Claude Fable 5 Review: Anthropic's Most Powerful Public Model Ever (First Look)

Anthropic shipped Claude Fable 5 on June 9, 2026, and the pitch is genuinely new: this is the first publicly available Mythos-class model — the internal frontier tier that, until yesterday, only existed behind Anthropic's own walls and a government-adjacent preview program. This Claude Fable 5 review is a first look, not a verdict carved in stone: I've had roughly a day with it across the Claude app, Claude Code, and the API. But a day was enough to see the gap.

Short version: the benchmark bragging — state-of-the-art on nearly everything tested, 80.3% on SWE-Bench Pro against Opus 4.8's 69.2% — matched what I saw on real tasks more closely than launch-day numbers usually do. And until June 22, it's included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost, which makes the next two weeks the cheapest this model will ever be.

How I Tested This

Why Claude Fable 5 is the biggest AI story of the week

Three things make this launch different from the usual model-version bump.

First, the Mythos backstory. Anthropic has been openly telling regulators and the public that its frontier internal models were getting too capable to release as-is — TechCrunch's launch coverage pointedly noted Fable 5 arrived days after the company warned AI is getting dangerous. Fable 5 is Anthropic's answer to its own warning: the same Mythos-class capability, wrapped in safeguards judged strong enough for general availability. A sibling model, Claude Mythos 5 — same underlying model, some safeguards lifted — went only to a small group of cyberdefenders and infrastructure providers through Project Glasswing, in collaboration with the US government. That two-track release is unprecedented for a mainstream AI product.

Second, the benchmarks aren't the usual single-digit creep. On SWE-Bench Pro, Fable 5 posts 80.3% against Opus 4.8's 69.2% and GPT-5.5's 58.6%. On Cognition's brutal FrontierCode benchmark it scores 29.3% — more than double Opus 4.8's 13.4% and five times GPT-5.5's 5.7%. Anthropic claims state-of-the-art on nearly every benchmark it tested, spanning software engineering, knowledge work, vision, and scientific research.

Third, the free window. From June 9 through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23 it comes off those plans and requires usage credits. Anthropic is effectively running a two-week public trial of its flagship — and pulling the ladder up after.

What Claude Fable 5 actually is

Strip the mythology and Fable 5 is a new top tier in the Claude lineup — a step above Opus, not a replacement for it.

• Model: the first public Mythos-class model, positioned above Opus 4.8 in capability and price
• Access: Claude apps (web, desktop, mobile), the API as claude-fable-5, and Amazon Bedrock for AWS shops
• Specs: 1M-token context window, 128K max output — same envelope as Opus 4.8
• Safeguards: in specific high-risk areas (notably cybersecurity and biology), Fable 5 declines and your query is answered by Claude Opus 4.8 instead — Anthropic says this triggers in under 5% of sessions
• Strengths per Anthropic: state-of-the-art software engineering, knowledge work, vision, and scientific research

For developers, one practical note from my API testing: Fable 5 keeps the modern Claude API surface — adaptive thinking only, no temperature/top_p/top_k — and adds one quirk of its own: explicitly sending thinking: {type: "disabled"} returns an error where Opus 4.8 accepts it. You omit the parameter instead. Small thing, but it'll catch people migrating scripts on day one.

My honest testing experience

The moment that earned the rating came forty minutes in. Last month, refactoring this blog's table-of-contents logic, Opus 4.8 needed three rounds: first pass broke an edge case, second pass fixed it but duplicated logic, third pass cleaned up. I kept the prompt. I gave Fable 5 the identical task on the identical files. It produced the final-round-quality answer on the first try — including catching the edge case I hadn't mentioned, with a one-line note explaining why it mattered.

The pattern held across the day: fewer rounds, less hand-holding, noticeably better at catching things I didn't say out loud. The 60-page contract analysis was the other standout — it didn't just summarize, it built the comparison I actually needed and cited clause numbers that checked out. Vision was quietly excellent: it read a cramped analytics dashboard screenshot dead accurately, including the small print Opus tends to fumble.

The reroute is real but rare. I went looking for it — security-adjacent research questions are normal work for this blog — and triggered it once: a brief notice, and Opus 4.8's answer in place of Fable's. For my use, a non-event. If your daily work is security research or the life sciences, that sub-5% will not be evenly distributed across your sessions, and you should test against your actual workload during the free window.

What I liked

• The coding gap is visible in real work, not just benchmarks — one-shot results where Opus 4.8 iterates
• It reads intent — repeatedly handled the thing I forgot to specify, which is the most "next-tier" behavior of all
• Document analysis with citations that survive spot-checking — clause-level accuracy on a 60-page stack
• Vision is the best I've used — dense dashboards and fine print extracted without error
• The free window is genuinely generous — two weeks of the frontier model on a $20 Pro plan
• 1M context at launch — no long-context premium games

What frustrated me

• Usage limits bite faster than Opus — "included free" on Pro is not "use freely"; heavy sessions hit the ceiling
• The June 23 cliff — it leaves the plans entirely and moves to usage credits, and credit pricing for app use wasn't clearly published at launch
• API price is double Opus 4.8 — $10/$50 vs $5/$25 per million tokens; for many workloads Opus remains the rational default
• The reroute can be invisible until you notice the notice — if you missed it, you'd attribute Opus 4.8's answer to Fable 5
• Slightly slower than Opus on long generations in my testing — the extra thinking shows up as latency

Pricing — is Claude Fable 5 worth it?

The honest math: through June 22, the answer is trivially yes — a $20 Pro plan buys you the most capable public model on Earth, and you should be stress-testing it on your hardest work right now. After June 23 it becomes a real decision. At double Opus pricing, Fable 5 makes sense for the work where the capability gap pays for itself — hard engineering, high-stakes analysis, anything where a failed Opus attempt costs more than the token premium. For routine drafting and summarization, Opus 4.8 at half price remains the sane default. Worth noting: $10/$50 is less than half what the Mythos Preview cost the few who had it, so by Anthropic's own ladder this is the discount tier.

Who should use Claude Fable 5

Use it (at minimum, this free fortnight) if you are:

• A developer with a backlog of hard problems cheaper models keep fumbling — this is the model for the tickets you've been avoiding
• A knowledge worker doing contract, research, or document-heavy analysis where citation accuracy matters
• Anyone already paying for Pro/Max/Team — testing it costs you nothing until June 22
• A team deciding its post-June model budget — you need this data point now, while it's free to gather

Who should avoid Claude Fable 5

Skip it (or stay on Opus) if you are:

• Running high-volume API workloads where 2x token cost compounds — Opus 4.8 handles most of it
• Working primarily in security research or biology — the safeguard reroute concentrates exactly on your topics, and you may effectively be paying Fable prices for Opus answers
• A casual user whose tasks never stressed Opus in the first place — you won't feel the difference, but you'll feel the limits
• Latency-sensitive — the extra capability shows up partly as extra thinking time

How Fable 5 compares to the alternatives

The benchmark deltas are unusually lopsided for 2026: 80.3% vs 58.6% on SWE-Bench Pro, and a five-fold gap on FrontierCode, put GPT-5.5 a full tier behind on engineering work. OpenAI will answer — they always do — but today, the coding crown isn't close. For the wider model landscape, my ChatGPT vs Claude vs Gemini comparison covers how the three ecosystems differ beyond raw capability.

What I'd change about Claude Fable 5

• Publish the post-June-22 credit pricing now. Asking users to fall in love during a free window without telling them the wedding costs is the launch's one genuinely user-hostile choice.
• Make the Opus fallback louder. The reroute notice is easy to miss. When I'm paying Fable prices, I want an unmissable flag — and ideally a session-level toggle to fail closed instead of falling back.
• A Fable-tier usage meter. On Pro you discover the limits by hitting them. A visible remaining-budget indicator would turn frustration into planning.
• Match Opus latency on short tasks. The thinking overhead is worth it on hard problems and pure tax on easy ones; smarter routing of trivial queries would fix the everyday feel.

Final verdict — 4.7 out of 5

Fable 5 earns the highest rating I've given a model on this site. The deductions are specific: double-Opus API pricing that demands workload-by-workload justification, usage limits that undercut the "included free" framing, and a credit-pricing cliff on June 23 that Anthropic has left needlessly vague. None of that touches the core finding — on the hardest work I could throw at it in a day, this is the most capable AI tool publicly available, by a margin you can feel without a benchmark chart.

The advice is time-boxed: if you have a paid Claude plan, the next two weeks are a free trial of the frontier. Bring your worst problems. Decide with your own data whether the gap is worth 2x when the bill arrives.

FAQ: Claude Fable 5 review

What is Claude Fable 5?

Claude Fable 5 is Anthropic's most powerful publicly available AI model, launched June 9, 2026. It's the first public release of a Mythos-class model — Anthropic's internal frontier tier — with safeguards added for general availability. It exceeds every previous Claude model on benchmarks, with state-of-the-art results in software engineering, knowledge work, vision, and scientific research, and is available in the Claude apps, via the API as claude-fable-5, and on Amazon Bedrock.

Is Claude Fable 5 free?

Temporarily, on paid plans. From June 9 through June 22, 2026, Fable 5 is included at no extra cost on Claude Pro, Max, Team, and seat-based Enterprise plans — though normal usage limits still apply and tighten faster than with Opus. From June 23, it leaves those plans and requires usage credits. There is no Fable 5 access on Claude's free tier.

How much does the Claude Fable 5 API cost?

$10 per million input tokens and $50 per million output tokens — exactly double Claude Opus 4.8's $5/$25, and less than half what the earlier Mythos Preview cost its limited audience. It ships with a 1M-token context window and 128K max output at standard pricing. The model ID is claude-fable-5, and it's also available through Amazon Bedrock.

Why does Fable 5 sometimes answer as Claude Opus 4.8?

By design. Fable 5 ships with safeguards that block responses in specific high-risk areas — notably cybersecurity and biology — and in those cases your query is answered by Claude Opus 4.8 instead, with a notice. Anthropic says this triggers in under 5% of sessions. In my testing it fired once, on a vulnerability-analysis question. If your work concentrates in those fields, test against your real workload before paying Fable rates.

Is Claude Fable 5 better than GPT-5.5?

On the published engineering benchmarks, decisively: 80.3% vs 58.6% on SWE-Bench Pro, and 29.3% vs 5.7% on Cognition's FrontierCode. My one-day testing matched the direction of those numbers on real coding tasks. GPT-5.5 retains advantages elsewhere — ecosystem breadth and native image generation among them — but for software engineering and hard analytical work, Fable 5 currently leads the public field.

What is Claude Mythos 5 and can I get it?

Almost certainly not. Mythos 5 is the same underlying model as Fable 5 with some safeguards lifted, released only to a small group of cyberdefenders and infrastructure providers through Project Glasswing, a collaboration with the US government. For everyone else, Fable 5 is the Mythos-class model — the safeguarded public version.

Related reviews

• ChatGPT vs Claude vs Gemini — how the big-three ecosystems compare beyond this launch.
• Anthropic IPO 2026 — the business story behind the company shipping this model.
• Cursor 3 review — the coding IDE where a model jump like this lands hardest.
• AI Agents — topic hub — every frontier-model and agent review in one place.

Testing Fable 5 on something specific before June 22 and want a second opinion? Get in touch — reader questions shape the next round of reviews.