What happened
On October 29, 2026, LlamaIndex announced Max Effort for Extract’s agentic_plus tier. Extract is the company’s document-extraction product: users provide documents and a schema, then receive structured fields from the files. The new setting is intended for critical and complex documents where a missing or incorrect value carries a higher cost than the extraction service itself.
The setting changes the commercial model for a job. Standard Extract tiers adapt their work while keeping a fixed price per page. With Max Effort, billing varies with the complexity of the extraction and the computation required. LlamaIndex says some documents may become substantially more expensive when extra work produces a large quality gain; when standard methods suffice, the increase can be marginal. The feature is currently marked beta.
The company’s published ExtractBench Pro results place standard Agentic Plus at 77.7% Value F1 and 7.5 cents per page, while Agentic Plus with Max Effort is listed at 94.3% Value F1 and 42.2 cents per page. ExtractBench Pro contains 100 documents. These are vendor-published benchmark results, but the accompanying category examples make the intended routing problem clearer. On seven long, dense single-page documents, LlamaIndex reports Value F1 rising from 12.7 to 95.0, with mean billed cost per document increasing from $0.075 to $4.19. For custom terminology, it reports a smaller change: 98.6 to 99.7 Value F1 and $3.25 to $3.37 per document.
LlamaIndex attributes part of the difference to visual material in PDFs. A chart may need a page rendered as an image before a model can read it. Under predictable standard pricing, the service makes image-resolution trade-offs that work for most documents. Max Effort can inspect relevant regions at higher resolution when needed. The release also shows different results for relational diagrams, document conventions, charts, and counting, suggesting that the additional compute is meant to be applied selectively rather than as a universal quality mode.
To use it, LlamaIndex says teams should upgrade llama-cloud and configure an extraction run with tier set to agentic_plus, effort set to max, and version set to latest. Its example retains the existing schema and options, including confidence scores and source citations. The company recommends testing a representative sample of existing documents and schemas before moving production workloads to the new setting.
What the archive adds
This is a new fact on a known story. Extract v2.5 framed LlamaIndex’s earlier release as schema-aware agents allocating reasoning and verification effort to document structure. Version 2.6 adds an explicit high-compute setting and variable billing to that same extraction surface.
Trend signal
The release supports Agent economics moves to completed work. A document-extraction buyer is not being offered a cheaper token or a faster model call; they are being offered a more expensive attempt at a correct field on an especially hard page. The operational unit becomes the verified extraction outcome and its cost, with document type and error consequences determining how much compute is justified.
Why it matters
Document pipelines often fail unevenly. A low-cost configuration can be adequate for clean invoices and still fail on a dense page, a diagram, or a visually complex chart. Max Effort gives teams a supported way to buy more work for the exceptional cases instead of rebuilding their entire flow around the most difficult file.
That flexibility also turns budgeting into an engineering responsibility. A fixed per-page estimate no longer describes every job. Teams need representative test sets, per-job cost observation, and routing rules that separate routine documents from costly exceptions. The useful question is not whether Max Effort has the highest published score, but whether the value recovered from fewer bad fields exceeds the additional spend for the document classes that trigger it.
What happens next
- Teams running Extract on heterogeneous PDF collections will add document-class or confidence-based routing before enabling Max Effort broadly. (3 months, 70%). Watch for: examples of selective production routing or budget controls.
- LlamaIndex will publish more guidance or controls for forecasting variable-cost extraction jobs. (6 months, 55%). Watch for: cost caps, estimates, or job-level billing dashboards.
Limits and uncertainty
The quality and price figures come from LlamaIndex’s own ExtractBench Pro presentation, not an independent evaluation, and they may not transfer to a team’s documents, schema, error definition, or production latency requirements. The release gives mean costs for benchmark categories, but it does not establish a maximum cost for a Max Effort job or a general rule for predicting which files will require extra work. It also does not show how the setting compares on every real-world document type. Teams should treat the published numbers as a reason to run a controlled sample, not as a deployment forecast.
Skill to build now
Create a small evaluation set from the documents that currently cause the most manual review. Keep the production schema unchanged, run it through standard Agentic Plus and Max Effort, and record field-level correctness, confidence, cited sources, latency, and billed cost. Group results by document type: charts, diagrams, dense pages, and routine forms. Then define a routing rule such as “use Max Effort only after low confidence on a high-value field” or “only for a named document class.” This turns a variable-compute feature into a bounded operational policy.
Reader takeaway
Extract v2.6 makes high-accuracy document extraction a variable-cost option rather than a single fixed tier. The valuable deployment pattern is selective escalation: measure hard documents against your schema, then spend additional compute only where the avoided extraction error is worth more than the added job cost.
Open question
What job-level caps, estimates, or routing signals will LlamaIndex provide so teams can use variable-compute extraction without exposing production pipelines to unpredictable document costs?