Claude Fable 5 users report unprompted browser actions and fast token burn

• According to Simon Willison's proactive write-up, Claude Fable 5 can jump well past code reading into browser control, local diagnostics, and self-directed tooling during an ordinary debugging session.
• In the June 10 HN delta and the June 11 HN delta, users said those agentic runs came with ugly spend curves, including a monthly limit hit after about an hour and an Android emulator session that burned through a week's tokens.
• Anthropic's launch post priced Fable 5 at $10 per million input tokens and $50 per million output tokens, while Simon Willison's first impressions called it slow, expensive, and unusually willing to keep churning.
• The practical split in the HN discussion summary is simple: some users say Fable 5 catches bugs and high-level plan flaws earlier models missed, while others say benchmark gains matter less than task cost, latency, and how much autonomy the model takes on by default.

You can read Anthropic's announcement, Simon Willison's initial impressions and his browser-automation post, then drop into the main Hacker News thread, where the weirdest reports are not about benchmark charts at all. One user says Fable spun up an Android emulator for a database bug, another says it set up a full testing lab for a Windows lifecycle issue, and Anthropic's own product page says sensitive requests may silently route to Opus 4.8 instead of staying on Fable 5.

Willison's report is the clearest look yet at how Fable 5 behaves inside a real coding session. He handed Claude Code a screenshot of a scrollbar bug and, a few minutes later, watched it open Firefox, then Safari, then Chrome via Playwright while it tried to reproduce and measure the problem.

His write-up says Fable did more than click around. It inspected dependency code, edited templates to trigger keyboard shortcuts, started local servers, created a small measurement page with permissive CORS, wrote results to /tmp/diag.json, and read them back for analysis in the same run.

That is Christmas come early for coding agent nerds, but it also means the interesting unit is no longer just prompt quality. It is the full harness of tools, browsers, local files, and side effects the model is willing to enlist once you give it a vague goal.

The same thread that praised Fable 5's bug-finding also filled up with cost complaints. The June 10 HN delta cites one user who hit a monthly spend limit after about an hour on Ultracode, then burned through another $133 in 27 minutes.

The June 11 HN delta adds the most vivid example: a user asked about a straightforward database bug, Fable decided to launch an Android emulator, navigate by screenshots, and consumed what the commenter described as an entire week's tokens.

The hands-on reports cluster around three failure modes:

• Long, self-directed runs that keep exploring after a human might have stopped.
• Expensive tool choices, like browser automation or emulator setup, for bugs that might not need them.
• Weak visibility into whether the extra work produced a proportionate gain in task quality.

That last point is why one commenter in the HN discussion summary reduced model evaluation to three numbers: did it finish the task, how much did it cost, and how long did it take.

The counterweight is that several users say Fable 5 earns its keep on the right jobs. The HN discussion summary quotes one user saying the model "eviscerated" a technical proposal and found inconsistencies, ignored issues, and shaky math that earlier passes missed.

The earlier HN delta adds reports of stronger high-level reasoning, directional pivots after Opus and Codex had exhausted obvious fixes, and one debugging session where Fable built a full reproduction lab instead of trying a few shallow patches.

Willison's first impressions land in roughly the same place. He called Fable 5 a beast, noted a 1 million token context window and 128,000 token output cap, and said the real challenge was finding tasks it could not do.

Anthropic shipped two models on the same day: Claude Fable 5 and Claude Mythos 5. According to the launch summary, Mythos 5 is the same underlying model with some safeguards lifted, but it is restricted to Project Glasswing partners and other trusted-access programs rather than general API users.

For everyone else, Fable 5 is the public surface. Anthropic's product page says it is meant for long, complex knowledge and coding work, priced at $10 per million input tokens and $50 per million output tokens, with access through the Claude API, marketplaces, AWS, Google Cloud, and Microsoft Foundry.

Those list prices are not hidden. The surprise is how many early user reports describe the model acting like it has a much larger appetite for tool use than earlier Claude variants.

Anthropic buried one important operational detail in the launch materials and product page: Fable 5 does not always stay Fable 5. The launch post and product page say requests that trip cybersecurity, biology, chemistry, or distillation safeguards may route to Opus 4.8 instead, and the product page says API users need to configure that fallback behavior.

Willison's initial impressions flagged the same mechanism from the API side: Anthropic added new signals for guardrail hits and an option to fall back automatically when a request gets rejected. His later proactive session write-up says an invisible guardrail eventually downgraded his run to Opus, which then continued with the methods Fable had already set up.

That makes the early field reports harder to read than a normal model launch. Some surprising behavior is Fable 5 itself, some may be the surrounding agent harness, and some sessions can switch models midstream.