25 Claude Fable 5 Prompts to Test Every Capability (2026)

Stripe ran a 50-million-line Ruby codebase migration in a single day using Claude Fable 5 — a job that would normally take a full engineering team more than two months. That's an impressive number, but it doesn't tell you what to actually type into the chat window.

Anthropic launched Claude Fable 5 on June 9, 2026, as the first publicly available model in its new Mythos class, sitting above Opus in capability. For two weeks — from June 9 through June 22 — it's included at no extra cost on Claude Pro, Max, Team, and seat-based Enterprise plans. If you're going to spend that free window on anything, spend it running real prompts against the model instead of reading another benchmark table.

This guide is the companion piece to Build Fast with AI's Claude Fable 5 review covering price, benchmarks, and API access, which breaks down the what and why. This article covers the how: 25 specific, copy-paste-ready prompts across coding, reasoning, research, content creation, data analysis, and agentic workflows — each designed to surface the exact capabilities Anthropic says Fable 5 leads on.

What Is Claude Fable 5? (Quick Recap Before You Start Prompting)

Claude Fable 5 is Anthropic's most capable publicly available model, launched June 9, 2026, as the first generally available model in the new Mythos class — a tier Anthropic positions above Opus. It runs on the model ID claude-fable-5, supports a 1-million-token context window with up to 128,000 output tokens, and is priced at $10 per million input tokens and $50 per million output tokens, exactly double Claude Opus 4.8's $5/$25 rate.

Fable 5 ships with safety classifiers that Anthropic says trigger in fewer than 5% of sessions on average. When a query touches offensive cybersecurity, controlled biological or chemical synthesis, or model distillation, it's automatically rerouted to Claude Opus 4.8 — and you aren't charged Fable 5 pricing for that response. None of the 25 prompts in this guide are designed to trip that classifier; they're all standard development, research, and content work, the categories Anthropic explicitly says run on full Fable 5 capability.

My honest read after going through Anthropic's own announcement and the AWS and TechCrunch coverage: Fable 5 isn't a new brain, it's the Mythos brain with a leash. For everyday coding, research, and writing — which is what this guide focuses on — the leash basically doesn't matter. If you want the full picture of every Claude model and product released in 2026, the Claude AI complete guide covers how Fable 5 fits alongside Opus 4.8 and Sonnet 4.6, and the Claude AI hub on Build Fast with AI is the place to bookmark for everything Anthropic ships next.

One practical note before you start: the free window runs out fast. June 9–22 is a trial period, not the long-term pricing reality — after June 23, Fable 5 usage on subscription plans draws from usage credits at the standard API rate. Run the highest-value prompts from this list first.

Coding Prompts to Test Claude Fable 5's Agentic Engineering

Coding is where Claude Fable 5's lead over previous models is largest. It scores 80.3% on SWE-Bench Pro versus 69.2% for Claude Opus 4.8 and 58.6% for GPT-5.5 — a 21.7-point gap that Anthropic says widens further on longer, messier, multi-file tasks. The five prompts below are built to surface that gap rather than test trivial single-function snippets.

Prompt 1 — Large-scale codebase migration (the Stripe-style test)

Here's a [language/framework] codebase with [N] files. Migrate it to [target language/framework], preserving all existing tests and behavior. Work through it file by file, run the test suite after each batch of changes, and give me a summary of every breaking change you had to handle.

Run this on a real internal repo, even a mid-size one of 50–200 files, and watch whether Fable 5 plans the migration order before touching code. That planning-then-verifying loop is what AWS described as Fable 5's “proactive self-verification” — the model develops its own checks and verifies its work as it goes.

Prompt 2 — Multi-file refactor with test-driven verification

Refactor the authentication module across these files to use [new pattern — e.g., dependency injection instead of singletons]. After each file change, write or update unit tests to cover the new structure, run them, and fix any failures before moving to the next file.

Prompt 3 — Vision-to-code: rebuild a UI from a screenshot

Here is a screenshot of a web app's dashboard page [attach image]. Rebuild this page as a working React component using Tailwind CSS, matching the layout, spacing, and color palette as closely as possible. List any elements you're estimating because they're not fully visible in the screenshot.

Anthropic specifically calls out that Fable 5 can rebuild a web app from screenshots alone with minimal scaffolding — this prompt tests that claim directly.

Prompt 4 — Autonomous flaky-test triage

This test suite has 3 tests that fail intermittently. Run the suite 5 times, identify which tests are flaky and why (race condition, shared state, timing, etc.), and propose a fix for each with your reasoning.

Prompt 5 — Token-efficiency comparison (medium vs. high effort)

Implement [a moderately complex feature, e.g., a rate limiter middleware with Redis backing] at medium effort. Then implement the same feature again at high effort. Compare the two implementations and tell me what changed and why.

This is the closest you can get in a single chat to reproducing the FrontierCode Diamond result Anthropic published: Fable 5 scored 29.3% at high effort and was still the most token-efficient frontier model even at medium effort — which matters because Fable 5's output tokens cost $50 per million.

To run any of these in Claude Code, start a session with claude --model claude-fable-5, or pick Fable 5 from the model picker in Claude Code for web or in GitHub Copilot (Pro+, Max, Business, Enterprise). If Claude Code is your daily driver, the Claude Code vs Codex comparison is worth a read before you commit Fable 5 credits to a long agentic run — it'll help you decide whether the task in front of you actually needs Mythos-class capability.

Reasoning and Knowledge-Work Prompts

For senior-level analytical reasoning — finance, legal, and root-cause work — Fable 5 holds the top score on Hebbia's Finance Benchmark for senior-level reasoning and leads the Legal Agent Benchmark at 13.3% versus GPT-5.5's 2.1%, on a benchmark Anthropic intentionally designed to be barely solvable.

Prompt 6 — Financial model audit

Here's a 3-statement financial model for [company/scenario — attach spreadsheet or paste the data]. Audit it for circular references, unrealistic growth assumptions, and inconsistencies between the income statement, balance sheet, and cash flow statement. Rank the issues by how much they'd distort the valuation.

Prompt 7 — Legal contract risk pass

Read this [NDA / vendor contract / lease — attach document] and identify every clause that's unusually favorable to the other party compared to market standard. For each one, explain the risk in plain English and suggest a redline.

This sits squarely in the Legal Agent Benchmark territory Anthropic tested — factual lookup, conceptual reasoning, and expected-value tradeoffs on real documents rather than toy examples.

Prompt 8 — Incident root-cause analysis from raw logs

Here are the error logs from the last 4 hours of an outage [paste logs]. Walk through a root-cause analysis: what happened first, what cascaded from it, and what's the most likely single point of failure. Then suggest 3 monitoring changes that would have caught this earlier.

Prompt 9 — Expected-value decision framework

I'm deciding between [option A] and [option B] for [business decision]. Build an expected-value comparison: list the possible outcomes for each option, estimate probabilities and impact, and tell me which option wins under the EV framework — and where your probability estimates are weakest.

IMC reportedly told Anthropic that Fable 5 “aced” their trading-analysis evaluations across factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis. In practice, prompts 6 and 8 are only as good as what you paste in — feed Fable 5 the messy, real logs or model, not a sanitized summary, or you won't see the reasoning gap Anthropic is describing. If these results impress you and you're curious how much further the unrestricted version goes, the Claude Mythos 5 review covers the Project Glasswing-only model that shares Fable 5's underlying architecture without the safety classifiers.

Research and Long-Document Prompts (1M Context + Vision)

Fable 5's 1-million-token context window and upgraded vision mean it can hold hundreds of pages of documents and read embedded charts, diagrams, and tables in a single pass. Anthropic says this opens up research and document-heavy work in finance, legal, analytics, architecture, and gaming.

Prompt 10 — Multi-document synthesis

I'm uploading [5–10 PDFs — research papers, reports, or filings on the same topic]. Read all of them and build a synthesis: where do they agree, where do they directly contradict each other, and what's the strongest claim each one makes that the others don't address?

Prompt 11 — Chart and figure extraction

Here's a scientific figure [attach image of a chart or graph]. Extract the exact numeric values for each data series, including axis units, and put them in a table. Flag any values you're estimating versus reading directly.

This is one of the vision strengths Anthropic specifically highlights for Fable 5 — extracting precise numbers from detailed scientific figures, the kind of task that used to require manual digitization.

Prompt 12 — Cross-document contradiction hunt

Compare these two versions of [a policy document / contract / spec — attach both]. List every substantive difference, not just wording changes, and explain what each change would mean in practice.

Prompt 13 — Competitor page rebuild and strategy read from a screenshot

Here's a screenshot of a competitor's pricing page [attach image]. Analyze the pricing structure, tier naming, and feature gating strategy, then suggest 3 ways our pricing page [paste your current copy] could be restructured to compete more directly.

These research prompts build on general prompt-structuring principles — clear role, explicit deliverable, explicit format. If you want the broader foundation before specializing for Fable 5, 150 Best Claude Prompts That Work in 2026 is a solid starting library.

Content Creation Prompts

Content creation isn't where Anthropic's headline benchmarks point, but Fable 5's larger context window and stronger instruction-following make it noticeably better at holding a consistent voice across long outputs and multi-format repurposing — the kind of work that breaks down when a model drifts halfway through.

Prompt 14 — Voice-matched long-form writing

Here are 3 examples of my previous blog posts [paste or attach]. Study the voice, sentence rhythm, and how I open and close sections. Then write a new 1,500-word post on [topic] in that same voice — don't tell me what the voice is, just write in it.

Prompt 15 — Multi-format repurposing in one pass

Take this [blog post / transcript — paste full text] and repurpose it into: a 6-tweet thread, a LinkedIn post under 200 words, and a 60-second video script with scene directions. Keep the core argument identical across all three but adapt the structure to each platform's norms.

Prompt 16 — Long-form video script with scene directions

Write a 10-minute YouTube script on [topic] structured as: hook (0:00–0:30), problem setup, 3 main points with on-screen examples, and a close with a call to action. For each section, include both the spoken script and a one-line visual direction for the editor.

Prompt 17 — Reverse-engineer a brand style guide

Here are 5 pieces of our existing content [attach docs/screenshots]. Reverse-engineer a one-page brand style guide: tone, vocabulary to use and avoid, sentence length patterns, and formatting conventions, so a new writer could match our style without seeing the original examples.

For more structured templates beyond these four, Claude AI Prompt Codes That Actually Work has reusable formats you can adapt for Fable 5's larger context window.

Data Analysis Prompts

Data analysis benefits directly from the same gains that pushed Fable 5 to the top of Hebbia's Finance Benchmark — stronger document-based reasoning and chart interpretation translate well into exploratory data work, SQL generation, and forecasting.

Prompt 18 — Full exploratory data analysis on a CSV

Here's a CSV with [N] rows of [describe data — e.g., 18 months of e-commerce orders]. Do a full exploratory analysis: summary statistics, missing data patterns, outliers, and at least 3 visualizations of the most interesting relationships you find. Explain what each chart shows before showing it.

Prompt 19 — SQL generation and optimization at scale

Here's my database schema [paste DDL or describe the tables]. Write a query that [describe the business question — e.g., finds customers whose order frequency dropped by more than 50% in the last quarter compared to the prior quarter]. Then explain how you'd optimize it if this table has 50 million rows.

Prompt 20 — Time-series anomaly detection with explanation

Here's daily [metric — e.g., signups, revenue, server latency] for the last 12 months [paste or attach]. Identify any anomalies, classify each as seasonal, one-off, or trend-break, and for the 2 most significant ones, hypothesize what business event might explain them.

Prompt 21 — Forecasting with methodology explanation

Using this historical data [attach], build a 6-month forecast for [metric]. Explain which forecasting approach you're using and why it fits this data better than the alternatives, and give me a confidence range, not just a point estimate.

These four prompts pair well with the 40 Python and SQL prompts in How to Use AI as a Data Analyst: 40 Python, SQL & ChatGPT Prompts, which work across models but get noticeably sharper output when run on Fable 5's larger context window, since you can paste in more of the schema and sample data at once.

Agentic Workflow Prompts for Claude Code, Cowork, and MCP

Fable 5's standout feature for agentic work is sustained autonomy. Anthropic says it can work for days at a time in an agent harness like Claude Code — planning its approach, checking progress against the goal, and refining its work as it goes.

Prompt 22 — Multi-day feature build with self-verification (Claude Code)

I want you to build [a moderately complex feature — e.g., a notification system with email and in-app channels, user preferences, and a digest mode]. Before writing any code, write out your plan and the test cases you'll use to verify each piece. Then build it incrementally, running tests after each piece, and keep a running log of what you've completed and what's left.

Run this with claude --model claude-fable-5 and let it run unattended for a while — this is the prompt pattern that exercises the long-horizon, self-checking behavior Anthropic is marketing as Fable 5's headline feature.

Prompt 23 — MCP-connected research-to-report pipeline

Using [connected tools — e.g., Google Drive and Slack via MCP], find our last 3 product roadmap docs and the last 2 weeks of #product-feedback messages. Cross-reference them and write a one-page memo: which roadmap items have the strongest user-feedback support, and which have none.

Prompt 24 — Cowork file-management automation

Go through the [folder/project] and organize these files into subfolders by [client/project/date]. Rename files that don't follow our [naming convention] and flag any duplicates you find before you move anything.

Prompt 25 — GitHub Copilot agent-mode PR review

Review this open PR end to end: check for logic errors, missing test coverage, and whether it matches the linked issue's requirements. Leave inline comments on anything that needs a fix before merge, and summarize the overall risk as low, medium, or high.

Set Copilot's agent mode model to Fable 5 in VS Code, Visual Studio, or JetBrains — Business and Enterprise admins need to enable the Fable 5 policy first, since it's off by default.

If you haven't set up Cowork yet, the Claude Cowork complete guide walks through getting started, and the Claude Managed Agents review covers running Fable 5 in a hosted agent harness instead of locally. For the MCP-connected prompt above, the MCP Workshop notebook in Build Fast with AI's gen-ai-experiments cookbook repository walks through wiring up connectors like Google Drive and Slack step by step.

Here's the honest caveat that applies to all five agentic prompts above: data retention. Fable 5 requires 30-day data retention to operate its safety classifiers, and zero data retention isn't available for this model. If your organization has a ZDR requirement — common in regulated industries — none of these agentic prompts are usable on Fable 5 yet, and Opus 4.8 remains your option for that work.

Frequently Asked Questions

What is Claude Fable 5 used for?

Claude Fable 5 is built for long-horizon autonomous coding, large codebase migrations, complex document analysis, senior-level financial and legal reasoning, vision tasks, and multi-day agentic workflows in harnesses like Claude Code and Cowork. Anthropic says its lead over previous models grows the longer and more complex the task gets.

How do I access Claude Fable 5?

Use the model string claude-fable-5 on the Claude API, select it in the Claude Code CLI with claude --model claude-fable-5 or via the model picker in Claude Code for web, or use it through Amazon Bedrock, Vertex AI, Microsoft Foundry, or GitHub Copilot (Pro+, Max, Business, and Enterprise plans).

Is Claude Fable 5 better than Claude Opus 4.8?

For long, complex, agentic tasks, yes by a clear margin: 80.3% versus 69.2% on SWE-Bench Pro and the top score on Hebbia's Finance Benchmark. For routine chat or cost-sensitive high-volume API use, Opus 4.8 at half the price ($5/$25 versus $10/$50 per million tokens) is often the better choice.

What prompts work best with Claude Fable 5?

Prompts that give Fable 5 room to plan, verify, and iterate — multi-step migrations, multi-document research, and agentic builds with explicit test-driven checkpoints — show the biggest gap over previous models. Short, single-turn prompts won't show much difference from Opus 4.8.

Can Claude Fable 5 build an entire app?

Yes. Within an agent harness like Claude Code, Fable 5 can plan, build, test, and refine a feature or app over an extended session with minimal intervention, which Anthropic describes as proactive self-verification — the model develops its own evaluations and checks its own work as it goes.

Is Claude Fable 5 free to use?

Only from June 9 through June 22, 2026, on Claude Pro, Max, Team, and seat-based Enterprise plans. After June 23, usage on subscription plans draws from usage credits at the standard $10/$50 per million token rate, and the claude.ai free tier doesn't include Fable 5 at all.

What happens if my prompt triggers Fable 5's safety classifier?

Anthropic says fewer than 5% of sessions trigger it. When one does, the query is automatically rerouted to Claude Opus 4.8 and you're notified that a different model answered — and you aren't charged Fable 5 pricing for that response.

Recommended Blogs

If you found this useful, these posts from Build Fast with AI go deeper on related topics:

Keep Building

Want to build production agentic workflows with models like Claude Fable 5? Join Build Fast with AI's Gen AI Launchpad — an 8-week structured program to go from 0 to 1 in Generative AI. Register here.

References

Anthropic — Claude Fable Product Page