On June 9, 2026, Anthropic released Claude Fable 5. It is a Mythos-class model, the same tier they originally built for government and defense partners. Mythos was considered too powerful to release publicly, so Anthropic added safety classifiers and guardrails, then shipped it under the name Fable.
Fable 5 sits above the Opus class in capability. On the coding benchmarks that actually matter, it beats every other publicly available model by double-digit margins. Stripe used it to complete a 50-million-line codebase migration in a single day that would have taken a team over two months.
The catch: it is expensive. Fable 5 costs $10 per million input tokens and $50 per million output tokens, roughly double what Opus 4.8 costs. It also burns through tokens fast because it thinks longer and harder than any previous model. For most everyday tasks, the older models are still the right choice. But when you need something done that no other model can handle, Fable 5 is in a class of its own.
How It Compares
- SWE-Bench Pro: Fable 5 80.3%, Opus 4.8 69.2%, GPT-5.5 58.6%, Gemini 3.1 Pro 54.2%
- FrontierCode Diamond: Fable 5 29.3%, Opus 4.8 13.4%
- Knowledge and Reasoning: Fable 5 1932, GPT-5.5 1769, Gemini 3.1 Pro 1314
- GPQA Diamond: Gemini 3.1 Pro 94.3%, GPT-5.5 92.8%, Fable 5 91.3%
- Price (input / output per million tokens): Fable 5 $10 / $50, GPT-5.5 $5 / $30, Opus 4.8 $5 / $25, Gemini 3.1 Pro $2 / $12
Fable 5 leads on coding and multi-step engineering work. Gemini 3.1 Pro edges it out on graduate-level science. GPT-5.5 is competitive on reasoning but falls behind on code. If you are building something complex, Fable 5 is the best model available right now.
What People Are Already Building
Within 24 hours of launch, X was flooded with demos. Some were real, some were people repackaging old videos to ride the hype. Here are the verified ones.
Minecraft clone in 20 minutes. A developer gave Fable 5 a single prompt: "Create a clone of Minecraft." Twenty minutes later, it generated a playable version with multiple biomes, a day/night cycle, ores, a cave system, and crafting. About 3,000 lines of code, $12 in API costs.
Boeing 747 from one prompt. Someone used Claude Code with Fable 5 to build a procedural Boeing 747 in Three.js from a single /goal prompt. The agent built the model, set up a headless screenshot camera, and iterated in a vision self-verification loop until it looked right. About 30 minutes.
Solar system physics simulation. Fable 5 built a solar system simulation that derived planetary orbital motion from physics first principles and used it to predict actual solar eclipses. Not a canned animation. Real physics.
Pokemon FireRed beaten with vision only. Anthropic showed Fable 5 playing through the entire game using nothing but screenshots. No game data, no maps, no helper tools. Previous Claude models could not do this even with helper frameworks.
Stripe's 50M-line migration. Stripe reported that Fable 5 completed a codebase-wide Ruby migration across 50 million lines of code in a single day. A human team estimated it would take over two months.
Full 3D websites from one prompt. Multiple developers showed Fable 5 generating complete 3D interactive websites from a single prompt description. Real visual hierarchy, working animations, professional design.
Heads up: some viral demos on X were fake. People repackaged old projects or used other tools and claimed it was Fable 5. Verify before you share.
How to Access Fable 5
Fable 5 is available now on all paid Claude plans.
- Open Claude. Go to claude.com or open the Claude desktop app. You need a paid plan (Pro, Max, Team, or Enterprise).
- Open the model selector at the top of the conversation.
- Select Fable 5. Your conversation will now use Fable 5 for responses.
For API access, use the model ID claude-fable-5 at $10 per million input tokens and $50 per million output tokens.
Tip: Fable 5 burns through tokens fast, especially at higher effort levels. For quick questions, casual conversations, or simple tasks, stick with Sonnet or Opus. Save Fable 5 for the hard stuff.
Effort Levels: Your Cost Control
Effort is the biggest dial you have for controlling Fable 5. It determines how long the model thinks before responding.
- Low (0 to 2,000 thinking tokens): summaries, formatting, simple Q and A, classification.
- Medium (2,000 to 8,000): multi-step analysis, code review, structured reports.
- High (8,000+, default): complex reasoning, novel problems, adversarial tasks.
- xHigh (extended): the hardest research and capability-sensitive work.
Fable 5 at medium effort often outperforms older models at their highest settings. Start at High and only go higher when the task demands it. Running everything at xHigh is the fastest way to drain your tokens.
Anthropic's Prompting Guide (Condensed)
Anthropic published a 14-section prompting guide specifically for Fable 5. The biggest takeaway: everything you learned about prompting other models might actually hurt you here.
1. Stop over-prompting. Long lists of rules and examples make Fable 5 worse, not better. Say what you want plainly and let it think.
2. Lead with the outcome. When Fable 5 finishes a task, its first sentence should answer "what happened" or "what did I find." Supporting detail comes after. Readability matters more than being concise.
3. Ground progress claims. Before reporting progress, have it audit each claim against a tool result from the session. Only report work it can point to evidence for. If something is not verified, it should say so explicitly.
4. State the boundaries. Fable 5 can occasionally take actions you did not ask for: drafting an email when you only asked it to analyze one, or creating backup branches you did not need. Tell it explicitly what it should and should not do.
5. Tell it to finish the work. Before ending its turn, have it check its last paragraph. If it is a plan, an analysis, or a promise about work it has not done, it should do that work now with tool calls. End the turn only when the task is complete.
6. Reassure it about context. Deep into long sessions, Fable 5 can worry about running out of context and suggest starting a new session. Tell it: "You have ample context remaining. Do not stop, summarize, or suggest a new session on account of context limits. Continue the work."
When to Use Fable 5 (and When Not To)
Use Fable 5 for: long-running engineering work, full app builds from one prompt, codebase migrations, complex 3D and game projects, multi-step research, anything where the cheaper models keep failing.
Use Sonnet or Opus for: quick questions, simple edits, formatting, translations, social posts, email drafts, data extraction, classification. Sonnet handles 80% of everyday tasks at a fraction of the cost.
Anthropic's own recommendation: the teams seeing the best results apply Fable 5 to their hardest unsolved problems. Testing it only on simpler workloads tends to undersell its capability range.
Copy-Paste Prompts
Build an entire app
Build me a [type of app] that does [core function]. Use [tech stack]. Make it production-ready with proper error handling, responsive design, and clean code. Do not stop until it works end to end.
Recreate a UI from a screenshot
[Paste screenshot] Recreate this UI exactly. Match the layout, spacing, colors, typography, and interactions. Use [React/HTML/etc]. Make it fully functional, not just a visual copy.
Build a game
Build me a [type of game] that runs in the browser. Use Three.js or plain Canvas. Include: player movement, collision detection, scoring, sound effects, and a restart mechanic. Make it actually fun to play, not just a tech demo.
Minecraft-style voxel world
Create a Minecraft-style voxel world with multiple biomes, a day/night cycle, procedural terrain generation, cave systems, and basic player movement with block placing and breaking. Make it run in the browser at 30+ fps.
Overnight autonomous task
# Task Description I need you to [describe the full task]. # Definition of Done [Describe exactly what 'done' looks like] # Available Context [Tools, files, constraints] Work autonomously. Verify your own progress against tool output before reporting it. Do not stop until the definition of done is met.
Long-form content
Write a [length] [type of content] about [topic]. This is for [audience]. The goal is [what you want readers to do/feel/understand]. Use a [tone] voice. Include specific examples, data points, and real stories where possible. Do not use generic filler. Every paragraph should earn its place.
The Safeguards
Fable 5 is a restricted version of Mythos. Anthropic added safety classifiers that block responses in certain high-risk areas. These trigger in less than 5% of sessions. The full unrestricted Mythos 5 model exists, but is only available through Anthropic's "Project Glasswing" program for infrastructure providers and vetted cybersecurity researchers. For everyone else, Fable 5 is the model you get.
Get the Full Guide
Download the PDF below for the full walkthrough, all copy-paste prompts, and the model selection cheat sheet.