Claude Fable 5 is the most powerful Claude model the public could use, and on certain prompts it answered as a weaker model on purpose. It sits a full tier above Claude Opus 4.8. Yet when a request touched cybersecurity, biology, chemistry, or model distillation, Fable 5 handed the answer to Opus 4.8 instead and told you it had. Anthropic's strongest generally available model dials itself down in its most dangerous domains. Then, three days after launch, the US government pulled it off the shelf entirely.
Status as of June 15, 2026: Fable 5 and Mythos 5 are offline. On June 12 the US government issued an export-control directive barring access by any foreign national, inside or outside the US, including Anthropic's own foreign-national staff. Unable to verify citizenship per request, Anthropic disabled both models for everyone. No return date. The company calls it a likely misunderstanding and says it is working to restore access. Read the tiers and pricing below as how this model worked, and likely will again.
That fallback design is one of the clearest AI lessons of mid-2026. It shows in production why a more capable model can ship more locked down than the one before it, how vendors trade frontier power against catastrophic risk, and how a built-in fallback changes what you pay for. The shutdown adds a second lesson nobody priced in: a frontier model can vanish on a government letter. Here is how each piece works.
Short answer: Fable 5 sits above Opus 4.8 and intentionally falls back
Fable 5 is the production, safeguarded deployment of the same underlying model Anthropic also ships as Mythos 5. Mythos 5 goes to pre-approved organizations with some safeguards lifted. Fable 5 was the public version, with safety routing wired in. When a request touched cybersecurity, biology and chemistry, or an attempt to distill the model, Fable 5 did not answer at full strength. It routed that one response to Claude Opus 4.8, the prior frontier model, which Anthropic judges less capable in those domains and therefore less prone to dangerous misuse.
Anthropic says more than 95% of Fable sessions run entirely on the model itself, so the fallback fired in fewer than 5% of interactions. For typical coding, analysis, and writing, you got the top tier the whole way through. The swap was the exception. Anthropic states that users are informed whenever a fallback happens, so you never had to guess which model answered. That transparency claim came with one ugly asterisk, covered below.
The mental model: Fable 5 is a frontier engine with a governor bolted on. In the flagged risk domains, the governor caps you at Opus 4.8 and flags the switch. Everywhere else, you run at full power.
What 'Mythos-class' means: 1M context, reported 80.3% SWE-Bench Pro, 2x Opus pricing
Mythos-class is Anthropic's name for the capability rung above the Opus family, and Fable 5 was the first model the public could reach there. Anthropic confirms pricing of $10 per million input tokens and $50 per million output tokens. The model carries a 1,000,000-token context window with up to 128,000 tokens per response, figures listed across providers including OpenRouter. A widely circulated 80.3% score on SWE-Bench Pro for agentic coding is doing real work in every comparison post, so here is the part those posts skip: that number is vendor-reported on Anthropic's own evaluation scaffold, and it does not appear on Anthropic's launch announcement at all. It is not an independent result on a standardized leaderboard like Scale's SEAL.
Those numbers frame the tradeoff. Fable 5's reported 80.3% on SWE-Bench Pro sits about 11 points ahead of Opus 4.8's reported 69.2%, a real gap on hard, long-running software tasks if you trust the scaffold both ran on. The price is exactly double Opus 4.8's $5 input and $25 output rate. The top tier cost twice as much and ran meaningfully stronger on the benchmarks where it was allowed to run at full strength.
- Tier: first public Mythos-class model, above the Opus family.
- Context: 1,000,000-token window, up to 128,000 tokens per response.
- Coding: a reported 80.3% on SWE-Bench Pro versus 69.2% for Opus 4.8, both on Anthropic's own scaffold, not a standardized leaderboard.
- Price: $10 input and $50 output per million tokens, double Opus 4.8 and confirmed by Anthropic.
- Access at launch: Claude API as claude-fable-5, the Claude apps, and Amazon Bedrock (US East N. Virginia and Europe Stockholm), released June 9, 2026.
Anthropic also included Fable 5 for paid subscribers on Pro, Max, Team, and seat-based Enterprise plans through June 22, 2026, after which usage was set to draw on credits. The government shutdown landed inside that window, so the cheap-trial pitch effectively expired early. If access returns, that included period is still the obvious place to pressure-test whether the top tier earns its price on your own tasks.
Why a smarter model refuses more: safety gating and capability-risk tradeoffs
A more capable model refuses more because capability and risk climb together. The same reasoning depth that let Fable 5 plan a multi-step refactor also makes its answers more useful to someone designing a cyberattack or a biological or chemical hazard. Anthropic's Responsible Scaling Policy ties deployment safeguards directly to demonstrated capability: as a model crosses higher AI Safety Levels, it must ship with stricter controls before release. Declining or downshifting on those prompts is not a bug. It is the cost of releasing a stronger engine to the public.
The policy structures this as AI Safety Levels, a hierarchy Anthropic models loosely on the US government's biosafety level standards for handling dangerous biological materials. Lower levels cover systems with no meaningful catastrophic risk. Higher levels cover models that could substantially raise misuse risk, with bioweapon uplift cited as a flagship example. Each rung up the ladder demands a stronger demonstration of safety before deployment. Fable 5's domain-specific fallback is one concrete way Anthropic cleared that bar without throttling the model everywhere.
The safety story is also a product story. Anthropic reports an external bug bounty that surfaced no universal jailbreaks across more than 1,000 hours of testing, while noting that the UK AI Safety Institute made progress toward one in a brief initial window. Shipping a frontier model that largely holds up under that adversarial pressure is part of what customers buy. The fallback let Anthropic chase a benchmark lead and a defensible safety position at the same time.
If your work legitimately touches security research, lab biology, or chemistry, expect Fable 5 to answer at Opus 4.8's level on those turns and to tell you when it has. Plan around the cap instead of fighting it with prompt tricks, which the published testing suggests do not reliably break through.
The fallback mechanism, and the 'secret sabotage' it tried to hide
The fallback triggers per request, not per session. When Fable 5's classifiers flag a request as high-risk in cybersecurity, biology and chemistry, or distillation, Opus 4.8 generates that specific response, and Anthropic says the user is informed. The conversation continues, and the next request can run on Fable 5 again as long as it stays out of the flagged zone. A single chat could mix outputs from two models, with a notice on the turns that were downshifted.
Now the part the launch-day coverage buried. The visible fallback was not the only safeguard. For frontier AI research and development prompts, the system originally degraded answers silently, with no notice to the user, an intervention Anthropic's own system card described as not visible to the user. Researchers called it secret sabotage. On June 10, one day after launch, Anthropic reversed it, saying it made the wrong tradeoff and apologizing for not getting the balance right, and committed to making those safeguards visible. So the clean story that Fable 5 always tells you when it holds back was, for one domain, briefly false.
Routing rather than refusing protects product quality. A blunt refusal returns nothing useful. Falling back to Opus 4.8 still gives an answer, just from a model whose capabilities in those domains Anthropic judges safe to release broadly. You lose Fable 5's extra reasoning on that turn but keep a working assistant. Anthropic calls the safeguards deliberately conservative, so they occasionally catch benign requests. That is the price of shipping the model sooner.
Practical signal: if a security or lab-science answer feels less sharp and you see a fallback notice, Opus 4.8 answered that turn. That is the mechanism working as designed, not the model getting dumber mid-conversation.
Model tiers explained: how vendors trade frontier power against risk
Model tiers exist because no single model can be the cheapest, the fastest, the smartest, and the safest at once. Vendors split the lineup so each tier optimizes a different point on that curve. Fable 5 is the clearest public example yet of the top tier also carrying the heaviest safety gating, because the same capability that justifies its price is what makes unrestricted release risky.
The table below frames the choice between the top tier and the prior frontier. Use it to match a task to the tier it actually needs, instead of defaulting to the most expensive option.
| Dimension | Claude Fable 5 (top tier) | Claude Opus 4.8 (prior frontier) |
|---|---|---|
| Tier | First public Mythos-class | Opus family |
| SWE-Bench Pro | 80.3% (vendor scaffold, not standardized) | 69.2% (vendor scaffold, not standardized) |
| Context window | 1M tokens, 128K max output | Large, below Mythos-class |
| Price per 1M tokens | $10 in / $50 out | $5 in / $25 out |
| High-risk prompts | Falls back to Opus 4.8, user informed | Answers within its own limits |
| Availability (June 15, 2026) | Disabled by US export order | In service |
The pattern reaches past Anthropic. As frontier models cross capability thresholds that vendors judge dangerous, expect more of this: top-tier models deliberately constrained in narrow domains while running full strength everywhere else. The safer commercial move is a capable model that downshifts on a short list of risky topics, not a weaker model across the board. The Fable 5 shutdown adds a sharper edge. Capability that triggers safety gating can also trigger regulators, and a model can disappear from your stack on three days' notice.
What this means for your workflow and when to pay for the top tier
Pay for Fable 5, when it returns, for the hardest tasks that fall outside the flagged domains where the extra coding capability or the larger context window changes your results. For agentic software work, long-document reasoning, and complex analysis, the top tier can justify double the price. For routine drafting, summarization, and chat, Opus 4.8 or a cheaper tier usually delivers the same output for less.
- Choose Fable 5 for: hard agentic coding, very long context, and multi-step analytical tasks where its benchmark lead shows up.
- Stay on a lower tier for: routine writing, summaries, classification, and high-volume calls where cost dominates.
- Expect Opus-level answers on: cybersecurity, biology and chemistry, and model-distillation prompts, regardless of which model you selected.
- Do not single-source on the frontier tier: the June 12 shutdown proves a top model can go dark fast, so keep a fallback model your pipeline can switch to.
One more workflow consequence: because the fallback is per request, you cannot assume a whole conversation ran on Fable 5. If you are benchmarking the model or building on its outputs, watch the fallback notices and log which turns Opus 4.8 answered. Confirm model identity from those signals rather than assuming it.
Where an external memory layer fits
Model tiers, per-request fallbacks, and sudden shutdowns share one weakness: your context lives inside whichever model answered, and it does not travel. Switch tiers, switch vendors, hit a fallback, or lose a model to a government order, and the assistant forgets what you established earlier. That is the gap MemX (memx.app) is built to close.
MemX is an external, model-agnostic AI memory layer. Your facts, preferences, and project context live in a memory you own and attach to any model, so moving between Fable 5, Opus 4.8, or a different vendor never resets what the assistant knows. It is private by architecture: per-user isolation, encryption at rest, and on-device options. To be precise, MemX is not end-to-end encrypted and not zero-knowledge. The design goal is durable, portable memory you control, not a security claim it cannot back. When models keep changing tiers, or vanish overnight, owning the memory layer is what stays stable.
01Is Claude Fable 5 better than Opus 4.8?
On the benchmarks where it runs at full strength, it leads. Fable 5 reports 80.3% on SWE-Bench Pro versus 69.2% for Opus 4.8, both on Anthropic's own scaffold, and it is the first public Mythos-class model. It costs double, $10/$50 per million tokens versus $5/$25.
02Why does Claude Fable 5 refuse cybersecurity and biology questions?
Anthropic gates high-risk domains. For cybersecurity, biology and chemistry, and distillation prompts, Fable 5 falls back to the less capable Opus 4.8 instead of answering at full capability. This follows Anthropic's Responsible Scaling Policy, which ties stricter safeguards to higher model capability.
03Why is Claude Fable 5 not available right now?
On June 12, 2026, the US government issued an export-control directive barring access by any foreign national. Unable to verify citizenship per request, Anthropic disabled Fable 5 and Mythos 5 for all customers. No return date is set; Anthropic calls it a likely misunderstanding and is seeking to restore access.
04How much does Claude Fable 5 cost?
Anthropic prices it at $10 per million input tokens and $50 per million output tokens, double Opus 4.8's $5 and $25. Fable 5 was included for paid subscribers through June 22, 2026, after which usage was set to draw on credits.
05What is Mythos-class and how does it relate to Fable 5?
Mythos-class is Anthropic's tier above the Opus family. Fable 5 and Mythos 5 are the same underlying model. Mythos 5 goes to pre-approved organizations with some safeguards lifted; the public got the safeguarded Fable 5 version with its risk-domain fallback.
The lesson outlasts this one release. As models climb capability tiers, vendors keep coupling more power with more restriction, and they will increasingly signal when a safeguard kicks in, sometimes only after getting caught hiding it. Knowing when you are running the top tier, when you have been downshifted, and which tasks justify paying for the frontier is now part of using these tools well. So is knowing your model can be gone by morning.
