ChatGPT is a general-purpose large language model developed by OpenAI (San Francisco). Claude is a general-purpose LLM developed by Anthropic (San Francisco). Gemini is a general-purpose LLM developed by Google (Mountain View). All three read documents through multimodal models, and all three can return output constrained to a JSON schema you define.
The honest version of this comparison is narrower than it was two years ago. These models extract well, and anyformat uses LLMs under the hood. The question is not whether a model can read an invoice. It is what surrounds the model when a wrong value costs money: where the value came from, how likely it is to be right, who reviews it when it is not, and under whose jurisdiction the document was processed.
Consumer tier and API tier are different products
Most of the confusion in this comparison comes from treating "ChatGPT" and "the OpenAI API" as one thing. They are not, and the difference is a data-governance one.
The consumer apps — ChatGPT, Claude.ai, the Gemini app — retain conversations by default, and depending on plan and settings that content may be used to improve models. Uploading customer documents there is a governance decision, not a technical one.
The API and enterprise tiers act as processors under a DPA with standard contractual clauses, do not train on submitted content by default, and offer approval-gated zero-retention modes. All three vendors' platforms hold ISO 27001 certification.
So the blanket claim that LLMs are not compliant is wrong. The narrower claim — that consumer chatbots are the wrong place for production documents — holds.
The provenance gap
When a model returns an invoice total, it returns a number. It does not return the coordinates that number was read from. Some vendors return citations to a source passage in text; none bind an extracted field to a bounding box on the page image.
Production document processing requires auditability. When an extracted value flows into your ERP and triggers a payment, auditors, compliance officers and downstream systems all need to verify where that value originated.
anyformat attaches a visual citation to every extracted field: click the field, land on the exact position in the source document. That is what makes AI decisions auditable under EU regulations like ViDA, DORA and country-specific e-invoicing mandates.
Fluency is not calibration
LLMs are fluent. They give confident-sounding answers even when they are uncertain. A model might extract a PO number and present it with the same apparent certainty as an invoice total, even though it is guessing from a blurry scan.
Token log-probabilities are not a substitute. They describe how sure the model was about the next token, not how likely a field is to be correct, and they are not calibrated against outcomes. So there is no threshold you can set and no field you can safely auto-approve on that basis. This is the silent failure problem: a wrong value presented confidently passes every downstream check until reconciliation fails.
anyformat assigns a calibrated confidence score to every extracted field, measured at 99.1% calibration accuracy with an Adaptive ECE of 0.009 (anyformat benchmark, 2026). Calibration is what makes a threshold mean something: the 98% field flows through automatically, the 62% field routes to a reviewer.
Constrained decoding closed this gap. Structured Outputs, tool schemas and response schemas make output structurally deterministic — the same field names, the same nesting, the same types, every run. Anyone still describing LLM output as structurally unpredictable is describing 2023.
What structural validity does not tell you is whether the value is right. A well-typed field can be confidently wrong, and a schema will not catch it. That is what evaluation and calibrated confidence are for, and neither ships with a model API.
anyformat ships an evals suite: run a labelled set, get per-field accuracy, catch regressions before they reach production, and compare pipeline changes against a fixed baseline rather than a vibe check.
Cost at scale
LLM token pricing makes ad-hoc extraction cheap. At production volume, the arithmetic changes: cost scales with page count, with retries, and with every re-prompt spent coaxing a stubborn document into shape. The larger line item is usually engineering — prompt maintenance, output validation, retry logic and the review tooling you end up building anyway.
anyformat bills in credits, per page and per operator: 10 credits cost €0.01, so Parse is 25 credits per page (€0.025) and Extract 35 credits per page (anyformat pricing, 2026). The number to compare is cost per correct output, not cost per token.
European sovereignty and jurisdiction
This is the dimension where nothing has changed. OpenAI, Anthropic and Google are US-headquartered. EU processing regions and data-residency options exist, but the vendor operating the service remains under US jurisdiction, and that is what European sovereignty requirements test — not where the datacentre sits. In the feature matrix above, EU sovereignty is the row where every competitor scores no.
anyformat is EU-native. European entity, European infrastructure, ISO 27001 certified, GDPR-compliant by architecture, with zero-retention processing and on-premise or air-gapped deployment on every plan.
Workflow orchestration and human review
Model APIs process one document at a time, one call at a time. Classification, splitting, routing, validation, review queues, conditional logic and downstream integration are custom engineering you build and maintain — and the review interface, the part nobody scopes, is usually the largest piece.
anyformat includes Studio, a no-code workflow canvas with branching, conditions, splitting, routing, extraction operators, cross-referencing against internal data sources and human-in-the-loop review built in. Operations teams change the process without a code deploy.
Multimodal models read tables and charts competently, and paired with a response schema they will return rows as typed JSON. The remaining failure modes are the operational ones: tables spanning dozens of pages, where positional bias drops rows in the middle; merged cells and multi-page spans; and the absence of any per-cell provenance to check a suspect row against.
anyformat's multi-stage pipeline held ~99% row recovery on tables running to 50 pages and roughly 2,400 rows (anyformat benchmark, 2026), and returns figures as classified, schema-bound descriptions with citations. These are engineering problems you cannot solve by prompting alone.
The right mental model
LLMs are the engine inside modern document processing platforms, anyformat included. Using a raw model API for production extraction is like buying a diesel engine and calling it transport. The engine is excellent. It is not a vehicle.
anyformat is the vehicle: the model wrapped in calibrated confidence, visual citations, workflow orchestration, human review, evals and EU-native compliance controls.
For production, yes — but not for the reasons this comparison used to give. The model APIs now handle schema enforcement, table reading and chart interpretation, run on ISO 27001-certified platforms, sign processor DPAs and offer zero data retention on the API tier. What they do not provide is a calibrated confidence score per field, a visual citation binding each value to its position on the page, a workflow layer with human review, an evals harness to measure accuracy on your own documents, or an EU-jurisdiction option. anyformat uses LLMs as its extraction engine and supplies exactly those five things. If you are answering ad-hoc questions about a handful of documents, a raw model is the right tool. If you are running thousands a day under European compliance obligations, the surrounding infrastructure is the product.
When to use a raw LLM
Prototyping, ad-hoc questions, one-off analysis. Use a raw model when you are exploring, not operating.
When documents hit production: every value traceable to its position on the page, every uncertain field routed to a reviewer, accuracy measured rather than assumed, and the whole thing under EU jurisdiction.
anyformat is the agentic document intelligence platform built for European enterprises. ISO 27001 certified, GDPR-compliant, with zero-retention processing and on-premise deployment. Get started at anyformat.ai
