How to combine Fable 5, Opus 4.8 and Sonnet 5 in Claude Code so the premium tokens only go where judgment matters. Planning and reasoning stay on the strong models, volume work goes to the cheap ones, and a packaged workflow does the routing for you. Up to 70% lower token spend.
Prices checked against Anthropic's pricing page, September 2026. The workflow file is on page 5.
Fable 5 is the strongest model available, and also the most expensive one. But in any large task, most tokens are burned on volume work: reading files, summarizing, mechanical edits, boilerplate tests. A mid tier model does that work just as well at a fraction of the price. The structure in this guide splits every big task into phases, routes each phase to the cheapest model that can do it well, and reserves Fable for the few moments where its judgment changes the outcome: planning, tricky debugging, conflict resolution. Quality is protected by an adversarial Opus review loop at the end. The blended bill drops sharply while the output stays at Fable level where it counts.
| MODEL | MODEL ID | INPUT $/1M | OUTPUT $/1M | VS FABLE |
|---|---|---|---|---|
| Fable 5 | claude-fable-5 | $10.00 | $50.00 | baseline |
| Opus 4.8 | claude-opus-4-8 | $5.00 | $25.00 | 50% less |
| Sonnet 5 | claude-sonnet-5 | $2.00 | $10.00 | 80% less |
| Haiku 4.5 | claude-haiku-4-5 | $1.00 | $5.00 | 90% less |
Anthropic API list prices per million tokens, checked September 2026. Sonnet 5 launched at $2 / $10 as an introductory price, and Anthropic has since made that its standard price.
Every token moved from Fable to Opus costs 50% less, to Sonnet 80% less, to Haiku 90% less, on both input and output. A hard task with several features run through the tiered pipeline keeps its planning and implementation on Fable, so it lands lower: typically 45 to 55% under an all Fable run (worked example on page 4). Routine work delegated entirely to Sonnet saves the full 80%. Across a week of mixed work the blended saving sits between those two, which is why the structure is summarized as up to 70% less.
The second lever is effort: implementers run at medium effort instead of high, which produces fewer thinking and output tokens on the same model without touching the plan quality.
When torn between two tiers, take the cheaper one for reversible work (a weak exploration report costs one retry) and the stronger one for work that is hard to undo (a bad merge costs an afternoon).
| TIER | USE IT FOR | NEVER USE IT FOR |
|---|---|---|
| Haiku 4.5 | Mechanical sweeps: file listing and renaming, log parsing, screenshot capture, format conversion, running scripted QA checks. | Anything that needs code judgment. |
| Sonnet 5 | Codebase exploration, ordinary implementation, test writing, routine QA, summarizing, first drafts of docs. | Architecture decisions, subtle debugging. |
| Opus 4.8 | Adversarial code review, verification passes, judging between options. | Work Sonnet already handles (Opus costs 2.5x Sonnet). |
| Fable 5 | Planning and task decomposition, design judgment, tricky debugging, merge conflict resolution, orchestration decisions. | Mechanical work of any kind. |
Effort controls how deeply a model thinks on a given call (low, medium, high, xhigh, max). It
works on the same model, so it stacks with tiering. In the workflow it is the
effort option on each agent call.
| EFFORT | WHEN | WHY |
|---|---|---|
| high | Planning, review, anything where intelligence matters. | Judgment work deserves full depth. |
| medium | Implementation against a written plan with concrete acceptance criteria. | The plan already carries the judgment; the implementer executes it. |
| low | Mechanical steps: opening a PR, running a checklist. | Fewer, more consolidated tool calls and terser output. |
For hard coding tasks with several parts, the tiering is packaged as a Claude Code workflow
(the file .claude/workflows/fable-five-dynamic.js, download on page 5). It splits
the task into features with explicit dependencies, implements them in parallel git worktrees,
and loops review failures back to the implementers until everything passes or the loop cap
is hit.
| PHASE | MODEL / EFFORT | WHY THIS TIER |
|---|---|---|
| Explore | Sonnet 5, parallel read only agents, one per angle | Reading and reporting is volume work; 80% cheaper per token. |
| Plan | Fable 5, effort high | The one place where judgment sets up everything downstream. |
| Implement | Fable 5, effort medium, one git worktree per feature | The plan carries the judgment; medium effort executes it with fewer tokens. |
| Merge | Fable 5, effort medium | Conflict resolution needs care but not deep planning. |
| Review | Opus 4.8, effort high, one adversarial reviewer per feature | A strong reviewer at half the Fable price; failures loop back, bounded by maxLoops (default 3). |
| PR | Fable 5, effort low | Mechanical. |
In a Claude Code session inside a git repository, ask for the workflow by name ("run the fable-five-dynamic workflow on: add a user settings page"), or invoke it directly:
Useful arguments: angles overrides the four default exploration angles;
maxLoops caps the review and fix loops (default 3); mode is
"merge" (integrate branches, default) or "separate" (leave one branch
per feature); openPR opens a GitHub PR when a remote is configured;
branch names the integration branch. Preflight commits a clean baseline
automatically before any worktree is created.
The same map applies to individual subagents and one off workflow scripts:
model: "haiku" or "sonnet") for work that belongs on a cheaper tier. An agent spawned without an override inherits the session model, which in a Fable session means paying Fable prices for work that did not need them.A four feature task, counting output tokens (the dominant cost; input tokens scale the same way at each tier):
| PHASE | TOKENS (APPROX.) | ALL FABLE, HIGH | TIERED |
|---|---|---|---|
| Explore (4 agents) | 200k | $10.00 | $2.00 (Sonnet) |
| Plan | 30k | $1.50 | $1.50 (Fable high) |
| Implement (4 features) | 350k high / 240k medium | $17.50 | $12.00 (Fable medium) |
| Review (4 reviewers) | 160k | $8.00 | $4.00 (Opus) |
| Total | $37.00 | $19.50 (47% less) |
The pipeline saves close to half on a hard task because planning and implementation deliberately stay on Fable. Routine standalone work routed entirely to Sonnet saves 80%, and Haiku sweeps save 90%. Across a week of mixed work the blended saving sits between those bounds, which is why the structure is summarized as up to 70% less.
| SITUATION | DO THIS |
|---|---|
| Coding task with several parts | Run the fable-five-dynamic workflow. |
| Search or read heavy question | Sonnet explorer agent, structured output. |
| File sweep, log parsing, capture | Haiku agent. |
| Ordinary coding or QA in isolation | Sonnet agent. |
| Verifying or judging finished work | Opus reviewer, effort high. |
| Plan, architecture, gnarly bug, merge | Fable, inline or effort high. |
| Tiny change (under about 3 files) | Do it inline; do not spawn anything. |
Open the link below (it opens as plain text) and save it as
fable-five-dynamic.js, or run this one line from the root of your repository:
Direct link: easyhive.pro/lm-tokens-step1/fable-five-dynamic.js
Claude Code loads workflows from .claude/workflows/ inside the repository you
are working in. The file name is the workflow name, so keep it as
fable-five-dynamic.js.
Start Claude Code in that repository and ask: "Run the fable-five-dynamic workflow
on: [WHAT TO BUILD]". Describe what done looks like in the same message. Watch
the phases with /workflows.
The run ends with an integration branch (fable-five/your-task), the merge and
test status and, when review passed and a remote exists, a pull request. Review it like
any PR from a colleague.
gh authenticated. The file contains no keys, no
accounts and no secrets; it only decides which model runs which phase and at what effort.
| MINUTES | DO THIS |
|---|---|
| 0 to 5 | Drop the file into .claude/workflows/ in one repository. |
| 5 to 15 | Pick a task that genuinely has three or more features. Write the request in one paragraph, including what done looks like. |
| 15 to 50 | Run the workflow. Do not intervene; let the review loop do its job. |
| 50 to 60 | Compare the cost of the run with your last all Fable session of a similar size. Write both numbers down; that is your real saving. |
All trademarks belong to their owners. Easy Hive is not affiliated with or endorsed by them. This site is not part of Facebook, Instagram or Meta Platforms, Inc., and is not endorsed by Meta. Savings and results shown are examples, not guarantees; yours will vary. Educational content only, not legal, tax or financial advice. Full disclaimers.