Claude Fable 5: Review, Benchmarks and Pricing

On June 9, 2026, Anthropic released two models built on the same weights: Claude Fable 5 and Claude Mythos 5. Mythos 5 is the unsafeguarded, restricted-access configuration. Claude Fable 5 is the one you can actually call: the same Mythos-class model wrapped in production safeguards and shipped through the public API.

Anthropic describes Mythos as a new capability tier above Opus, and the numbers back it. Fable 5 posts 95.0% on SWE-bench Verified, 80.0% on SWE-bench Pro, a leading 1932 Elo on GDPval-AA, and takes #1 on FrontierCode. The price is $10 / $50 per million input / output tokens, double Opus 4.8 but less than half what the Mythos Preview cost. The catch is the safeguard layer, and it matters more than any single benchmark.

At a Glance

• Release date: June 9, 2026. Generally available.
• Model ID: claude-fable-5 on the Claude API.
• Tier: Mythos-class, a step above Opus. Same weights as the restricted Claude Mythos 5.
• Pricing: $10 per 1M input tokens, $50 per 1M output tokens.
• Context window: 1M input tokens / 128K output tokens.
• Modalities: Text + vision input, text output. Extended thinking supported.
• Safeguards: Classifiers for cyber, bio/chem, and distillation route to a Claude Opus 4.8 fallback.
• Deployment: Claude API, Claude apps, and major cloud platforms.

Fable 5 vs Mythos 5

This release is unusual because it ships one model under two names. The weights are identical. What differs is the wrapper around them.

Claude Mythos 5 is the raw model with restricted, trusted access, the configuration Anthropic uses internally and with vetted partners. It carries the full capability profile, including in domains Anthropic considers sensitive. Claude Fable 5 is the same model made safe for general use. When a request touches a guarded domain, Fable 5 does not answer with Mythos-class capability. It falls back to Claude Opus 4.8.

The practical consequence: a single published benchmark table can carry two very different stories depending on whether the task trips a safeguard. Anthropic reports that Fable 5 and Mythos 5 land within one to three points of each other on most evaluations. The exceptions are the guarded categories, where Fable 5 behaves closer to Opus 4.8. Keep that split in mind for the rest of this review.

Benchmarks

All scores are self-reported by Anthropic in the launch announcement and the system card. The cleanest read on Fable 5 is the set of benchmarks that do not trip a safeguard, where the published number is genuinely Fable 5 and not a fallback. On those, it is a tier above Opus 4.8.

Software engineering and agentic work

The standout is FrontierCode, a hard, less-saturated coding set where Fable 5 takes first place at 29.3% on the Diamond subset (46.3% on Main) and, per the system card, beats every other model even at medium effort. SWE-bench Verified is approaching its ceiling, so the +10.8 points on SWE-bench Pro over Opus 4.8 is the more meaningful coding signal.

Knowledge work

GDPval-AA, Artificial Analysis's knowledge-work Elo, puts Fable 5 in front at 1932, ahead of Opus 4.8 and well clear of the GPT-5.5 and Gemini fields. The absolute numbers on AutomationBench and the Legal Agent Benchmark are low across the board, which is the point: these are deliberately hard, long-horizon business workflows where every model still has headroom, and Fable 5 leads them.

The safeguarded benchmarks

On the guarded categories, the launch table reports the Mythos 5score. Fable 5's real-world result is lower because its safeguards trigger and the trajectory falls back to Opus 4.8. The clearest measured example is Terminal-Bench, where 20.9% of trials hit a safety refusal and reverted to Opus 4.8 for the rest of the run.

So when you see Fable 5 quoted at 78% on a cybersecurity benchmark, read it as the Mythos 5 number. The model you call through the public API will, on those tasks, perform much closer to Opus 4.8.

Safeguards & Fallback

The safeguard layer is the defining feature of Fable 5, not a footnote. Anthropic runs classifiers over requests and, when one fires, declines to serve the full Mythos-class model. What happens next depends on the surface you are calling from.

Three classifier categories are in scope: cybersecurity, biology and chemistry, and model distillation attempts. In Claude client apps, a flagged request transparently falls back to the latest Claude Opus model and the user still gets an answer. On the Messages API, the default is to block: developers have to implement or opt into the Opus fallback themselves, otherwise the request is refused outright.

A separate set of competitive-use safeguards targets work that looks like frontier LLM development. Anthropic estimates these affect roughly 0.03% of traffic, so most users will never see them. The cyber, bio, and chem classifiers are the ones worth testing against your own workload before you commit.

Pricing & Availability

At $10 / $50, Fable 5 sits at double Opus 4.8's rate and 2.5x below the Mythos Preview's $25 / $125. For teams that were priced out of Mythos-class capability, that is the headline: a tier above Opus, generally available, at a rate that fits a production budget. See Anthropic's pricing page for batch and caching discounts.

Using Fable 5

Fable 5 is a drop-in model id on the Claude API. Three things are worth checking before you route production traffic to it.

Decide your fallback posture

If you call the Messages API, the safeguard default is to block, not to fall back. If your application can tolerate an Opus 4.8 answer on guarded requests, opt into the fallback so users are not hit with hard refusals. If correctness on those requests matters, handle the refusal explicitly rather than silently degrading.

Reserve it for the hard jobs

At $10 / $50, Fable 5 is not the model to put behind every call. Its edge over Opus 4.8 shows up on long-horizon coding, agentic workflows, and dense knowledge work. For routine traffic, Opus 4.8 or a Sonnet-class model is the better cost fit. Route to Fable 5 where the task is genuinely at the frontier.

Budget for output tokens

Like every recent frontier Claude release, Fable 5 leans on extended thinking and high effort to reach its top scores. Output token consumption rises accordingly. Measure token-per-task on your own traffic before extrapolating cost from the per-token rate.

Outlook

Fable 5 is the first time a Mythos-class model is available to the general public, and Anthropic has clearly thought hard about how to do that without handing out a frontier offensive-cyber tool. The result is a genuinely strong model with an asterisk: in unsafeguarded domains it is the best thing Anthropic ships, and in guarded domains it quietly becomes Opus 4.8.

For most developers that trade is fine, even invisible. The work that benefits most from Fable 5, large-scale software engineering and knowledge work, sits well outside the guarded categories. The teams that need to weigh it carefully are the ones doing legitimate security or life-sciences work, where the fallback turns a tier-above-Opus model back into Opus. For the full methodology and per-benchmark detail, see Anthropic's launch post and system card, and compare it on LLM Stats.