that can expose both text and page image information to supported vision models. Google documents PDF understanding and structured extraction via Gemini API routes. An Anthropic deployment can participate in a document extraction pipeline using supplied text or visual input, but the specific supported PDF endpoint and model compatibility must be confirmed separately. Do not infer three-way feature parity from the fact that each supplier markets multimodal AI.
Define extraction as a contract
Require a JSON object with invoice_id, issuer, currency, subtotal, tax, total, invoice_date and page_reference fields. Values must be strings or null unless a reliable canonical parser defines the units. Add a validation rule that subtotal plus tax must reconcile with total after permitted rounding, but do not silently “correct” a number by calculation when the source says otherwise. Preserve bounding boxes or page references where supported; otherwise store a human-review flag.
A fair corpus
Build a version-controlled corpus containing digitally generated invoices, flattened scans, low contrast images, multi-page tables, handwritten notes and duplicate invoice numbers. Label a ground truth field map independently. Freeze the exact model ID, SDK version, input processing mode and schema. Report precision and recall by field, false-positive amounts, abstention rates, preprocessing time and retry costs separately. There are no hands-on evaluation results in this article.
Failure and privacy boundaries
A visually plausible header can trick a model into treating a sample invoice as a real one. Embedded instructions inside the PDF must be treated as untrusted content, not tool calls. Quarantine all files; limit integrations to read-only extraction until a validator approves the structured data. Before processing personal or financial data, confirm vendor terms, region, retention and subprocessors for the exact deployment. The public documentation above does not by itself answer those procurement questions.
Decision
Shortlist the API that accepts your required input types and provides a reviewable evidence trail, then test it against the same annotated corpus. Choose a conventional OCR-plus-rules path if the forms are highly standardized. Choose a vision model only if its error handling and auditability justify complexity. API price comparisons belong to dated AI 10 snapshots, not unsourced estimates.
Pricing, regions and comparable plans
Version-, region- and subscription-tier-specific availability is not established here. Product interfaces and entitlements can change. Access requires direct verification for the buyer’s country, account, contract and deployment as of the purchase decision. Consumer subscriptions, API billing, seats, credits and enterprise agreements must not be reduced to one “monthly price.” Link any future cost table to an AI 10 snapshot with currency, region, billing unit, effective date and source URL. No unverified numerical plan cost is included in this article.
Limits of the available evidence
The sources confirm the documented features listed above; they do not establish which provider produces better results on the models & apis tasks of document-automation engineers and auditors. No controlled product trials, comparative benchmark scores, account-level entitlements, universal language coverage, legal determinations or image rights assessments were conducted for this article. Any deployment must check the actual model version and market, record failures and obtain human approval before publication or procurement.
Sources and Further Reading
- OpenAI — File inputs. Accessed 2026-10-25.
- Google — Document understanding: Gemini Generate Content API (Legacy). Accessed 2026-10-25.
- Anthropic — How tool use works. Accessed 2026-10-25.