Anthropic released Claude Fable 5 on June 9, 2026. It is the first Mythos-class model made available to the public: Fable 5 and Claude Mythos 5 share the same underlying model, with Mythos restricted to approved organisations and Fable carrying additional safety measures for general availability. If your team is on a Claude Pro, Max, Team or per-seat Enterprise plan, you can use Fable 5 free until June 22. After that it moves to usage-based credits, and Anthropic has not confirmed whether it returns as a standard plan feature.
We covered the launch when it happened. This post is the practical follow-up for Australian businesses: is Fable 5 worth routing real work to, and which work? The short answer is that it is a specialist, not a daily driver, and the free fortnight is the cheapest evaluation you will ever run on it.
What Fable 5 costs
There is no published rate card yet for API access. Community reports put pricing at US$10 per million input tokens and US$50 per million output tokens. Treat those as unconfirmed, but they are consistent across multiple builders. If accurate, that is less than half what the old Mythos Preview cost, yet still roughly 1.7 to 2 times GPT-5.5 depending on your token mix, and about twice Claude Opus 4.8.
The bills move quickly on agentic workloads. A few community-reported data points:
One builder reported spending US$371 (about $570 Australian) in half a day of agentic coding, even with cache hit rates around 92 per cent.
A Max 20x subscriber reported exhausting a full five-hour usage window in 45 minutes.
Extend that arithmetic and a two-person Sydney dev team using Fable 5 for everything could plausibly clear $5,000 a month. That is illustrative, not a quote, but it shows why routing matters.
Where the benchmarks say it wins
These numbers are community-compiled rather than official, so read them as indicative:
SWE-bench Pro (complex code repair): Fable 5 scored 80.0 against GPT-5.5 at 58.6, a 21.4 point gap.
FrontierCode Diamond (the hardest frontier coding problems): 29.3 versus 5.7, roughly five times the score.
Terminal-Bench 2.1 (terminal agent tasks): 84.3 versus 83.4, effectively a tie.
The pattern matters more than any single number. The gap opens up on long-horizon, cross-module engineering work, the kind where an agent has to hold a large system in its head across hours of changes. On routine tasks, most frontier models feel about the same, which means paying a premium there buys you nothing.
What early users are reporting
Beyond benchmarks, the early field reports are specific and mostly consistent. One builder found that a flaky document-formatting skill which misbehaved across machines ran correctly on the first try under Fable 5, suggesting noticeably better instruction-following on complex procedures. Another reported that a project crash Opus 4.8 had worked on unsuccessfully for a week was fixed in ten minutes. Frontend and UI implementation quality also comes up repeatedly as a strength.
Two reports deserve extra weight. Stripe reportedly migrated a 50-million-line Ruby codebase in a single day, work estimated at more than two months for a human team. And Fable 5 is currently the highest-scoring model on Cognition's ViBench evaluation. One operational caveat: safety classifiers route some cybersecurity and biomedical topics back to Opus 4.8, under 5 per cent of sessions per the launch notes. If you build security tooling, test whether your workload trips this before you commit.
The routing rule: send it the hard problems
At reported pricing, Fable 5 does not make sense as the default model for everything. The sensible setup routes by difficulty: send hard, multi-module, long-horizon engineering work to Fable 5, and keep everyday drafting, analysis and routine coding on Opus 4.8, where the economics still work. A $570 half-day is a bargain if it closes out a problem that has blocked a release for a fortnight, and a waste if it spent the budget reformatting documents.
What to test before June 22
The free window is the test. Here is what we would put through it:
Pick your gnarliest stuck problem, the bug or migration nobody has cracked, and give Fable 5 a full session on it. This is where the model differentiates, so it is the fairest test.
Run the same task through Opus 4.8 in parallel and compare outcomes, not vibes. If both succeed, the cheaper model wins your routing table.
Track token burn during the trial so you can project a realistic monthly spend in Australian dollars before the meter starts.
If you work in security or biomedical domains, check whether the classifier rerouting touches your actual workload.
Revisit your build-versus-buy maths. If Fable 5 turns a two-month engineering project into a day, some projects you shelved as too expensive may now be viable.
Nothing about your data handling obligations changes with the new model: the Privacy Act applies to what you send any AI system, so the same governance you applied to Opus applies here.
If you would rather get a structured answer than spend the fortnight on ad hoc prompting, this is exactly the kind of evaluation we run for Australian businesses: same tasks, both models, measured outcomes and a routing recommendation. Book a short scoping call before the window closes on June 22.
