Photo credit: anthropic.comThe most powerful model Anthropic has ever sold went on general sale on 9 June, and the company bolted a governor to it before turning the key. Claude Fable 5 is the first Mythos-class model the firm has made available to everyone - a capability tier Anthropic places above its Opus flagships - with performance the company says exceeds any model it has previously released to the public, and a lead that widens the longer and more complex the task. The governor is the story. When a request touches offensive cybersecurity, sensitive biology or chemistry, or an attempt to copy the model’s capabilities, the answer quietly comes from the older Claude Opus 4.8 instead, and the user is told the handover happened.Strip away the launch-day noise, and a precedent sits underneath it. A frontier lab has decided the safest way to ship its best engine is to ship it detuned, with the high-compression parts locked out for anyone without clearance. Whether the rest of the industry copies that decision or routes around it is the question that outlasts this week.Key Takeaways Claude Fable 5 is Anthropic’s first public Mythos-class model, priced at 10 dollars per million input tokens and 50 dollars per million output - double Claude Opus 4.8.Fable 5 and the restricted Claude Mythos 5 are the same model. Safeguards alone separate them; risky queries on Fable 5 fall back to Opus 4.8 in under 5 per cent of sessions, billed at the lower Opus rate.Independent testing backs the coding claims: on the Every Senior Engineer benchmark, Fable 5 scored 91 out of 100 against Opus 4.8’s 63, according to AY Automate’s analysis.India sits inside the related Project Glasswing programme, with CERT-In and NCIIPC among the agencies understood to have access to the unrestricted Mythos line.Subscription access is free only until 22 June 2026; after that, Fable 5 needs usage credits.So what is Claude Fable 5, in plain terms? Fable 5 is one of a matched pair, and the pair shares a single engine block. Claude Mythos 5, launched the same day for a small set of cyberdefenders and infrastructure providers, is the same underlying model with the safeguards lifted in some areas, and Anthropic calls it the strongest cybersecurity model in the world. The two names mark the same hardware in two states of tune. Fable comes from the Latin fabula, “that which is told”, a cousin of the Greek mythos, and the safeguards are the only thing that earns the two models separate badges.Think of it the way a manufacturer thinks about homologation. The race team builds one engine. The version that reaches the public road carries a restrictor and a remapped ECU, so it stays street-legal; the version that stays in the paddock runs unrestricted for drivers with a licence to handle it. Fable 5 and Mythos 5 run the same weights - Mythos 5 is the unsafeguarded configuration, while the publicly callable Fable 5 adds the layer that falls back to Opus 4.8 in guarded domains, as the model-tracking site LLM Stats puts it.The road this engine drove in on matters. Anthropic first revealed Mythos in April 2026, finding it adept at locating vulnerabilities across every major operating system and web browser, and held it back to a select group of companies under a programme called Project Glasswing. Fable 5 honours what the company had called its “eventual goal” of deploying Mythos-class models at scale. Two months from paddock to public road.How much faster is the engine than Opus 4.8? On the launch-day numbers from Anthropic and its partners - treat these as the manufacturer’s own dyno sheet - the gains are large rather than marginal. Stripe reported that Fable 5 compressed months of engineering into days: in a 50-million-line Ruby codebase, it ran a codebase-wide migration in a single day that a team would have spent over two months completing by hand.The independent track times, now in, hold up better than launch dyno sheets usually do. On the Every team’s Senior Engineer benchmark - the same test the firm gives human senior engineers in its hiring loop - Fable 5 scored 91 out of 100 against 63 for Opus 4.8, a 45 per cent absolute jump on a test built to be hard for senior engineers, according to AY Automate’s analysis. The cross-model picture lands the same way. On SWE-Bench Pro, a test of real software engineering tasks pulled from public GitHub repositories, Fable 5 hit 80.3 per cent against 69.2 per cent for Opus 4.8, 58.6 per cent for GPT-5.5, and 54.2 per cent for Gemini 3.1 Pro, per figures reported by The Decoder. On Terminal-Bench 2.1, the model reached 88 per cent, ahead of GPT-5.5 at 83.4 per cent and Gemini 3.1 Pro at 70.7 per cent.BenchmarkFable 5Opus 4.8GPT-5.5Gemini 3.1 ProSWE-Bench Pro80.3%69.2%58.6%54.2%Terminal-Bench 2.188.0%82.7%83.4%70.7%Every Senior Engineer91/10063/100N/AN/ASources: Anthropic launch post; The Decoder; Weights & Biases; AY Automate, 9 June 2026The leash itself costs a fraction. Fable 5 trails Mythos 5 by a sliver on most coding benchmarks - 95 against 95.5 on SWE-bench Verified - with the system card attributing the gap to safety-fallback rather than capability, and Anthropic reports the two land within one to three points of each other on most evaluations outside the guarded categories.Where the gap against Opus narrows is as telling as where it widens. On smaller, well-scoped tasks - a one-function refactor, a tightly specified algorithm - the difference shrinks; it shows up most on multi-file refactors, debugging across module boundaries, and long async runs where the model plans, executes, reviews its own output, and iterates with zero supervision. The engine pulls hardest at the top of the rev range. For a quick errand, it is overkill; for the long haul, it changes what one person can finish.That stamina is the genuine shift. Fable 5 holds focus across millions of tokens, and when Anthropic had it play the deck-builder Slay the Spire with persistent file-based memory, its performance improved three times more than Opus 4.8’s, and it reached the final act three times as often.Why does it sometimes hand you a different model? It reroutes rather than refuses, and the mechanism is the most novel thing in the release. A set of classifiers - separate AI systems that scan each request - sits in front of the model, and when one flags offensive cybersecurity, sensitive biology and chemistry, or an attempt to distil the model into a competitor, the response is handled by Opus 4.8 instead.The system card discloses a fourth lock, the launch post left in the footnotes. Anthropic has added safeguards limiting Claude’s effectiveness on requests targeting frontier LLM development itself - building pretraining pipelines, distributed training infrastructure, or ML accelerator design - citing concern about accelerating other AI developers in building powerful systems while lacking commensurate safeguards, as AI researcher Nathan Lambert highlighted from the document. Read that twice. The company restricting its rivals’ access to AI-building help is also the company whose own engineers merged eight times the code they did in 2024. The restrictor plate, it turns out, also covers the engine-design manual.For most users, the governor never trips. More than 95 per cent of Fable sessions involve zero fallback, and for those, Fable 5 performs the same as Mythos 5. Anthropic concedes the trade: the safeguards are tuned cautiously and will sometimes catch harmless requests, a friction the company says it aims to reduce after launch. The honest verdict from independent reviewers is that the catch lands unevenly. In unguarded domains, Fable 5 is the best thing Anthropic ships; in guarded domains, it quietly becomes Opus 4.8 — invisible to most developers, but a real downgrade for teams doing legitimate security or life sciences work, as LLM Stats framed it. CyberScoop’s counterweight keeps the leash framing honest: even when Fable 5 draws its cybersecurity answers entirely from Opus 4.8, users still receive impressive - though hardly unique - dual-use cyber capability, since Opus 4.8 itself improved on prior models even while staying far below Mythos Preview. The governed car still does highway speeds.The lock itself appears to hold. Anthropic ran an external bug bounty over more than 1,000 hours, in which nobody found a universal jailbreak, and external red-teaming groups also failed. One caveat survives in the fine print, and it belongs in any straight account: the UK AI Security Institute made progress towards a universal jailbreak within a brief initial testing window.The system card’s alignment data lands on the reassuring side of the ledger. Anthropic’s automated assessment found Mythos 5’s level of misaligned behaviour - including deception and cooperation with misuse - was low and similar to Opus 4.8’s, and since the two share a model, Fable 5’s alignment profile matches. The sharper differentiator is resistance to prompt injection, the attack class that matters most once models run as autonomous agents. On the Grey Swan / UK AISI agent red-teaming benchmark with thinking enabled, Fable’s attack-success rate at k=100 was 4.8 per cent, against 9.6 per cent for Opus 4.8, 30.8 per cent for GPT-5.5, and 45.5 per cent for Gemini 3.1 Pro, per the system card figures reported by Digital Applied. For anyone wiring an agent into email or a browser, that table is arguably more consequential than any coding score.What did the unrestricted version already do? This is where the product stops being a chatbot upgrade and starts being an infrastructure event. In Project Glasswing’s first month, Anthropic and roughly 50 partners used the model to find more than ten thousand high- or critical-severity vulnerabilities across the world’s most systemically important software. The individual hauls read like a Q Branch inventory. Cloudflare found 2,000 bugs across its critical-path systems with a false-positive rate its team rated better than human testers; Mozilla found and fixed 271 vulnerabilities in Firefox 150, over ten times more than the previous version surfaced with Opus 4.6.One case sits close to an ordinary customer’s life. At a Glasswing partner bank, the model helped detect and stop a fraudulent 1.5 million dollar wire transfer after an attacker compromised a customer’s email and made spoof phone calls.The volume created a fresh problem, and Anthropic names it: security progress used to be limited by how quickly vulnerabilities could be found; it is now limited by how quickly humans can verify, disclose, and patch the large numbers the model surfaces. The gadget that finds every weakness is only half a defence when the people who fix them are outnumbered. Some open-source maintainers, already buried under low-quality AI-generated bug reports, asked Anthropic to slow its disclosures so they had time to design patches.Why ship these days after warning that AI is getting dangerous? Anthropic’s defence is a race-against-proliferation argument. It expects that within 6 to 12 months, many other AI companies will have Mythos-class models, and that those rivals could release them without safeguards. Ship the governed version first, the reasoning runs, and defenders get a head start before the unrestricted clones arrive - and that clock has already started ticking. Since Mythos appeared, OpenAI released its own cybersecurity-focused model, GPT-5.5-Cyber, rolled out to a large group of partners, per TechCrunch. A second clock runs alongside the safety one: the Glasswing expansion came a day after Anthropic filed confidentially for an IPO, following a 65 billion dollar funding round at a valuation near 1 trillion dollars.What does it cost, and how long is it free? Both models cost 10 dollars per million input tokens and 50 dollars per million output tokens - under half the price of the earlier Mythos Preview. Measured against the rest of the range, though, the premium bites: that is, double the price of Opus 4.8, and the cost alone may deter wide use at a moment when enterprises are exhausting yearly AI budgets early, as TechCrunch noted. The model carries a 1 million-token context window and a 128k-token maximum output, with batch pricing at 5 dollars and 25 dollars per million tokens.Subscription access works on a closing window. From launch through 22 June, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost; on 23 June, it leaves those plans and requires usage credits, with the company aiming to restore it as a standard feature once capacity allows. Developers reach it through the Claude-Fable-5 model string. One consumer-relevant footnote from the product page: rerouted requests are billed below Fable’s rates - the fallback to Opus 4.8 carries Opus billing rather than Fable’s premium. The governor trips, and the meter slows with it.One policy shift deserves more scrutiny than the price tag. Anthropic will require 30-day data retention on all Mythos-class traffic, across first- and third-party surfaces, using the data to defend against novel attacks rather than to train models. TechCrunch argues this could set an industry precedent in which access to more powerful models arrives bundled with mandatory retention framed as a safety measure - a clause every enterprise legal team should read before the engineers start spending tokens.What does this mean for India? India is already inside the tent on the cybersecurity side, which sharpens rather than settles the question. Anthropic expanded Project Glasswing to 150 organisations across 15 countries, including India, after the software industry body Nasscom and the Ministry of Electronics and Information Technology held meetings seeking access, Business Standard reported. CERT-In, the NCIIPC under the National Security Advisor’s office, and the Department of Telecommunications’ Digital Intelligence Platform are understood to have received or are in line to receive access to Mythos. The unrestricted defensive engine is reaching Indian critical-infrastructure guardians; the throttled public one is reaching everyone else.On cost, the economics point at exactly the layer where India earns its keep. At 50 dollars per million output tokens, Fable 5 prices itself for the long autonomous job - the codebase migration, the multi-day analysis - measured in engineer-months saved rather than tokens spent. If the Stripe-style migration figures replicate inside Indian IT services firms and global capability centres, the staffing maths shifts at the precise point where the services industry books its margin. The senior-engineer benchmark result is the one to watch here: a model scoring 91, where human seniors sit, reframes what a mid-tier engineering team has to hire for.Exposure runs the other way. The same proliferation forecast that justifies the launch is a warning for every Indian operator outside the Glasswing cohort - the mid-sized banks, the state utilities, the telecom vendors who build on the same open-source code Mythos has been quietly auditing. India’s named agencies hold the shield; the long tail of Indian infrastructure waits outside it, and the cheap, unguarded clones Anthropic predicts will reach attackers and defenders on the same day.What happens next? Watch one number and two dates. The number is the sub-5-per-cent fallback rate: hold it, and Anthropic has proved a frontier model can ship publicly with a governor most users never feel; let it drift upward as adversaries probe the classifiers, and the industry learns the opposite lesson. The first date is 23 June, when the free subscription window shuts and real demand shows itself at double the Opus price. The second is whenever the promised trusted-access programme for biology opens, the test of whether the governor loosens on schedule or stays bolted down. A lab that built a frontier engine, fitted a restrictor, and drove it onto the public road days before its own IPO has made a bet that the restrictor holds. Everyone building on top of it is now a passenger in that bet.Frequently Asked Questions1. What is Claude Fable 5? Claude Fable 5 is Anthropic’s first publicly available Mythos-class AI model, a capability tier the company places above its Opus flagships. It launched on 9 June 2026 with safeguards that route high-risk queries to the older Claude Opus 4.8, and prices at 10 dollars per million input tokens and 50 dollars per million output tokens.2. How is Fable 5 different from Claude Mythos 5? They share the same underlying model. Mythos 5 has safeguards lifted in areas such as cybersecurity and stays restricted to vetted government and infrastructure partners through Project Glasswing. Fable 5 keeps all safeguards active and is open to the general public.3. Is Claude Fable 5 better than GPT-5.5 and Gemini 3.1 Pro? On the coding benchmarks available at launch, yes. Fable 5 scored 80.3 per cent on SWE-Bench Pro against GPT-5.5’s 58.6 per cent and Gemini 3.1 Pro’s 54.2 per cent, and led on Terminal-Bench 2.1. Broader independent testing across other task types is still arriving, so the picture may fill out over the coming weeks.4. Can people in India use Claude Fable 5? Yes. Fable 5 is available everywhere through the Claude API and paid plans. Separately, India was added to the restricted Project Glasswing programme, with agencies including CERT-In and NCIIPC understood to be among those granted access to the more powerful Mythos line.5. Why does Fable 5 give some answers from Opus 4.8 instead? Its safety classifiers flag requests involving cybersecurity, biology, chemistry, model distillation, or frontier AI development and route them to Opus 4.8. Anthropic says this affects fewer than 5 per cent of sessions, and users are notified each time the handover occurs.6. Does the Opus 4.8 fallback cost the same as Fable 5? No. Rerouted requests are billed below Fable’s rates, so a query handled by Opus 4.8 avoids the Fable premium. Anthropic notifies users whenever a fallback occurs.end of article