Which Claude model and effort level should I use in Claude Code?

Default to Opus 4.7 at xhigh effort. Most of the real decisions are about when to leave it.

The decision in 5 steps

1. Set Opus 4.7 + xhigh as your baseline.

/model opus
/effort xhigh

Or in ~/.claude/settings.json:

{
  "model": "opus",
  "effortLevel": "xhigh"
}

As of April 23, 2026, default resolves to Opus 4.7 on Max, Team Premium, Enterprise pay-as-you-go, and the Anthropic API. Pro, Team Standard, and most subscription Enterprise tiers still default to Sonnet 4.6. Bedrock, Vertex, and Foundry default to Sonnet 4.5 unless you pin a newer model. Set opus explicitly if you want Opus and aren’t sure which tier you’re on.

2. Drop to Sonnet 4.6 for routine work.

Switch when the task is mechanical:

/model sonnet

• Renaming, find-and-replace, mechanical refactors
• Writing tests for code you’ve already shipped
• Markdown / config / docs edits
• Any batch or CI usage where latency stacks up

Sonnet 4.6 is $3/$15 per MTok versus Opus 4.7 at $5/$25, so about 1.7x cheaper per token. The bigger win is throughput: Sonnet runs noticeably faster, and on a long task that compounds. Stay on it whenever the task isn’t intelligence-bound.

3. Drop effort, not model, when the task is small.

/effort medium

medium is right for short, scoped edits where you’d waste tokens on xhigh deliberation. The effort scale is calibrated per model, so medium on Opus 4.7 is not the same as medium on Sonnet. Both lower the spend, neither makes the model dumber on the actual code part.

4. Use opusplan when planning is most of the work.

/model opusplan

opusplan runs Opus during plan mode, then switches to Sonnet when you exit plan and execute. Use it when the architecture is the hard part and the implementation is mostly typing. Don’t use it when the implementation has subtle invariants (concurrency, migrations, anything where Sonnet might miss an edge Opus would catch).

5. Use Fast mode for live debugging only.

/fast

Fast mode runs Opus 4.6 at 2.5x speed for 6x the per-token cost ($30/$150 per MTok versus standard Opus 4.6 at $5/$25). Toggle it on at the start of a session, not mid-conversation. It’s billed entirely to extra usage from token one. Your subscription doesn’t cover it. Fast mode is not available on Bedrock, Vertex, or Foundry.

Footguns

Effort levels silently downgrade across models. If you set xhigh and then run /model sonnet, you get high. Sonnet 4.6 doesn’t support xhigh. The session shows with high effort next to the spinner; it does not warn you that you asked for more. Always glance at the spinner after switching models.

Mid-session model switches invalidate your prompt cache. The /model picker prompts you for confirmation when there’s prior output for exactly this reason: the next turn re-reads the full conversation history without any cache hits. On a long session that’s a real bill. Switch at the start, not in the middle.

opusplan doesn’t get the 1M context window. If you’re using opusplan on a long codebase session, the plan-mode Opus phase runs in a 200K window even when your account supports 1M. To get 1M during planning, use /model opus[1m] directly, but then you stay on Opus through execution. There’s no opusplan[1m].

max effort on Opus 4.7 is a trap on familiar problems. Anthropic’s own docs say it’s “prone to overthinking”, which is polite for “burns tokens dithering on choices already made.” Use max only when an xhigh run produced a wrong answer that suggests the model didn’t think hard enough. If xhigh reasoned to a wrong conclusion confidently, max won’t fix it.

default resolves differently by account type. Max, Team Premium, Anthropic API, and Enterprise pay-as-you-go get Opus 4.7 (post April 23, 2026). Pro and Team Standard get Sonnet 4.6. Bedrock, Vertex, and Foundry get Sonnet 4.5. If you’re pairing with someone and tell them “just use default”, they may be on a much smaller model than you. Always say the model name out loud.

Fast mode + 1M context is a budget detonator. Fast mode pricing is flat across the full 1M window. A long session at $30/$150 per MTok adds up fast. Use Fast mode in fresh sessions or short ones; don’t combine with extended context unless you’re prepared.

Subagents inherit your model unless you tell them not to. Set CLAUDE_CODE_SUBAGENT_MODEL=haiku and your parallel research subagents stop running on the same Opus 4.7 you’re paying premium for. Per-subagent override goes in the subagent frontmatter:

---
name: research
model: haiku
effort: medium
---

When NOT to use this default

• Long autonomous CI runs. Stay on Sonnet 4.6 at high. Latency stacks across thousands of tool calls and Opus pricing on a 6-hour CI job will hurt.
• Vision-heavy work (dense screenshots, diagram analysis). Opus 4.7 is genuinely better here, but xhigh overshoots. Use high and save the difference.
• Pure content generation (README rewrites, doc drafts, commit messages). Use Sonnet at medium. Opus on prose is wasted budget.
• You’re on Bedrock/Vertex/Foundry and haven’t pinned a version. Pin ANTHROPIC_DEFAULT_OPUS_MODEL and ANTHROPIC_DEFAULT_SONNET_MODEL to specific IDs before rolling out. Aliases there default to older versions and silently fall back when new ones aren’t enabled.