Inside Fable 5 at Anthropic: Catherine Wu on Faster Model Shipping and AI Workflows

Catherine Wu, who leads product for Claude Code and Cowork, describes Fable 5 as "the strongest coding model we’ve ever had." She says it is especially well suited to the kinds of work earlier models struggled with: long-duration tasks and careful changes inside complex codebases.

In areas where teams previously had to specify every detail, Fable 5 can now handle conversations at the level of "Here is the problem, we have a few options, what should we do?" It can brainstorm like a design partner, align on an approach, and then carry execution through with something close to human-level success rates. Wu says that when she faces a complex task, her first step now is to brainstorm with Fable 5.

In the program, Wu highlights internal use cases such as a "self-doing to-do list" in Claude Desktop, where Claude is launched for each task, and automated assembly of marketing and sales materials by feeding in multiple product specs and launch assets to generate presentation decks.

The most notable example is customer support triage. Anthropic is already running workflows at the level of: monitor this channel, and whenever a customer issue comes in, identify the root cause, find the engineer who last touched that part of the codebase, tag that person in an issue, and generate the fix PR as well. Wu says this kind of support is increasingly important because teams get too large for any one person to track what everyone has changed. Fable can read the code, identify who touched it most recently, and explain the context behind that change.

Anthropic had already been shipping new models at roughly one per month before Fable 5. Wu attributes that pace in part to productivity gains from Quad, including Claude Code and Cowork. She cites an internal metric that, compared with the 2021-2025 baseline, employees ship eight times more code per person after adopting the latest models.

The workflow of handing a final job to an agent before bed and waking up to completed work is also spreading inside the company. The message is simple: humans work eight hours, but agents can work 24. Anthropic encourages teams to increase automation, replace repetitive tasks, and reserve human time for the interesting work. At the same time, the company insists on close supervision, warning that the first pass at automation fails about 30% of the time. Humans are expected to inspect the errors, fix the prompts and harnesses, and keep supervising until stability reaches 100%.

Wu repeatedly says that safety drives every decision at Anthropic. Her example is Auto Mode, released earlier this year. Internal data showed that users hit "Yes" on permission prompts 97% of the time, often within milliseconds, which led the company to consider handing those permission decisions to Claude itself.

Anthropic accumulated logs of both "Yes" and "No" decisions and tuned the classifier until it could make better judgments than a human. The company then hired external red-team firms to attack the system under adversarial scenarios and only released it publicly after hardening the classifier. Wu says the concept had existed since December of the previous year, but Anthropic waited months to confirm it was safe enough.

On token costs, she says the product direction is to make each user’s consumption visible. Her rule of thumb is: if it is a task you would otherwise want a human to do, use Claude. At the same time, users should exercise judgment and avoid waste, such as asking for 1,000 options when 10 is enough. She also notes that Fable 5 was priced below the Mythos preview because Anthropic wanted it to be usable for everyday work.

Inside Anthropic, Claude Code is used primarily by engineers, some PMs, designers, and data science staff. Cowork is the main tool for teams whose deliverables are not code, including marketing, legal, and sales.

But Wu expects that boundary to blur as the models improve. One example featured in the segment is a lawyer on the legal team who built an internal app with Claude Code that answers which member of the Product Counsel team should handle a product-related legal question. Wu predicts that within the year, teams will likely have Claude build role-specific Cowork interfaces for them. She closes with a challenge to engineers: if Claude is not getting stuck, you probably still are not giving it hard enough tasks.