More Dials, More Questions: Configuring Claude for Dependable Work

By Jacob Hodges

When Claude Opus 4.8 was released last Thursday, the headline was honesty rather than raw capability. Anthropic described it as its most candid model yet, more willing to flag its own uncertainty and, according to its internal coding measure, roughly four times less likely than Opus 4.7 to let a flaw in its own work pass unremarked. That last figure is specific to code review, but the direction matters to anyone who relies on the model for work that must hold up.

For lawyers, this update lands on a tool many are already folding into their practice. The Opus line became the frontier model legal users increasingly gravitated to starting with Opus 4.6 in early 2026, and the dedicated “Claude for Legal” push in May only accelerated that.

The 4.8 release also shipped a new product feature: an effort control giving users a dial over how hard the model works on a given prompt. That dial is one more setting on top of the several that lawyers were already managing (e.g., Model choice, Research). Now, with more settings than ever, which combination produces more dependable work?

The problem

More settings mean more ways to be both over- and under-configured, with each lever labeled from a seeming cloud of near-synonyms for “the AI works more.” Crank everything to maximum and you burn through usage limits with little insight into any gains achieved. Dial the mix up wrong and credibility quietly degrades, perhaps manifesting as a confident-sounding answer drawn from stale training data, or a “reviewed” conclusion that rested on a partial read. Against this backdrop, the goal is the mix that best reduces the risk of an error slipping past review without spending capacity for marginal returns.

A word on what no setting changes: the lawyer must stay in the loop. Even the most honest model still errs and can sound confident while doing so, so the aim is a better-calibrated, better-sourced first pass, not a substitute for review.

The settings

Model choice comes first because it is the threshold decision, generally setting the ceiling every other setting operates under. It applies to all work, including where the toggles below may be irrelevant. For the highest confidence results, lawyers are likely best served using the most capable model available, which today is Opus 4.8. Beyond raw capability, Anthropic positions Opus 4.8 as tuned to more often surface uncertainty and catch its own errors rather than declaring victory early, a behavior that increases in value with the cost of an unflagged mistake.

Effort is the newest dial, and it is where Anthropic’s own documented guidance is clearest. Claude offers five settings, from Low to Max. High is the default, and Anthropic frames it as the balanced choice that still engages the model’s deep reasoning. Extra is suited to genuinely complex, multi-part work where thoroughness matters more than speed. Max, per Anthropic’s guidance, rarely justifies its cost on typical workloads; it adds little beyond slower, pricier output and can even prompt overthinking on structured tasks. More effort is not automatically better.

One clarification of a common confusion: the “Adaptive thinking” toggle lets the model decide whether and how much to reason before answering, while Effort guides how deep that reasoning goes. In practice, leave Adaptive thinking on and let Effort set the depth.

Web search is plausibly the single biggest lever available whenever a claim needs to be grounded in a source or checked against current information and is best turned on first; leaving it off is where confident-sounding errors most often appear, because the model is then answering from patterns learned during training, frozen at a cutoff. Web search helps produce citations and verify claims with current sources. In contrast, for the review of materials supplied directly, it does little; there, credibility hinges on the model’s own calibration and on effort.

Research is the heavier, agentic cousin of Web search, and it requires Web search to be on. Research runs many web queries that build on one another and then synthesizes the results into a thorough, cited review of its investigation. Research’s increased search breadth can also mean increased risk. Pulling in and synthesizing many sources gives Research more chances to absorb a weak one, and its synthesis, even in Opus 4.8, can fail to distinguish a claim relying on a single shaky source from well-founded ones.

Choosing to enable Research is less about the difficulty of a prompt than its scope. A narrow question that likely resolves against a single authoritative source, such as what a rule currently says or the latest version of a standard, is usually better served by just a Web search. Conversely, Research normally earns its added cost when there is a need to survey broadly, querying across many sources and aggregating what it finds.

Projects and custom instructions are where standing expectations belong. Anything you would otherwise re-type every session is better encoded once as project or profile instructions, so those instructions then automatically apply to each use within either the project or the Claude product, respectively.

Finding the mix

Each of these settings may impact how dependable a draft is before it reaches your review. Understanding each can help a legal user find the mix where a setting’s incremental “cost” in slower responses and consumed usage pays off with a more dependable result.

This article was created by Jacob Hodges, Member, Fulton Jeang PLLC, with the assistance of Claude Opus 4.8 and is for informational purposes only. It reflects on the perspective and experience of its author and does not constitute legal, or expert technical advice.