Anthropic has released its clearest guidance yet on when to upgrade to a larger AI model versus simply working the existing model harder. On July 7-8, 2026, Lydia Hallie, a member of the technical staff on the Claude Code team, published a framework that answers a question every developer using Claude Code asks: should I switch to a more powerful model like Opus, or just increase the effort level to get better results ?

What's the Difference Between Model Size and Effort Level?

The distinction matters because they solve different problems. A model swap changes which set of frozen weights, or learned parameters, handle your request. Effort, by contrast, controls how much work those weights perform before checking back in with you. Higher effort means Claude reads more files, runs more tests, performs deeper verification, and loops through more tool calls before presenting results. This is fundamentally different from "thinking longer," Hallie explained.

When you press enter in Claude Code, everything packs into a single API request: your system prompt, available tools, context files, conversation history, and any files you mention. On Anthropic's servers, text converts to integer IDs that the model processes one token at a time. A 200-token response requires 200 separate forward passes through the model's weight stack. The model never sees raw strings; it only sees probability distributions shaped by your prompt.

When Should You Upgrade Your Model Instead of Just Trying Harder?

Anthropic's prescribed sequence for troubleshooting disappointing output follows a clear hierarchy. First, fix context. If your prompt is vague, you're using the wrong tools, or your task scope is poorly defined, no amount of model power or effort will help. The solution is better context: a more specific prompt, a well-written CLAUDE.md file, the right Model Context Protocol (MCP) servers, or clearer task boundaries.

Only after context is optimized should you consider raising effort. If Claude skipped steps despite having good context, increasing effort makes sense, especially if you're currently below the default setting. Finally, if Claude tried thoroughly and still failed, that's when a larger model becomes necessary.

The inverse rule applies too: if you're running routine edits on a larger model like Opus for extended periods, dropping to Sonnet at default effort saves money without sacrificing quality on mechanical work.

How to Choose the Right Model and Effort Combination

• Routine edits and code already in context: Use Sonnet 4.6 at default effort. This handles mechanical work efficiently without paying premium prices for renames and simple changes.
• Tried hard, still confidently wrong: Switch to a larger model like Opus 4.8 or Fable 5. If Claude made thorough attempts and still failed, the issue is capability, not effort.
• Simple task, overthinking: Lower effort or use a smaller model. Don't pay Opus prices for work that doesn't require it.
• Vague prompt, wrong tools, bad scope: Fix context first. Better prompts, CLAUDE.md files, MCP servers, and task clarity solve more problems than model or effort adjustments.

Anthropic's default rule is straightforward: use each model's default effort for most tasks. Tune effort as a general preference, not task-by-task. The team trains against artificial overthinking on simple tasks, so higher effort doesn't mean endless loops; it means more thorough verification before declaring a task complete.

The metaphor Anthropic uses is instructive. Low effort on Opus is like a five-minute consultation with an expert: pattern recognition without reading every file. High effort on Sonnet is like a generalist spending the afternoon on your codebase: exhaustive, thorough, reading everything and running tests. Fable, the newest and most expensive model, finishes jobs that Opus and Sonnet cannot complete at any effort level in Anthropic's testing.

What Changed With Opus 4.8 and New Effort Tiers?

Anthropic reports that default effort on Opus 4.8 matches roughly the same token spend as Opus 4.7 default but delivers better results on the same work. This means you're not paying more for the same throughput; you're getting better quality at the same cost.

On routine work at the same effort level, both Sonnet and Opus succeed, but the larger model adds extra verification tokens at a higher per-token price. On hard multi-step work, smaller models grind toward their capability ceiling, requiring many iterations and high total tokens, while larger models hit quality in fewer steps. Total cost per task can actually favor the larger model even at higher unit price. More importantly, only the larger model may finish the job at all.

Fable pulls furthest ahead on long-horizon work, aligning with loop engineering and Goal command patterns. The key insight is that none of these models is universally "better." Model choice determines capability; effort determines thoroughness.

How to Adjust Model and Effort in Claude Code

• In-session command: Type "/effort high" (or low, medium, max, xhigh where supported) to adjust effort mid-conversation without restarting.
• CLI flag: Use "claude --effort max" when launching Claude Code from the command line to set effort for the entire session.
• Settings file: Add "effortLevel": "high" to your.claude/settings.json file to make effort changes persistent across sessions.
• Prompt keyword: Use "ultrathink" for one-off deep reasoning tasks as a community pattern that pairs high effort with standing permission for multi-agent workflows.

Defaults vary by model and version. Opus 4.8 and Sonnet 4.6 default to high effort on Claude Code, while Opus 4.7 historically defaulted to xhigh. Always verify your version's defaults in Claude Code settings.

Why This Matters Beyond Just Saving Money

The framework Hallie published reflects a deeper shift in how developers think about AI-assisted coding. Rather than treating model upgrades as the primary lever for better results, Anthropic is encouraging developers to think systematically about context, effort, and capability as separate, stackable problems. This distinction matters because it prevents unnecessary spending on model upgrades when the real issue is task decomposition or prompt clarity.

The same day Anthropic shipped Claude Reflect, a usage dashboard with AI Fluency scores and quiet hours for chat users with Memory enabled. Reflect measures how much and how developers delegate work to AI; the model-versus-effort framework measures capability and thoroughness per task. Together, they give developers both macro visibility into their AI usage patterns and micro control over individual task execution.

Developers should audit whether failures stem from LLM burnout from reviewing poor-quality work. Effort and model cannot fix bad task decomposition. Pair effort increases with verification hooks, tests in continuous integration, and thorough review processes so "try harder" means running tests, not writing more prose.

The practical takeaway is permission to stay on Sonnet plus default effort for mechanical work and reserve Fable for genuine stretch tasks that smaller models cannot complete. Don't flip model and effort on every failed diff. Instead, treat Hallie's article as a decision tree that prioritizes context fixes first, then effort adjustments, then model upgrades.