As of 2026-05. A skeptic-disarmer companion to decision-memes.html. Every entry below is a misread that drives a measurable mis-budget, mis-architecture, or mis-staffing call. If a myth doesn’t change a decision, it’s not in this list.
How to read this. Format: Myth → Reality → What you’d mis-decide → Cite. Pinned to the current model surface (Opus 4.7 / Sonnet 4.6 / Haiku 4.5). Refreshed monthly against docs/feature-inventory.md — the single source of truth.
1. Model + capability misreads
“Claude’s context window is still 200K tokens.”
- Reality. Opus 4.7, Opus 4.6, and Sonnet 4.6 ship 1M tokens (GA). Sonnet 4.5 and Haiku 4.5 remain at 200K.
- Mis-decide. Architecting RAG chunking + retrieval ceilings against 200K when the workhorse model already supports 1M — over-engineered retrieval, under-used cache.
- Cite. docs.claude.com — context windows;
docs/feature-inventory.md Models table.
“Bigger context = better answers, so always use 1M.”
- Reality. Long-context recall quality varies by position and task. Stuffing the window often hurts both latency and answer quality vs. retrieval + caching. Use the window deliberately, not by default.
- Mis-decide. Skipping retrieval design entirely; paying for tokens that erode answer quality and inflate latency.
- Cite. docs.claude.com — context windows;
feature-decision-matrix.html RAG row.
2. Claude Code (CLI) surface
“Claude Code is just a chat wrapper for the terminal.”
- Reality. It exposes hooks (PreToolUse / PostToolUse / SessionStart / Stop / UserPromptSubmit), a settings hierarchy (user/project/local), Skills, subagents, plugins, MCP, headless / CI mode, and scheduled background tasks.
- Mis-decide. Treating the CLI as a developer toy; under-funding rollout. Engineering managers who skip the platform surface ship per-engineer plugins (“5 engineers, 5 plugins, 0 reuse” — see
claude-code-adoption-guide.md failure modes).
- Cite. docs.claude.com — Claude Code;
docs/feature-inventory.md Claude Code surface table.
“Custom slash commands and Skills are different systems.”
- Reality. In the 2026 line, custom commands were unified into Skills. Existing
.claude/commands/ files keep working, but Skills add frontmatter-based auto-invocation and bundled-file management on top. Treat new work as Skills.
- Mis-decide. Building parallel infrastructure for both; missing auto-invocation triggers; team Skills repo splintered across two folder conventions.
- Cite. docs.claude.com — slash commands;
claude-code-starter-skills.md.
“Hooks just run shell scripts.”
- Reality. Hooks support shell commands, HTTP endpoints, and LLM prompts as actions, plus async hooks and MCP-tool-scoped hooks. Configure in
.claude/settings.json.
- Mis-decide. Forcing all governance through bash; missing the HTTP-endpoint pattern that lets central security own the hook body without per-repo bash sprawl.
- Cite. docs.claude.com — hooks;
hooks-starter-pack.md.
- Reality.
sandbox.denyRead blocks bash subprocesses only at the OS level. It does not stop Claude’s built-in Read tool. Block both layers: sandbox.denyRead for shells, plus a permission deny rule like Read(./.env) for the tool. Inversely, Read(./.env) does not stop cat .env from a hook or sub-shell.
- Mis-decide. False sense of security on secrets. The most common Week-1 governance gotcha — discovered in audit, not in design.
- Cite. docs.claude.com — sandboxing;
hooks-starter-pack.md block-secrets pattern; governance-overlay.md.
“Claude Code has a free tier.”
- Reality. Claude Code requires a paid Claude.ai subscription (Pro / Max / Team / Enterprise) or API credits. Usage limits scale by plan; Max plans buy more headroom, not unlimited.
- Mis-decide. Procurement assumes individual Pro covers an engineering team. Scale-out blocked when team usage limits hit mid-pilot.
- Cite. docs.claude.com — pricing;
cost-calculator.html.
3. API economics
“Pro / Max subscriptions include API credits.”
- Reality. Claude.ai consumer subscriptions (Pro, Max, Team, Enterprise) and Anthropic API credits are separate billing surfaces. Max does not unlock API quota. Buy API credits explicitly.
- Mis-decide. Engineering team billed twice — or worse, finance assumes the seat license covers production inference and is surprised by the API invoice.
- Cite. docs.claude.com — pricing.
“Prompt caching is too complex / not worth it.”
- Reality. Cache reads bill at ~10% of the input rate; 5-min writes 1.25×, 1-hr writes 2×. For chat copilots and RAG patterns with stable system prompts or shared retrieved context, steady-state cost drops 60–80%.
- Mis-decide. Skipping the dominant cost lever; the cost-calculator numbers in
executive-briefing.html cease to apply, and unit economics quietly fail at scale.
- Cite. docs.claude.com — prompt caching;
cost-calculator.html.
“Long context costs 2× once you cross 200K tokens.”
- Reality. The long-context surcharge has been eliminated on Opus 4.7, Opus 4.6, and Sonnet 4.6. A 900K-token request bills the same per-token rate as a 9K request on these models. Older / legacy models still apply the 200K break — verify per model.
- Mis-decide. Ruling out long-context patterns on cost grounds that no longer hold for the workhorse models.
- Cite. docs.claude.com — pricing;
docs/feature-inventory.md Models table.
“Batch API is just a queue — same price.”
- Reality. Batch API delivers 50% off all rates with a 24-hour SLA. Stackable with caching. Async by design — fire-and-forget for evals, enrichment, and regression suites.
- Mis-decide. Running overnight evals on real-time pricing; forgoing the cheapest path for every workload that does not need synchronous latency.
- Cite. docs.claude.com — batch processing;
eval-starter-pack.md.
“Haiku is too weak for production.”
- Reality. Haiku 4.5 prices ~5–15× cheaper than Opus and is the right tier for triage, classification, format conversion, and the cheap leg of cascade patterns. Production systems commonly run Haiku-first and escalate to Sonnet or Opus only when triage demands it.
- Mis-decide. All-Opus deployment for workloads where Haiku would meet the bar at a fraction of the cost — an Opus tax that compounds at scale.
- Cite. docs.claude.com — pricing;
feature-decision-matrix.html cascade row.
4. Safety + refusal
“Claude refuses everything sensitive.”
- Reality. Anthropic publishes Claude’s constitutional principles; refusal is calibrated against actual harm potential, not topic surface. Most reported refusals trace to prompt shape (no role context, no safety frame, no allow-listed scope) rather than the topic.
- Mis-decide. Ruling out viable use cases (clinical decision support, security research, content moderation) on a stale “Claude will refuse” prior — when the same workload runs cleanly with proper system-prompt scoping. See
anti-use-cases.md for use cases that genuinely should be ruled out.
- Cite. anthropic.com — Claude’s Constitution.
5. Rate limits + Computer Use
“Rate limits are just the dashboard percentage.”
- Reality. Three independent limits apply per organization + tier: requests-per-minute, input-tokens-per-minute, and output-tokens-per-minute. The dashboard surfaces the binding one. Hitting any single limit throttles the whole org for that model.
- Mis-decide. Sizing capacity off RPM only; throttled by token-per-minute headroom on a long-context workload that the RPM number didn’t predict.
- Cite. docs.claude.com — rate limits.
“Computer Use is a paid product feature on its own SKU.”
- Reality. Computer Use is an API capability available to API customers, not a separate purchase. Billed by tokens like any other inference. Sandbox + governance posture is the customer’s responsibility.
- Mis-decide. Procurement waits for a “Computer Use SKU” that doesn’t exist; pilot blocked behind a non-existent contract step.
- Cite. docs.claude.com — computer use;
governance-overlay.md prompt-injection section.
How this artifact connects to the rest
- Runs before
decision-spine.html. Disarms the priors that distort which branch the reader picks. A reader who believes “Claude refuses everything sensitive” exits at the anti-use branch when their use case is actually viable.
- Companion to
decision-memes.html. Same skeptic-disarmer role, text-form for execs and architects who want substance under the joke.
- Sourced from
docs/feature-inventory.md. Every numeric claim trails the inventory. When inventory drifts, this artifact follows in the same change.
- Cross-linked from
executive-briefing.html Slide 9 (Risks + mitigations). The “I heard X about Claude” objection has its own page.
© gmanch94 · CC-BY-4.0 · As of 2026-05. Verify pricing/models at anthropic.com.