AI Tools · Model Review

Claude Fable 5 Review: Features, Pricing, Pros & Cons

Anthropic’s first Mythos-class model, tested against real coding and writing work — the benchmarks, the $10/$50 pricing, the June export-control shutdown, and whether it’s actually worth using over Opus 4.8.

I’ve now run Claude Fable 5 through a 40,000-line refactor, a stack of client contracts, and about a dozen “can you just fix this bug” sessions that turned into multi-hour debugging marathons. It’s the most capable model I’ve personally used from any provider, and it’s also the first Claude model I’ve had to actively talk myself out of using, because the token bill adds up fast if you point it at trivial work. This review covers what it actually does well, where it falls short, what it costs in practice (not just on paper), and the mess around the government-ordered shutdown that most reviews either skip entirely or get wrong.

Fable 5 vs Opus 4.8 vs GPT-5.5, at a glance

If you’re comparing models for a coding or agentic workload, here’s the matrix I built before committing budget to Fable 5. Numbers below are from Anthropic’s published launch benchmarks and my own day-to-day usage tracking.

What Claude Fable 5 actually is

Until June 9, 2026, Anthropic had a tier of models it never released publicly. Internally it’s called Mythos-class — the class that includes capabilities in cybersecurity and biology research strong enough that Anthropic kept it restricted to a small group of vetted partners under something called Project Glasswing. Fable 5 is what happens when that tier gets a public release: the same underlying weights as Claude Mythos 5, wrapped in a layer of safety classifiers that watch every request for high-risk cybersecurity, biology, chemistry, or model-distillation queries and reroute those to Opus 4.8 instead. Anthropic says that reroute triggers in under 5% of sessions, and you’re not charged Fable pricing when it happens.

Mythos 5, the uncapped sibling, is still restricted to Glasswing partners. If your integration or account doesn’t have that access, Fable 5 is the ceiling — and honestly, for anything short of frontier cybersecurity or bio research, it’s plenty.

The model ID on the API is claude-fable-5. It ships with a 1M-token context window by default and up to 128K output tokens per request, at the same flat per-token rate whether you’re sending 9K tokens or 900K — no long-context surcharge, which is a real change from how a lot of competitors price things.

Claude Fable 5 benchmarks

Anthropic’s headline number is SWE-Bench Pro, a harder successor to the standard SWE-Bench coding benchmark. Fable 5 posted 80.3%, more than 20 points ahead of GPT-5.5’s 58.6% and well clear of Opus 4.8. Anthropic also cites it as the top scorer on Cognition’s FrontierBench coding eval, and reviewers who got early access reported it leading on nearly every published benchmark, not just coding — vision, finance and legal reasoning, and long-horizon agentic tasks all show a similar gap.

What the raw scores don’t show is token efficiency. In my own testing and in reports from other early users, Fable 5’s lead widens the longer and more complex the task gets — on a genuinely hard multi-day migration, one team reportedly finished in a day what would’ve taken over two months by hand. On short, simple tasks, though, the advantage mostly disappears, and you’re just paying more for the same output a cheaper model would give you.

One important caveat I’d flag to anyone quoting these numbers in a procurement decision: benchmark gains don’t always translate cleanly to your specific domain. I’d treat SWE-Bench Pro as a strong signal for coding work specifically, and run your own eval on a real task from your backlog before committing budget.

Claude Fable 5 free — is there a free tier?

Short answer: not really, not anymore. There’s no Fable 5 access on Claude’s free plan at all. When it first launched, Anthropic gave Pro, Max, Team, and seat-based Enterprise subscribers free access to Fable 5 within their existing plan from June 9 through June 22, 2026 — no extra usage credits required. That window closed, and then the export-control shutdown hit on June 12 anyway, so most people never got the full two weeks.

After access was restored on July 1, 2026, Anthropic reopened a short trial window again: paid subscribers could use Fable 5 through July 7 for up to 50% of their normal weekly usage limits, according to Claude’s help center. Past that, Fable 5 requires usage credits on top of a subscription until Anthropic has enough capacity to fold it back in as a standard included feature — and they’ve been clear they haven’t committed to a date for that.

My advice: treat any “free” window on Fable 5 as a trial period for testing your workload, not as the long-term pricing reality. If you want to evaluate it seriously, do it while a free window is open and log your actual usage so you’re not guessing at cost later.

Claude Fable 5 price

On the API, Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens — double Opus 4.8’s $5/$25 and roughly double GPT-5.5’s sticker input price. Prompt caching still applies and can cut input costs by up to 90% on repeated context, which matters a lot if you’re running long agentic sessions that reuse the same system prompt or codebase context across many calls. If your workload needs to run entirely within the US, US-only inference is available at a 1.1x multiplier on top of standard pricing.

The framing I keep coming back to: comparing Fable 5 to Sonnet or Opus on a per-token basis is misleading, because different models burn different numbers of tokens to finish the same task. On a quick code review, Fable 5 can cost 15x more than a smaller model and it’s hard to justify. On a task that would otherwise eat two hours of your own cleanup time, Fable 5 finishing it correctly on the first pass is often cheaper than your own hourly rate. Price it against the task, not the ticker.

Anthropic Fable 5 ban — the export-control shutdown, explained

This is the part most “review” posts either skip or get wrong, so here’s the actual timeline. Three days after launch, at 5:21pm ET on June 12, 2026, Anthropic received a letter from the U.S. Department of Commerce — reportedly from Commerce Secretary Howard Lutnick — issuing an export-control directive ordering the company to suspend all access to Fable 5 and Mythos 5 for any foreign national, anywhere in the world, including Anthropic’s own foreign-national employees. Because there was no practical way to filter access by nationality in real time across every cloud platform, Anthropic disabled both models globally, for every customer, effectively overnight.

Anthropic’s public statement said the government’s letter didn’t include specific details of the national-security concern, but that its understanding was the order followed a demonstrated method of bypassing Fable 5’s safeguards — a “jailbreak.” Anthropic reviewed the demonstration and said it surfaced a small number of previously known, minor vulnerabilities that other publicly available models could also find, and pushed back publicly on the idea that a narrow jailbreak justified pulling a model already in the hands of hundreds of millions of users. Reporting from Axios and Forbes later added color: Amazon researchers had reportedly flagged the cybersecurity risk to officials, and there was separate concern that a China-linked group may have accessed Mythos, raising fears the model could be reverse-engineered.

The suspension lasted nearly three weeks. On July 1, 2026, the Department of Commerce lifted the export controls, and Anthropic began restoring access the next day across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork, with AWS, Google Cloud, and Microsoft Foundry following shortly after. Anthropic also confirmed it had trained an improved safety classifier in the interim specifically targeting the reported jailbreak technique — with the honest caveat that the tighter filter may now flag more benign requests during routine coding and debugging, something the company says it’s continuing to tune.

If you’re building anything in production on top of Fable 5, the practical lesson isn’t about the politics — it’s that a frontier model can disappear from your stack with zero notice, and organizations that had silently auto-adopted “whatever the latest Anthropic model is” learned that the hard way. Build a fallback to Opus 4.8 into your integration now, even if you never expect to need it.

Claude Fable 5 system prompt

A quick note before this section: you’ll find sites claiming to publish Fable 5’s full “leaked” system prompt. I’m not going to reproduce or vouch for that content here — it’s unverified, Anthropic hasn’t confirmed it, and treating a scraped file as ground truth is a bad idea for anyone making real prompting decisions. What I’ll cover instead is Anthropic’s actual published prompting guidance for Fable 5, which is more useful anyway.

A few things Anthropic explicitly documents as different about prompting Fable 5 compared to Opus or Sonnet:

• Start harder than you’d normally dare. Anthropic’s own guidance is to pick a task more ambitious than what you’d assign a prior model, and let Fable 5 scope it, ask clarifying questions, and execute — undertesting it on trivial tasks tends to undersell what it can actually do.
• Don’t ask it to narrate its reasoning in the response. Prompts or skills that tell the model to echo or transcribe its internal thinking as response text can trigger a “reasoning_extraction” refusal category, which causes elevated fallbacks to Opus 4.8. If you need visibility into reasoning, use the structured thinking blocks instead.
• For long asynchronous agents, build a send-to-user tool. Fable 5 rarely calls this kind of tool on its own unless the system prompt explicitly instructs it to use it for verbatim, user-facing content — not for routine progress narration.
• Vision and dense technical images are a genuine strength. Fable 5 is trained to use bash and crop tools to work through blurry, flipped, or noisy screenshots, and interprets dense UI and technical diagrams more accurately than prior models, often using fewer output tokens to do it.

If you’re migrating existing skills or CLAUDE.md files from Opus to Fable 5, it’s worth auditing them specifically for “explain your thinking” or “show your work” instructions — that’s the single most common thing I’ve seen trip the reasoning-extraction fallback in practice.

How to use Claude Fable 5

Fable 5 is available across every major Claude surface. Here’s how to get to it on each one:

• Claude.ai (web, desktop, mobile): open the model picker in a new or existing chat and select “Fable 5.” It counts as 2x usage against your plan’s limits.
• Claude Code (CLI): pass the model flag when starting a session, e.g. claude --model claude-fable-5, or set it as your default model in your Claude Code config so you don’t have to specify it every time. Make sure you’re on a recent CLI version (2.1.170 or later) since earlier builds don’t support it.
• Claude Code for web: open the model dropdown in the session settings and select Claude Fable 5 from the list.
• Claude API: set model to claude-fable-5 in your Messages API request. There’s no default system prompt on API calls — that’s specific to the Claude.ai consumer chat interface.
• Claude Cowork, AWS Bedrock, Google Cloud, Microsoft Foundry: all support Fable 5 post-restoration, though availability on some cloud platforms rolled back in slightly after the July 1 restoration date.

If a request gets flagged by the safety classifier and rerouted to Opus 4.8, the API returns it as a normal successful response with a stop_reason of “refusal,” not an error — and you’re not billed Fable pricing for that specific call. If this happens a lot during routine coding, that’s expected behavior right now while Anthropic tunes the post-shutdown classifier; use the feedback option in Claude Code or the thumbs-down button in Claude.ai to flag false positives.

Pros and cons

My testing notes

I ran Fable 5 against three real workloads instead of synthetic prompts, because that’s the only way I trust a review — including my own.

A 40,000-line codebase migration. This is where Fable 5 earned its price. It planned the migration in stages, flagged three edge cases I hadn’t thought about, and got through it with one round of human review instead of the usual three or four. Opus 4.8 could have done this too, but it would have needed more hand-holding along the way.

Routine bug fixes. Here it was overkill. It’s thorough to a fault — it wants to investigate every corner of a problem before shipping a fix, which is great for a hard bug and genuinely annoying for a one-line typo fix. On a handful of small PRs, I found myself waiting longer for a fix Opus 4.8 would’ve shipped just as correctly and faster.

The classifier fallback. Twice during normal debugging (not anything security-adjacent), a request got silently rerouted to Opus 4.8 with a refusal stop reason. Both times the actual task was benign — one was a request to review error-handling code that happened to touch input sanitization. Not billed, not a huge deal, but worth knowing this can happen mid-session on ordinary work, especially right after the July restoration while the classifier is still being tuned.

Signup and access friction. If you’re on an older Claude Code CLI build, Fable 5 simply won’t show up in the model list with no clear error explaining why — I lost twenty minutes to this before realizing I needed to update past version 2.1.170. Worth checking your version first if it seems to be missing.

Who should actually use Fable 5

FAQ

The bottom line

Claude Fable 5 is the most capable model I’ve personally tested, and it’s also the easiest one to overpay for if you’re not deliberate about when you reach for it. Use it for your genuinely hard, ambiguous, long-running work — the stuff that would otherwise eat days of your own time — and route everything else to Opus 4.8 or Sonnet. The government-ordered shutdown is now resolved, but it’s a real reminder to keep a fallback model wired into anything you build in production, because a frontier model disappearing overnight isn’t hypothetical anymore.

Related reading

• Wondering whether Fable 5 or its predecessor is trustworthy for sensitive work? See Is Claude AI Safe and Trustworthy?
• For a deeper look at day-to-day coding performance before Fable 5 existed, read Is Claude AI Good for Coding?
• Comparing across providers? Is ChatGPT Good for Coding? covers the other side of the SWE-Bench conversation
• If Gemini is also on your shortlist, see ChatGPT vs Google Gemini
• Looking beyond Anthropic entirely? Check ChatGPT Alternatives
• Already running Claude Code day to day? Claude Code Use Cases is a good next read
• For the numbers behind the competition, see ChatGPT Statistics 2026 and Key Facts
• If you’re pricing out coding assistants specifically, Verdent Alternatives covers a few worth comparing