Claude Fable 5 is here. What it means for your work, in plain terms.

Anthropic released Claude Fable 5 on June 9. It is the most capable model they have ever made publicly available, and if you are on a paid Claude plan, you can use it right now at no extra cost, until June 22.

This post answers the questions that matter for people who use AI for client work, analysis, and writing. Not the coding benchmarks. Not the AGI debate. What it does for your week.

What is Fable 5, in one paragraph

Anthropic built a model called Mythos 5 that they consider too capable to release openly, staying restricted to vetted security and biomedical researchers. Fable 5 is the same model with a safety layer on top. Requests touching dangerous territory (think bioweapons, cyberattack tooling) get quietly rerouted to the older Opus 4.8 model. Anthropic says this happens in fewer than 5% of sessions. For consulting, finance, sales, and marketing work, you will essentially never hit it.

What it is best at

The pattern across every early benchmark and test: Fable 5 pulls ahead on long, complex, multi-step work. On short, simple prompts, you will barely notice a difference from Opus 4.8.

Where the gains are real:

Financial analysis. Fable 5 scored highest of any model on Hebbia's Finance Benchmark, which tests reasoning at senior-analyst level: reading documents, interpreting charts and tables, solving problems across them. Trading firm IMC reported it passed their trading-analysis evaluations almost across the board.

Data analysis. First model to break 90% on Hex's analytics benchmark, a 10-point jump over Opus 4.8 on long-running analytical tasks that need judgment, not just computation.

Large document sets. It holds focus across millions of tokens and writes its own working notes as it goes. If your work involves a 40-page report, six competitor documents, and a regulatory framework all at once, this is the upgrade you will feel.

Reading visuals. Big jump in chart, table, and screenshot interpretation. In one test it rebuilt a working web app from screenshots alone. Practical version: it reads the graphs in your client's PDF correctly more often.

Contract review. Lawyers in blind tests found its contract redlines matched or beat specialized legal AI tools. It is already live inside Harvey, the legal AI platform.

Long-running tasks. Anthropic markets "days-long" autonomous work. One caveat the press coverage skips: that applies to automated pipelines built on the API, not to you chatting in the Claude app. In normal use, what you get is a model that stays coherent much deeper into a long working session before it starts losing the thread.

What it is not better at

Honest list, because the launch coverage will not give you one.

Quick drafts and short prompts. Early reviewers at Every.to found the model's slower, deeper thinking "works against the rapid-iteration rhythm most writing depends on". If your main use is drafting emails and punching up copy, Opus 4.8 is faster and half the price.

Marketing and sales writing specifically. No independent test of proposal writing, marketing briefs, or sales decks has surfaced yet. The finance and analytics gains are confirmed. The writing gains are not. We will know more in a few weeks.

Finding a use for it at all. The same Every.to review noted that heavy AI users found it paradigm-shifting on their hardest tasks, while lighter users "struggled to find something to use it for". If you have not yet built AI into a real workflow, a stronger model does not change that. The bottleneck is the workflow, not the model.

The privacy catch Dutch professionals should know

Two things matter here if you handle client data.

Mandatory 30-day retention. All Fable 5 traffic is retained by Anthropic for 30 days for safety monitoring, including for customers who previously had zero-retention agreements. No EU regulator has ruled on how this sits with GDPR data minimization yet. If you process confidential client data through AI, this is a real change, not fine print.

No EU data residency on the direct API. Anthropic is GDPR compliant through its data processing agreement, but compliance is not the same as residency: Anthropic's own API runs inference in the US only, for every model, not just Fable 5. If EU processing matters for your contracts, the routes are AWS Bedrock EU endpoints or Google Vertex AI EU regions. Claude Enterprise offers EU data storage (Frankfurt, Dublin), but that predates this release.

None of this blocks you from using Fable 5 at claude.ai today. It does mean: do not paste sensitive client documents into it without checking your processing agreements first. Same rule as always, now with a sharper edge.

What it costs

Access	Price
Claude Pro / Max / Team / Enterprise	Included free through June 22
After June 22	Removed from plans, requires usage credits (pricing not announced)
API	$10 per million input tokens, $50 per million output, 2x Opus 4.8
Free plan	Not included

The June 22 cliff is the thing to plan around. Anthropic says they want to restore it as a standard plan feature "as soon as possible" but gives no date. Translation: the next two weeks are a free trial. Use them deliberately.

Should you switch today?

If you are on a paid Claude plan: yes, try it; it costs you nothing until June 22. But try it on the right task.

Do not test it with "write me an email." You will not see the difference and you will conclude it is hype.

Test it with the hardest analytical task on your plate this week. The 60-page tender you need to compare against three competitors. The quarterly data nobody has time to dig through. The contract you have been putting off reading properly. That is where this model separates itself.

And if the result does not impress you, that is useful information too: it means your bottleneck is setup and workflow, not model quality. Which, conveniently, is fixable.

Want to build AI into your actual workflows; your templates, your client work, your data? That is what we do in our hands-on workshops. You bring a real workflow, you leave with it running. Reserve a spot at aibl.to.