Claude Code settings balance — speed × accuracy × cost (2026-06-29 pin)

rule Mon Jun 29 2026 00:00:00 GMT+0000 (Coordinated Universal Time) claude-codesettingsbalancespeedaccuracycostpreference

Claude Code settings balance — 2026-06-29 pin

Locked via 2026-06-29 grill-me session, 4 batches × 4 questions = 16 picks. Captures the "max-quality posture" since Bedrock-through-Hr is corp-paid.

The 12 picks

Setting	Value	Speed	Accuracy	Cost
`effortLevel`	`"high"`	8/10	8/10	4/10
`alwaysThinkingEnabled`	`true`	-1/10	+2/10	+1/10
Adaptive thinking	enabled (default)	+1/10	-0 (with always-thinking floor)	-1/10
`MAX_THINKING_TOKENS`	DELETED	0	0	0
`model`	`"claude-opus-latest"`	6/10	10/10	9/10
`CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`	`"85"`	+1/10	9/10	-1/10
`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS`	`"1"`	+3/10	+2/10	+3/10
`switchModelsOnFlag`	`true`	(off-axis)	(off-axis)	(off-axis)
`CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS`	DELETED (betas ON)	-1/10	+1/10	0
`showThinkingSummaries`	`true`	-0	+2/10 (catches bad paths)	0
`autoMemoryEnabled`	`true`	-0	+1/10	0
`ANTHROPIC_AUTH_TOKEN`	? `settings.local.json`	0	0	0 (security only)

Cost-positive number = burns more. Speed-positive number = faster wall-clock.

Rationale per axis

Speed posture: medium-high

Hr proxy compresses prefix. ToolSearch defers MCP tools out of cache. 85% compact reduces rebuild frequency. PowerShell tool over Git Bash. Pinned model + pinned effort = stable cache key. Betas ON costs a few ms of HTTP fanout — acceptable for new features.

Accuracy posture: max

Opus 4.8 daily. Always-thinking + adaptive thinking combined: floor>0, depth auto-tuned. effortLevel=high (pre-March-2026 Anthropic default; the level Boris Cherny says Claude was tuned for). Agent teams for architecture work. Thinking summaries visible.

Cost posture: don't count

Corp Bedrock via Hr means cost is downstream concern. Pick Opus over Sonnet. Pick agent teams ON. Pick 85% compact over 70% (fewer rebuilds = more cache reuse = NET cheaper despite longer prefix).

The exception: /model sonnet and /fast available per-task when the work is mechanical. Don't let "cost is downstream" become "always burn Opus on file renames."

Why these specific overrides matter

Three picks OVERRODE my initial recommendations:

model: claude-opus-latest over Sonnet — overrides 2026-06-27 default-model-sonnet-4-6 memory. New memory: default-model-opus-2026-06-29. Captures the corp-pays insight that earlier grill missed.
Adaptive thinking ENABLED over disabled — Marco Lancini + Pasquale Pillitteri both recommend disabling adaptive. User insight: combining alwaysThinkingEnabled=true with adaptive gives floor>0 + auto-tune, better than either alone. This is a novel position the cited articles don't cover.
Agent teams ENABLED — Marco's setup; conflicts with the delegate-to-subagents-by-default rule. Resolution: when user invokes a slash command (/grill-me, /code-review), explicit delegation wins. Agent teams handles the "free-form prose multi-step" case where the explicit delegation rule didn't fire. They compose; no double-spawn.

Anti-patterns this pin avoids

? Cargo-culting Marco Lancini's xhigh default — costs 2× without 2× benefit at the corp-Bedrock margin
? Disabling adaptive thinking globally — works for Marco (always-xhigh) but wastes tokens on trivial turns when paired with high effort
? 70% compact threshold — fires too early on long sessions; 85% is the Anthropic default for a reason
? Pinning MAX_THINKING_TOKENS — legacy budget_tokens API, ignored on Opus 4.7+
? Auth token in ~/.claude/settings.json — gets committed/synced. Move to settings.local.json (gitignored by CC).

When to re-grill

Anthropic changes default effortLevel (next regression)
Bedrock corp billing model changes (cost re-enters the equation)
Agent teams graduates from experimental or gets renamed
User stops using Hr (cost becomes personal)

Cross-refs

claude-code-latency-techniques — settings reference + cache model
default-model-opus-2026-06-29 — memory file capturing model choice
minimum-claude-settings — "keep settings.json minimal" — partially superseded; this pin has more keys
grill-me-default — the discipline that produced this pin
Upstream: Marco Lancini's 2026 setup, Pasquale Pillitteri effort guide, Boris Cherny HN comment, AMD Stella Laurenzo 6852-session analysis