🏆 What Claude Fable 5 is
🧠 Imagine it this way: it’s like taking the most powerful car in Anthropic’s garage. It costs more at the pump, but it handles any tough stretch of the route without running out of steam.
O Claude Fable 5 is, today, the most capable model—and also the most expensive—in Anthropic’s public catalog. It was released in June 9, 2026, alongside Claude Mythos 5 (same capability, but without the safety classifiers public — restricted use through Project Glasswing, for defensive cybersecurity only).
🆕 New here? Safety classifier — an automatic filter that examines the request and response for signs of risk (e.g., a security exploit, biological harm). When it triggers, the model refuses part or all of the request. Mythos 5 runs without this public layer, so it’s used only in controlled settings.
The main capability shift is the context window (context window): 1 million tokens by default, and up to 128 thousand output tokens per request—without charging a price “premium” for it.
🆕 New here? Token / MTok — a token is the smallest unit of text the model processes (roughly a piece of a word). "MTok" = 1 million tokens, the unit used in the pricing table (e.g., "$10 per MTok of input"). We’ll use this unit in topic 4.
📊 By the numbers
In one sentence: Fable 5 is Anthropic’s most capable and most expensive model, with a 1M-token context window and up to 128k output.
⚡ Why it “feels different”
🧠 Imagine it this way: it’s like turning up the volume on a conversation that was already going well. You say less and are understood faster—the model already gets your intent without needing you to repeat the request in three different ways.
The first thing people notice when switching from another model is that Fable 5 follows short, direct instructions better than previous models, and supports a stronger reasoning on long tasks. This changes how you prompt—the next modules in this track teach exactly how to make the most of it (less enumeration, clearer intent).
✓ Officially highlighted capabilities
- • Long-horizon autonomy — preserves the instruction across runs lasting several days
- • Getting it right on the first try in complex systems—complete implementations in a single shot
- • Vision — reads dense technical images and screenshots more accurately with fewer tokens
- • Enterprise workflows — spreadsheets, slides, and documents
- • Code review/debugging — greater recall of bugs than Opus 4.8, outside the domains covered by the safety classifiers
✗ Typical limitation of earlier models
- • Needed constant supervision on long tasks
- • Several iterations to complete the “last step” in complex systems
- • Could read the image, but got fine details wrong or used more tokens
- • Spreadsheets/slides required manual rework afterward
- • Subtle bugs slipped through more often
📚 Official source
This list of capabilities comes from the official “Introducing Claude Fable 5” documentation, which compares it directly with Claude Opus 4.8 (the catalog’s previous most capable model).
💡 Practical tip
If you’re migrating from another model, resist the urge to write longer prompts “just in case.” Fable 5 responds better to short, clear instructions—the modules 1.2 and 1.5 in this track show exactly how that works in practice.
In one sentence: Fable 5 follows short instructions better and sustains more autonomy, vision, and code accuracy than previous models.
🕰️ The redeployment timeline
🧠 Imagine it this way: it’s like a new car recalled by the dealership the day after launch—not because it broke, but because the export paperwork wasn’t settled yet.
This topic is context, not the course’s central topic—the focus of this track is learning to prompt Fable 5 well. But it’s worth understanding what happened in the three weeks between launch and full availability today:
09/06/2026 — Launch
Claude Fable 5 and Claude Mythos 5 are publicly released. Mythos 5 is the variant without public safety classifiers, available for restricted use via Project Glasswing.
12/06/2026 — Export control and full suspension
The US government applies export controls both models, requiring access to be restricted by nationality. Since there was no way to verify nationality in real time, Anthropic suspended access for ALL users — not just foreigners.
06/26/2026 — Partial reactivation of Mythos 5
The government approves Mythos 5’s reactivation for Glasswing partners in the US.
06/30/2026 — Export controls lifted
The official announcement clears the way for both models to return.
01/07/2026 — Global redeploy (today)
Fable 5 and Mythos 5 are returning globally on Claude Platform, Claude.ai, Claude Code, and Claude Cowork. AWS/GCP/Microsoft Foundry are still being reactivated.
🆕 New here? Export control — a government rule that restricts access to a technology based on nationality. Without a way to check nationality in real time, the safest option for Anthropic was to suspend general access, not just access for those targeted by the rule.
The real reason behind the suspension is an important nuance: researchers at Amazon reported a method for bypassing Fable 5's safeguards, getting it to identify software vulnerabilities and, in one case, generate code demonstrating how to exploit one.
✗ Common misunderstanding
"Only Fable 5 could produce that exploit—which is why it was suspended; it's a risk unique to that model."
✓ What Anthropic’s own test showed
Much less capable models — Opus 4.8, GPT-5.5, Kimi K2.7, and even Haiku 4.5, Sonnet 4.6, Opus 4.6/4.7 — could reproduce the same exploit demonstration. The suspension was a regulatory precaution, not because Fable 5 was uniquely dangerous in this area.
In one sentence: The redeploy took 22 days because of export control, not because Fable 5 was uniquely risky—other models reproduced the same exploit.
💰 Exact pricing and comparison
🧠 Imagine it this way: Think of MTok as “1 million token-words” on your monthly bill—and Fable 5 charges twice as much as Opus 4.8 for every million that passes through it, both for reading and writing.
Fable 5 is, without a doubt, the most expensive model in the catalog. The table below is the official price quote per MTok (input = what you send, output = what the model generates):
Claude Fable 5 / Mythos 5 input $10.00 / MTok output $50.00 / MTok Claude Opus 4.8 input $5.00 / MTok output $25.00 / MTok Claude Sonnet 5 (até 31/ago) input $1.00 / MTok output $5.00 / MTok Claude Haiku 4.5 input $0.50 / MTok output $2.50 / MTok
Haiku 4.5
Sonnet 5
Opus 4.8
Fable 5
📚 Details worth knowing
- Batch API (asynchronous processing, 50% discount): Fable 5 drops to $5/$25 per MTok.
- Without a long-context premium: The price per MTok stays the same even when using the full 1M-token context window.
💡 Practical tip
For simple, high-volume tasks, consider Haiku 4.5 or Sonnet 5. Save Fable 5 for work that truly needs the extra capability—the higher price often pays off when it prevents rework and iteration.
In one sentence: Fable 5 costs $10/$50 per MTok—exactly twice as much as Opus 4.8, the most expensive model in the catalog today.
🎁 The promotional window
🧠 Imagine it this way: is a generous sample of the store’s most expensive product, for a limited time, before you have to decide whether to pay full price to keep using it.
In today’s redeploy, Anthropic opened a promotional window—but it applies only to Claude subscribers, not API users. It’s worth understanding exactly where the line is drawn:
✓ This applies to
Plans Pro / Max / Team (and part of Enterprise): up to 50% of the weekly usage limit in Fable 5, at no extra cost, up to July 7, 2026. After this date, additional usage starts consuming usage credits.
✗ This does NOT affect
The price of API ($10/$50 per MTok) has never changed. Anyone using Fable 5 through the Claude API or Claude Platform has always paid—and still pays—the full price. There is no "API promotion" here.
🆕 New here? Usage credits — credits you buy separately to keep using Claude after your plan’s weekly usage limit runs out. This is different from the API price per MTok.
📅 The date that matters
If you subscribe to Pro/Max/Team, use this week to really test Fable 5—the 50% discount on the weekly limit ends on July 7, 2026. After that, extra usage comes out of usage credits as usual (see support.claude.com about managing usage credits).
In one sentence: through July 7, Pro/Max/Team subscribers get a 50% discount on the Fable 5 weekly limit — the API has always charged (and still charges) the full price.
🗣️ What early adopters are saying
🧠 Imagine it this way: before the spec sheet, word of mouth from people who have actually tested it gives you the first real read on a new model.
In the first few days after the redeploy, some well-known early adopters in the development community shared their first impressions of Fable 5 for coding—it’s worth reading them with the right filter, explained at the end of this topic.
📚 Dan Shipper (Every)
No benchmark the internal “Senior Engineer” at Every, Fable 5 scored 91 out of 100, compared with Opus 4.8's 63 (the previous record). Shipper called the model "warp drive for coding": it handles large tasks over hours, but isn’t the ideal choice for quick, low-cost tasks (source: every.to, "Vibe Check: Fable 5 Is the Best Coding Model in the World").
EN (oficial): "the best model I have used for coding, by a wide margin" PT-BR (tradução): "o melhor modelo que já usei para programar, por uma margem ampla"
⚠️ Editorial caution note
The two citations above came from posts on X and an external blog— no from Anthropic’s official documentation. Treat them as enthusiastic, informative early adopter accounts, but not as officially confirmed facts from Anthropic.
In one sentence: The first reports point to major coding gains—but these are early adopter impressions, not an official claim from Anthropic.
🧾 Module Summary
Next module:
1.2 — Give the why + don’t ask for reasoning.