Source: raw/Stop_Using_Fable_5_in_Claude_Code_It_s_Holding_You_Back.md
Creator: Simon Scrapes (YouTube) · URL: https://www.youtube.com/watch?v=d9XCX0PcOq0 · Duration: 20:16 · Uploaded: 2026-07-12 · Platform: YouTube
Recorded the week Fable 5 was expected to leave Claude subscriptions, this video argues that most work doesn’t need Fable and shows three places it pays for itself (planning, review, goal-driven reasoning work) plus eight settings that cut the baseline context every message carries. The access situation it describes changed a week later; see the note below. The settings advice is model-independent. It closes with a pitch for the creator’s Agentic Academy and one-line-install Agentic OS.
Dated access claim
Simon says Fable 5 “is leaving your Claude subscription on July the 12th” and every token will bill at API rates (50 out). That was the deadline at recording. Anthropic extended it to July 19, then on 2026-07-20 made Fable included in Max and Team Premium at 50% of limits, with Pro and Team Standard on usage credits plus a one-time $100 credit (Claude Fable 5, status callouts). The per-token prices are unchanged.
Key Takeaways
- “Hi” costs tens of thousands of tokens. Before you type anything, Claude Code loads the system prompt,
CLAUDE.mdand memory files, tool and MCP schemas, custom agents and skills. On his repo, “Hi” used 37,200 tokens (4% of the window) per/context; a GitHub report he shows had 31,000 for “Hi” and 19,000 for “What color is the sky?”. On Fable he puts the first message at 30 to 40 cents and each cached follow-up at 3 to 4 cents./usageshowed Claude in Chrome taking 45% of his MCP token use. - Judge cost per finished task, not per token. He relays one team’s three workloads: a 200-line PR review (Fable about 10 cents, nearly eight times Sonnet, but more edge cases found); a Next.js pricing page (Sonnet 25 cents plus two hours of human rework, Fable 5.40 “with rough edges”, Fable $23.70 with 450,000 output tokens, tests and docs, merge-ready; Sonnet lost coherence).
- Pattern 1: plan on Fable, build on Sonnet.
/model opusplanplans on Opus and implements on Sonnet; there was no Fable version, so do it by hand: switch to Fable, plan in plan mode, have it writeplan.mdand stop, switch to Sonnet, “execute the plan in<dir>/plan.md”. His planning run used about 38,000 tokens over a couple of minutes and asked about edge cases. - Pattern 2: Fable as reviewer. A small subagent in
.claude/agents/code-reviewer.mdpinned to Fable: reviews changes for bugs, edge cases and security issues after implementation; “do not write code, just return the findings” to keep expensive output short. - Pattern 3: goal plus success criteria, no recipe. For reasoning-heavy work with many small judgment calls (his example: reconcile every Q2 receipt in
/receiptsagainst a bank export; every transaction matched or listed inunmatched.mdwith a best guess why; totals agree; uncertain items flagged; verify by summing both sides and confirming each bank line appears once). A Sonnet prompt for the same job needs explicit matching rules (amount first, date within four days, 15% tolerance only if the merchant matches, “Uber *trip”), and still misses cases. - Don’t chat with Fable. A 30-minute back-and-forth costs 3 by his estimate against about 15 cents on Sonnet.
The eight settings
- Default to Sonnet.
/model→ Sonnet as the default for new sessions. He relays that teams making this change cut bills by about half with no quality loss on everyday work. - Pin subagents to a cheap model. Subagents inherit the parent model, so a Fable session can spin up five Fable researchers. In
~/.claude/settings.jsonunderenv, set the subagent model variable (CLAUDE_CODE_SUBAGENT_MODEL) to Haiku. Do the planning on the strong model first so the workers only execute. - Trim
CLAUDE.mdto what Claude can’t infer. He relays a test: a 3,847-tokenCLAUDE.mdcut to 312 tokens gave about 92% less context with no quality regression. HTML comments cost nothing; an@AGENTS.mdimport loads that whole file every session. Aim for under 200 lines, ideally about 60. .claudeignoreat the repo root fornode_modules,dist,build,coverage, lock files andvendor. It is advisory. He relays an 85.5% context reduction from this file alone (attributed in the captions to “Fireclaw’s blog”).- Deny reads in permissions, which is enforced: add deny rules for
node_modules,distand lock files insettings.jsonand commit it so the team shares them. - Disconnect idle MCP servers (he removed Notion and Canva), or use a CLI for occasional calls. He estimates 10,000 to 20,000 tokens per server and 50,000 to 70,000 for four or five.
/mcpshows unauthenticated servers to disable. - Tool search: set
ENABLE_TOOL_SEARCH=trueinsettings.jsonenvso tool definitions load on demand. (His August video says tool search is now on by default and retracts item 6; see Power-User Tips.) - Auto memory off if you run your own memory system. Auto memory forks the conversation into a background call to extract memories at full input price. Toggle it in
/memory. And set a monthly spend limit on claude.ai → Settings → Usage if usage credits are on.
His before-and-after “Hi” test came out higher after the changes, which he attributes to something left in memory and to having switched to Sonnet; he did not show a saving on camera.
Try It
- Type “hi” in a fresh session in your main repo and run
/context. Note the total and the biggest line. - Set the subagent model to Haiku in
settings.json, add.claudeignoreplus deny rules, and rerun. - For your next Fable job, plan on Fable, write
plan.md, then build on Sonnet. - Add a Fable-pinned review subagent that returns findings only.
- Set a monthly spend cap before turning on usage credits.
Implementation
Tool/Service: Claude Code settings and commands: /context, /usage, /model, /mcp, /memory.
Setup: ~/.claude/settings.json (env block, permissions.deny), .claudeignore at the repo root, .claude/agents/code-reviewer.md with a model: field.
Cost: Fable 5 API pricing 50 per million tokens in / out (his figures, matching the model page).
Integration notes: The three workload costs and the 92% and 85.5% reductions are third-party figures he relays; none were reproduced in the video.
Open Questions
- The team behind the three-workload comparison is not named in the captions.
- “Fireclaw’s blog” may be a caption error; the source was not identified.
- Whether
.claudeignoreis a supported Claude Code file or a community convention was not checked against the docs; he himself calls it advisory.
Related
- Mythos 5 — pricing, access history, field reports.
- Claude Code Token Optimization — overlapping tactics from other creators.
- Which Claude Model for Which Job — routing guide.
- Advisor Strategy — the first-party way to mix a cheap executor with a strong model.
- Subagents — model inheritance and per-agent model fields.
- Claude Usage Limits Playbook — plan limits and usage credits.