Skip to main content

Docs

Effort Levels

Every K‑Dense Web run happens at one of three effort levels. The level decides how much compute the agent spends, whether it can revise its own plan, and how much of the output gets independently checked before you see it. It is the single biggest lever on output quality, and the one most new users miss.

You choose it with the effort selector before you start a run.

If you only read one thing

Use Standard for most work. Move up to Pro when the output has to hold up — a manuscript, a decision memo, anything with a number in it someone will act on — or when the task has steps that depend on what earlier steps find. Use Instant when you want an answer, not a research run.

Instant

A direct answer, the way a chat assistant would give it. No code, no execution plan, no research run.

Best for

  • Quick factual questions
  • Asking what the platform can do before committing to a run
  • Follow-up questions about work you already have

What it does

  • Answers straight from the model, in the conversation
  • Does not write or run code
  • Does not search databases or pull papers
  • Produces no files

Validation

None — this level is not doing research to validate.

Typical runtime: SecondsTypical cost: Free

Standard

Writes and runs code

A full research run from a single execution plan. The workhorse for most tasks.

Best for

  • Literature reviews and syntheses
  • Analysis of a dataset you upload
  • Figures, reports, and slides from a clear brief
  • Anything where you already know what you want

What it does

  • Asks clarifying questions, then builds one execution plan
  • Writes and runs code, queries databases, generates figures
  • Executes the plan start to finish without revising it
  • Produces downloadable deliverables

Validation

Moderate. Code is checked and sources are grounded, but the deeper review passes are skipped to keep runs faster and cheaper.

Typical runtime: Under 5 minutes to a few hoursTypical cost: Capped at $9 per run

Pro

Writes and runs codeRevises its plan

Maximum compute, full validation, and a plan the agent revises as it learns. For work you intend to publish, submit, or act on.

Best for

  • Manuscripts and publication-ready figures
  • Due diligence and decision memos
  • Multi-step computational work where early steps change later ones
  • Anything you would otherwise pay a specialist to produce

What it does

  • Builds an execution plan and revises it when steps fail or findings redirect the work
  • Runs the full validation suite over code, sources, and methodology
  • Iterates on output quality rather than stopping at first draft
  • Runs substantially longer than Standard

Validation

Full. Separate agents review the code, verify every citation resolves to real work that supports the claim, and critique the scientific methodology — then the system acts on that feedback.

Typical runtime: 30 minutes to several days, depending on the taskTypical cost: Capped at $29 per run

Side by side

CapabilityInstantStandardPro
Writes and runs codeNoYesYes
Revises its plan mid-runNoNoYes
Validation depthNoneModerateFull
Typical runtimeSecondsUnder 5 minutes to a few hours30 minutes to several days, depending on the task
Typical costFreeCapped at $9 per runCapped at $29 per run

Within a tier, cost scales with the work a run actually does — the figures above are the ceiling, not a flat fee. See pricing for credit rates.

Two mistakes worth avoiding

Running complex work at Standard

Standard commits to one plan up front. If you give it a wrong accession number, or an early analysis changes what the right next step is, it will carry on with the original plan rather than correct course. That reads as the model being stubborn when it is really the effort level doing what it is designed to do. Pro revises the plan; use it when the path is not fully known at the start.

Judging output quality without checking the level

The gap between Standard and Pro on a hard task is large, because Pro is where the citation and methodology reviewers run. Before concluding that a result is weak, check which level produced it — re-running the same prompt at Pro is often the whole fix.

Getting more out of any level

Effort level sets the ceiling; your prompt decides how close you get to it. The prompting guide covers the six things worth specifying before you hit go.

Read the prompting guideAll docs