Claude code effort models

Skill ucsandman/claude-code-capability-primer/skills/claude-code-effort-models

Claude Code plugin: injects a capability self-awareness card at session start so Claude actually knows and uses its built-in capabilities — skills, subagents, dynamic workflows, hooks, MCP, plugins, GitHub Actions, the Agent SDK, and more.

Install

npx -y skills add ucsandman/claude-code-capability-primer --skill claude-code-effort-models

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when deciding how to work on a task - choosing effort level, model, or fast mode to balance reasoning depth, speed, and token cost.

SKILL.md

7.6 KB, ~2.0k tokens by cl100k_base, as published. Nobody here has run it

Claude Code Effort, Models, and Fast Mode

Quick reference for tuning how Claude Code works: model selection, effort levels, and speed mode. Read this mid-session when deciding how to approach a task.

Models via /model

Switch with /model (opens picker) or /model <name> to set directly. Persists to next session.

Available models (aliases resolve to the latest version for your provider; on the Anthropic API opus→Opus 4.8, sonnet→Sonnet 5):

opus → Claude Opus 4.8 (strongest Opus-tier reasoning)
sonnet → Claude Sonnet 5 (daily coding, best speed/intelligence balance; native 1M-token context on the Anthropic API — no [1m] suffix needed) — requires Claude Code v2.1.197+
haiku → Claude Haiku 4.5 (fast, simple tasks)
fable → Claude Fable 5 (most capable; for tasks larger than one sitting — long autonomous sessions, root-cause debugging, architecture) — requires v2.1.170+
best → Fable 5 where your org has access, otherwise the latest Opus
opusplan → opus during planning, auto-switches to sonnet for execution
opus[1m] / sonnet[1m] → 1M token context window (no effect once sonnet already resolves to Sonnet 5's native 1M window)

Model IDs (full names):

claude-fable-5
claude-opus-4-8
claude-sonnet-5
claude-haiku-4-5-20251001 (alias claude-haiku-4-5)

Default model (via default / picker): Opus 4.8 on Max, Team Premium, Enterprise pay-as-you-go, and Anthropic API; Sonnet 5 on Pro, Team Standard, and Enterprise subscription seats. Fable 5 is never a default — select it explicitly.

When to pick:

Fable: your hardest, longest-running work — hand it the outcome and let it plan; it investigates and verifies with less prompting. Note: requests flagged by its cybersecurity/biology safety classifiers auto-fall back to Opus.
Opus: hard reasoning, architecture decisions, complex refactors, code review, debugging. Higher token cost.
Sonnet: most coding work, features, edits, tests. Good speed/capability tradeoff; 1M context suits large codebases.
Haiku: searches, simple formatting, routine tasks. Cheapest.
opusplan: complex feature design where planning matters but execution is straightforward.

Effort Levels via /effort

Controls adaptive reasoning depth per message. Raise it for complex problems; lower it for routine tasks. Persistent across sessions unless overridden by env var.

Available levels (varies by model):

low — minimal thinking, fastest, cheapest. Use: latency-sensitive, low-complexity tasks.
medium — lighter reasoning, cost-conscious work that trades some intelligence.
high — default on Fable 5, Sonnet 5, Opus 4.8, Opus 4.6, Sonnet 4.6. Balances tokens and capability.
xhigh — deeper reasoning, higher token spend. Default on Opus 4.7. Use: tricky architecture, intricate bugs.
max — deepest reasoning, unbounded tokens, session-only. Can overthink; test first.

Fable 5, Sonnet 5, Opus 4.8, and Opus 4.7 support low–max. Opus 4.6 and Sonnet 4.6 support low/medium/high/max only (no xhigh — it falls back to high). Setting an unsupported level falls back to the highest supported at or below it.

Special: /effort ultracode (Opus only, session-only) sends xhigh to model AND orchestrates dynamic workflows for substantive tasks. Reserved for ambitious multi-phase work.

Usage:

/effort — open slider picker
/effort high — set directly
/effort auto — reset to model default
Env: CLAUDE_CODE_EFFORT_LEVEL=xhigh
Skill frontmatter: effort: xhigh

Token tradeoff: low < medium < high < xhigh < max. Each step costs more tokens but enables deeper reasoning for complex tasks.

Fast Mode via /fast

Opus only. Same model quality, ~2.5x faster output, higher cost per token. Toggle with /fast or "fastMode": true in settings.json.

Pricing (per MTok):

Opus 4.8 fast: $10 input / $50 output (vs $5/$25 standard)
Opus 4.7 fast: $30 input / $150 output (vs $5/$25 standard)
Opus 4.6 fast: $30 input / $150 output (deprecated; migrate to 4.8 or 4.7)

How it works:

NOT a different model. Same Opus, different API config prioritizing latency.
Auto-switches you to Opus if on Sonnet/Haiku.
Persists to next session by default (admins can set per-session reset).
↯ icon shows it's active.
Shares rate limit pool across Opus 4.8/4.7/4.6; auto-falls back to standard speed if rate-limited.

When to use:

Rapid iteration, live debugging, tight deadlines. Cost matters less than speed.
NOT good for long autonomous tasks, batch work, CI/CD.

Cost gotcha: enabling fast mode mid-conversation re-caches full history at fast-mode price. Enable at session start for best cost.

Requirements:

Anthropic API or Claude subscription (Pro/Max/Team/Enterprise) with usage credits enabled.
NOT available on Bedrock, Vertex, Foundry, or AWS Platform.
Team/Enterprise admins must explicitly enable it; disabled by default org-wide.

Token and Cache Angle

Model choice drives cost (per MTok input/output, Anthropic API):

Haiku 4.5 — $1/$5. Cheapest; searches/format work.
Sonnet 5 — $3/$15 standard (introductory $2/$10 through Aug 31, 2026). Most coding; 1M context at standard pricing (no premium past 200K).
Opus 4.8 — $5/$25. Hard reasoning; justifiable when depth pays off.
Fable 5 — $10/$50. Reserve for the hardest, longest tasks.

Effort and fast mode interact:

Higher effort = more tokens on same model.
Fast mode = same tokens, higher per-token price, faster latency.
Combine lower effort + fast mode for max speed on straightforward tasks.
Don't raise both for cost-sensitive work.

Prompt caching: a stable warm prefix re-reads at ~10% of input price. Keep the prefix stable (don't rewrite early messages mid-session). Switching /model mid-session invalidates the cache for the next turn.

Quick Decision Tree

Task is routine (format, search, simple edit)? → haiku, low effort, standard mode.

Task is typical coding (features, tests, refactors)? → sonnet, high effort (or medium to cut cost), standard mode.

Task is hard (architecture, complex bug, design)? → opus, xhigh effort, standard mode.

You need output in seconds, not minutes? → Use fast mode on Opus (higher cost, lower latency). Not a model change.

You need the deepest reasoning on an ambitious task? → opus, max effort, standard mode. Session-only, unbounded tokens.

Environment Variables

ANTHROPIC_MODEL=<name> — set model for this session only.
CLAUDE_CODE_EFFORT_LEVEL=<level> — effort level; overrides session choice.
CLAUDE_CODE_DISABLE_FAST_MODE=1 — disable fast mode entirely.
ANTHROPIC_DEFAULT_OPUS_MODEL / ANTHROPIC_DEFAULT_SONNET_MODEL / ANTHROPIC_DEFAULT_HAIKU_MODEL / ANTHROPIC_DEFAULT_FABLE_MODEL — pin specific model versions for each alias (useful on Bedrock, Vertex, Foundry).
CLAUDE_CODE_DISABLE_1M_CONTEXT=1 — remove 1M variants from the picker; treats Sonnet 5 sessions as 200K.

Fallback models: set fallbackModel (settings, array) or --fallback-model sonnet,haiku to try backups in order when the primary is overloaded/unavailable. Chain capped at 3; lasts the current turn only.