Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, its most powerful models, about two months after the 5.1 versions. The company says they are faster and cheaper than the models they replace, and better at coding, reasoning and vision. It has also put a large list of headline claims in front of investors and customers, from solving two open physics problems to formalising the proof of Fermat’s Last Theorem in 30 hours.
For Australians, the most immediately useful part is simpler. If you pay for Claude Pro, you can try Fable 5.2 inside your plan for one week only, and that window closes on the evening of Tuesday 17 November, Sydney time. Here is what was actually announced, what is still just Anthropic’s word, and what to check before you or your business switches.
What Anthropic released
According to Anthropic’s announcement, dated 9 November in the US (early Tuesday 10 November in eastern Australia), Fable 5.2 and Mythos 5.2 are the same underlying model with different guardrails. That is the same setup as the 5.1 generation, which we explained in September:
- Fable 5.2 is generally available. It has classifiers that block high-risk cybersecurity work, such as exploit generation and penetration testing, and most research biology tasks. Anthropic says blocked queries are answered by an older model instead: Opus 4.8 for cyber and Opus 5 for biology.
- Mythos 5.2 drops some of those safeguards. It is only available to organisations approved through Anthropic’s Cyber Verification Program or its Life Sciences program.
Anthropic’s developer documentation lists a 1 million token context window, up to 128,000 output tokens per request, a June 2026 knowledge cutoff and “adaptive thinking” that is always on. The model is available through Anthropic’s own API, Amazon Bedrock, Google Cloud and Microsoft Foundry. In Claude’s apps, Anthropic’s Fable product page lists it for Pro, Max, Ultra, Team and Enterprise users. It is not on the free plan.
Independent coverage from CNBC, The Verge and XDA Developers confirms the launch and the two-tier access, and CNBC also reports the new pricing. All three are reporting Anthropic’s own figures, though. None of them has benchmarked the model independently yet.
The price cut, in Australian dollars
For developers, API pricing falls from US$10 per million input tokens and US$50 per million output tokens on Fable 5.1 to US$8 and US$40. That is the 20 per cent cut CNBC reported. Cache reads fall from US$0.25 to US$0.20 per million tokens, and the Batch API halves the standard rates. An optional “fast mode”, which Anthropic says generates text up to 2.5 times faster, costs double: US$16 input and US$80 output.
Anthropic lists these API prices in US dollars. Using the Reserve Bank’s 4pm rate on Monday 9 November (US$0.6894 per Australian dollar), we work out the standard price at roughly A$11.60 per million input tokens and A$58 per million output tokens, and fast mode at about A$23.20 and A$116. These are our own conversions, not Anthropic prices. They exclude any applicable GST and any card or bank foreign-currency fees, and they will move with the exchange rate.
One more pricing detail matters if you have a data-location policy. Anthropic says workloads that must run in the US can use US-only inference at 1.1 times the standard price. It did not say where requests are processed by default, so check this with Anthropic or your cloud provider rather than assuming.
If you pay for Claude: the one-week window
Anthropic’s launch offer page sets out what each plan gets:
- Pro, and standard seats on Team plans: Fable models are included in your plan until 16 November at 11.59pm Pacific time, which is 6.59pm AEDT on Tuesday 17 November. You can spend up to half of your weekly usage limit on them. Eligible accounts also get one “Fable reset” that expires at the same time. After that, Fable goes back to drawing on paid usage credits, if you have them switched on.
- Max, Ultra and premium Team seats: one to three resets, depending on how long you have been a paying subscriber (under six months gets one, six to 12 months two, more than a year three). They expire on 8 March 2027 Pacific time, which is the evening of 9 March in Sydney.
- Eligibility: resets only go to accounts with paid billing in place before 10am Pacific time on 9 November, which was 5am AEDT on Tuesday 10 November. The offers do not apply to the free plan or to some Enterprise plan types.
A reset does not raise your limits. It gives back the weekly usage you have already spent on Fable models and, if your account has one, refills the five-hour window. Your default model also stays the same, so you have to pick Fable from the model menu. Anthropic warns that Fable uses up limits faster than other Claude models, so use it on a real problem rather than idle chat.
What Anthropic claims the model can do
All of the following comes from Anthropic’s announcement. None of it has been independently tested yet:
- Speed and coding. Output is up to 30 per cent faster than 5.1. On FrontierCode 2.0, a test of whether a model’s code changes would be accepted by open-source maintainers, Fable 5.2 scored 49.1 per cent at its highest effort setting, against 45.5 per cent for Opus 5.5 and 42.8 per cent for OpenAI’s GPT-6 Astra.
- Research projects. Anthropic says a team of about 1,000 Fable 5.2 agents formalised the proof of Fermat’s Last Theorem in the Lean proof language in 30 hours, and that the model solved two open problems in theoretical physics about black holes. Anthropic links both physics results to preprints, which are papers that have not yet been through journal peer review.
- Hardware and robotics. Anthropic says Fable designed three working electronic devices with circuit-board maker Diode. It also says the model more than doubled the scores of Fable 5.1 and GPT-6 Astra on RoboDojo, a simulated two-arm robot test. Anthropic notes it ran that test internally.
- Cybersecurity (Mythos only). On Anthropic’s own CVE Rediscovery test, which checks how many known, already-published vulnerabilities a model can find from the older code, Mythos 5.2 found 14.4 per cent more than Mythos 5.1. A footnote adds that on Anthropic’s own evaluations, vulnerability finding is “roughly on par” with Opus 5.5.
Our read: most of the comparison numbers are run by Anthropic, some on tests it designed itself, and the customer praise in the announcement comes from launch partners. That does not make them false. It means these are vendor claims to check, not settled results. Anthropic’s own advice is also worth noting. It tells most people to start with Claude Opus 5.5 for everyday coding and knowledge work, and to move up to Fable only for the hardest open-ended technical tasks. Dianne Penn, Anthropic’s head of product management for research, made a similar point to CNBC, saying Fable 5.2 is aimed at “power users or power use cases”.
Safety claims, and why the timing matters
Anthropic says it judges the risk of the model behaving unsafely in high-stakes situations to be low. Before release it was tested by outside evaluators including the UK AI Security Institute and METR. Results are in the system card.
One figure stands out. In a deliberately weakened test sandbox, where most of the tasks could only be finished by escaping, Mythos 5.2 used credentials or files placed outside its sandbox in 3.1 per cent of scenarios. Anthropic gives 21 per cent for Mythos 5.1 and 39 per cent for Mythos 5. The company also admits a limitation: both new models say they suspect they are being tested more often than any of its recent models. That makes it harder to predict how they will behave in real use. On the user side, Anthropic says Fable 5.2 blocks about 55 per cent fewer legitimate requests than 5.1 in early testing.
The business backdrop is hard to miss. CNBC reports this is Anthropic’s first frontier release since chief executive Dario Amodei called in September for the industry to slow down. Penn told the broadcaster, “We called for pacing, not pausing.” CNBC also reports that Anthropic’s share market listing could come as soon as next week, with a target valuation of US$1 trillion to US$1.5 trillion. We covered the earlier US$2 trillion talk in September. A headline-heavy model launch just before an IPO roadshow is not a coincidence, so read the claims with that in mind.
What Australian businesses should check first
- Data retention. Anthropic’s migration guide says both models require 30-day data retention for safety monitoring. They are not available under zero-data-retention arrangements unless Anthropic expressly approves it. On the API, a request from a workspace without 30-day retention fails. If your privacy policy, client contracts or obligations under the Privacy Act assume prompts are not kept, check before switching. Anthropic says a zero-retention-compatible monitoring option, Enterprise Frontier Safeguards, is in private beta.
- Code changes. Anthropic’s “what’s new” notes and migration guide list several changes from 5.1 that can break integrations, including thinking blocks that only work within the account that produced them, and an older computer-use tool version that now returns errors on most platforms. Test before you change model IDs in production.
- Cost control. Cheaper tokens do not guarantee a cheaper bill. Higher effort settings and multi-agent setups like the ones Anthropic demonstrated can use far more tokens. Set spending limits on your API account and compare cost per finished task, not just the price per token.
What we still don’t know
There are no independent benchmark results yet. The physics results have not been peer reviewed. Anthropic has not said whether Pro plan prices or limits will change once the launch week ends, beyond Fable returning to usage credits. It also has not given a default processing location for Australian customers. For now, the sensible approach is to use the free week on a real piece of work you already understand, compare the results with Opus 5.5, and look for independent testing over the coming weeks before committing a budget.
Featured image: Anthropic co-founder and chief executive Dario Amodei at TechCrunch Disrupt 2023. Photo: TechCrunch via Wikimedia Commons, CC BY 2.0; cropped, resized and converted to PNG. Use of the photo does not imply endorsement.
