Claude Fable 5 Lands in Max Plan and May Have Disproved a 40-Year-Old Math Conjecture

Anthropic has quietly rolled Fable 5 into the Claude Max subscription as a standard feature, capping usage at 50% of a subscriber's weekly limit before requiring paid credits. On the same day, a tweet claiming Fable 5 helped disprove the Jacobian conjecture began circulating, drawing attention to the model's mathematical reasoning ceiling.

Why it matters

These two signals, one about access and one about capability, land together in a way that matters for builders. Fable 5 is now the de facto premium tier inside Max, which means teams evaluating Claude for research-heavy or math-intensive workloads no longer need a separate contract to test it. At the same time, the Jacobian conjecture claim, if it holds up under peer review, would mark the first time a language model contributed to disproving a major unsolved problem in algebraic geometry. That is a different category of result than passing a coding benchmark.

Separately, the Wandr Benchmark launched today to evaluate research agents on tasks requiring both broad coverage and deep drill-down, exactly the profile Fable 5 is being tested against in the conjecture story. The timing is not a coincidence: the field is converging on harder evals as models get more capable. Explore more context in Agents.

What changes in practice

• Max plan subscribers get Fable 5 access immediately, but the 50% weekly cap is a real constraint for high-volume pipelines. Plan for credit top-ups if you are running batch research tasks.
• Math and formal reasoning workloads just got a credible new option. The Jacobian claim is unverified, but even a plausible attempt at that level of problem signals strong symbolic reasoning.
• Agentic shopping and task-execution demos like the Claude shopping agent are proliferating, suggesting the ecosystem is moving from "can it do this" to "how do we ship this reliably."
• Benchmark expectations are rising. Wandr's wide-and-deep framing will pressure teams to move beyond single-hop retrieval evals when assessing research agents.

How to use it

1. Audit your weekly token distribution. If your Max-plan workflows are already near the usage ceiling, Fable 5 tasks will eat into that 50% cap fast. Separate high-stakes reasoning calls from routine completions to preserve headroom.
2. Stress-test Fable 5 on your hardest formal reasoning tasks now. The conjecture story is a signal, not a guarantee, but it is worth running your most complex math, logic, or proof-adjacent prompts against it while access is included.
3. Adopt a benchmark like Wandr for internal agent evals. Single-turn retrieval tests no longer reflect what production research agents actually do. Map your agent's task profile to wide-search vs. deep-search requirements before you ship.
4. Watch the Jacobian claim closely. If independent mathematicians verify the result, update your priors on where to use LLMs in formal verification pipelines. If it does not hold, note the failure mode for your own high-stakes reasoning use cases.

"The model helped disprove the Jacobian conjecture" is either the most important sentence in AI research this year, or a cautionary tale about trusting unverified tweets. Either way, it tells you something about where Fable 5 is operating.

Fable 5 in Max is the access unlock; the Jacobian story is the capability signal. Both deserve your attention this week.