Context capacity for the largest documents or multi-document batches you plan to send.
AI models for document extraction
Choose an extraction model for the documents and economics you actually have.
Document extraction is often less about buying the most capable frontier model and more about finding reliable structured output, enough context, acceptable error rates, and token economics that survive volume.
The decision
The answer changes with the workload.
A small batch of messy contracts, a million short receipts, and long regulatory filings are different model-selection problems. The right shortlist depends on document length, schema complexity, tolerance for retries, expected output volume, and whether tool or structured-output support is mandatory.
Structured-output and tool support when extraction must conform to a strict schema or downstream workflow.
Quality evidence appropriate to reasoning-heavy versus repetitive extraction tasks.
Input and output pricing at the actual document volume, including retry and validation overhead.
Independent performance evidence
Artificial Analysis
ModelShortlist uses Artificial Analysis benchmark and performance evidence when the model identity can be reconciled confidently. It does not create, relabel, or pretend ownership of those benchmarks.
How Artificial Analysis contributesCurrent operational facts
OpenRouter
The current catalog supplies model availability, context, supported parameters, pricing, and provider details. ZDR endpoint evidence is applied only when the workload explicitly requires ZDR.
Compare OpenRouter models by workloadModel selection is time-sensitive: new models launch, prices move, benchmark results change, tool support evolves, and providers add or remove endpoints. ModelShortlist surfaces freshness and degraded upstream state instead of silently presenting stale evidence as current.
Ask naturally
Prompts that work.
ModelShortlist sits behind the AI assistant you already use. Describe the job and hard constraints instead of translating them into a fixed ranking formula.
“I need structured extraction from thousands of long documents. Reliability matters, but output cost needs to stay low. What should I shortlist?”
“Compare current models for extracting a strict JSON schema from contracts up to 80k tokens each.”
“I have a high-volume OCR-to-LLM pipeline. Which current models give me the best balance of extraction quality and token economics?”
Why ModelShortlist
Current evidence, not a static leaderboard.
Starts from the current OpenRouter catalog rather than a fixed list of document models.
Can apply minimum context, tool support, creator, and price ceilings before the host AI weighs softer tradeoffs.
Uses Artificial Analysis evidence when the exact model can be reconciled confidently, without fabricating scores for unmatched models.
Keeps price and capability facts current enough for workloads where a small per-token difference multiplies across large document volumes.
Local BYOK architecture. No ModelShortlist account, hosted key vault, or telemetry in the MCP.
Put current model-selection evidence inside your assistant.
Install the local MCP, add your OpenRouter and Artificial Analysis keys, and ask the model-selection question in normal language.
Open the install configurator