Claude Fable 5, Explained for Skill Builders: What Changes When You Build on Anthropic's New Flagship

Anthropic announced Claude Fable 5 (claude-fable-5) on June 9, 2026, and called it its most capable widely released model — "built for the most demanding reasoning and long-horizon agentic work." After a brief pause in late June, it was redeployed globally around July 1. As of July 2026 it runs on the Claude API, in Claude Code, and across the major cloud platforms.

If you build skills, agents, or automation loops, "most capable model" is not the interesting part. The interesting part is that Fable 5 quietly changes the contract you code against. Thinking is no longer a knob you turn on — it's always on. The token budget you used to set is gone, replaced by an effort dial. There's a new stop_reason your code has to handle before it reads a single character of output. And there's a data-retention floor that will break some enterprise deployments outright.

None of that shows up if you only swap the model ID and eyeball the results. It shows up three weeks later when a skill that worked fine starts refusing, or a subagent silently times out, or a compliance review flags the model. This piece is the map. Each section below is a headline shift, and each links to a deeper article in the rest of this series where we work through the fix.

Key Takeaways

• Thinking is always on and adaptive. You can't disable it and you can't set a token budget — you steer depth with an effort level instead.
• Effort is the new dial: low, medium, high, xhigh, max. Start at high. Route subagents and routine work down, long-horizon work up.
• 1M-token context, up to 128K output tokens per request — the same tokenizer as Opus 4.8, so migration doesn't inflate your counts.
• Pricing is $10 / $50 per million input/output tokens — roughly 2x Opus 4.8. That changes the routing math for high-volume skills.
• Refusals arrive as HTTP 200 with stop_reason: "refusal". Check it before reading content, and wire up a fallback to Opus 4.8.
• Fable 5 requires 30-day data retention — a zero-data-retention org gets a 400 on every request.

Shift 1: Thinking is always on — the budget knob is gone

On older models you could turn extended thinking on or off and hand it a budget_tokens. On Fable 5, thinking is always on and adaptive, and both of those controls are gone. Send thinking: {"type":"disabled"} and you get a 400. Send {"type":"enabled","budget_tokens":N} and you also get a 400. The correct move is to omit the thinking parameter entirely and control depth through output_config.effort instead.

There's a second wrinkle that matters for any skill that inspects model output. The raw chain of thought is never returned. Responses carry thinking blocks, but their display field is either "summarized" (a readable summary) or "omitted" (the default — empty text). If you're continuing a conversation on the same model, pass those thinking blocks back unchanged; other models will drop them. This single behavior underpins two of the traps below.

Shift 2: Effort replaces the budget — and it's a routing decision

Because you can't set a token budget, effort is now how you trade capability for cost and latency. There are five levels — low, medium, high (default), xhigh, and max — and they live inside output_config, not at the top level.

Anthropic's guidance is concrete: start at high; reach for xhigh on capability-sensitive or long-horizon work (think runs longer than 30 minutes); drop to low or medium for routine tasks and subagents. The line worth memorizing is that "lower effort on Fable 5 still performs well and often exceeds prior models' xhigh." In other words, medium here is not a downgrade — it's frequently better than your old ceiling, at a fraction of the cost.

For skill builders, that turns effort into an architecture decision. A cheap classifier subagent should not run at max. A frontier refactor across a large repo should not run at low. We build the full routing table — including where Claude Code's "ultracode" mode sits — in Effort Is the New Dial.

Shift 3: Refusals are a stop reason you have to handle

This is the change most likely to cause a 2am page. Fable 5's safety classifiers can decline a request, and when they do you get HTTP 200 with stop_reason: "refusal" and a stop_details object naming the category — "cyber", "bio", "reasoning_extraction", or null. It's a 200, not an error, so naive code sails straight past it and reads empty or partial content as if it were a real answer.

The rule is simple: always check stop_reason before you read content. Refusals trigger in under ~5% of sessions on average (Anthropic-reported), and they target offensive cybersecurity and biology work — but benign adjacent tasks can trip a false positive, and a refusal can fire before any output (empty content, not billed) or mid-stream (partial output, billed — discard it).

The good news is that fallbacks are opt-in and built for exactly this. The server-side fallbacks parameter re-runs the same request on a fallback model — Opus 4.8 at launch — in a single call. We walk through both the server-side and client-side middleware paths, with the handling checklist, in Building Fallback-Aware Agents.

Shift 4: Your old skill prompts are probably too prescriptive

Here's the counterintuitive one. Anthropic's own prompting guidance says that skills and prompts written for older models are "often too prescriptive… and can degrade output quality" on Fable 5. The step-by-step scripts you carefully tuned for a weaker model now get in the way of a stronger one.

The fix is to audit those prompts and strip the over-specified procedure, stating the goal plus constraints and letting the model plan. There's a sharp special case: any instruction that tells the model to "explain your reasoning step by step," "show your work," or "output your chain of thought" can now be refused outright with category reasoning_extraction, and it elevates fallbacks. Two articles cover this — the general audit in Your Skills Are Too Prescriptive, and the specific trap in Don't Tell Claude to Show Its Work.

Shift 5: Turns can run for minutes — sometimes hours

Fable 5 takes minutes-long turns by default. Hard requests at higher effort can run for many minutes; autonomous runs can extend for hours. Anthropic is explicit that you should adjust client timeouts, use streaming and progress UIs, and move toward async or scheduled check-ins rather than blocking a request thread and hoping.

It also dispatches parallel subagents more readily, and asynchronous orchestrator-to-subagent messaging — long-lived subagents that keep their context — outperforms spawn-and-block. And it performs strongly with a markdown memory system: give it a file, tell it the format, and tell it to consult that file in future sessions. If you build multi-step agents or workflows, these three behaviors reshape how you structure a run.

Shift 6: Retention is now a hard gate

Fable 5 requires 30-day data retention and is not available under zero-data-retention (ZDR). A ZDR org gets a 400 invalid_request_error on every request — there's no partial mode. This has already produced real friction: at least one large enterprise removed Fable 5 from an internal tool selection because it conflicted with a zero-retention standard.

If you ship skills to teams, this is a deployment blocker you need to surface early, not a footnote. The workaround pattern — routing retention-sensitive traffic to a Covered-Model-free path while keeping Fable 5 for the rest — is the subject of a dedicated article later in the series.

The six shifts at a glance

Where this goes next

Fable 5 rewards builders who treat it as a new contract rather than a faster version of the old one. The through-line across every shift above is the same: give the model a goal, a boundary, and a cheaper-or-stronger effort setting, then get out of its way and handle the two new failure modes — a refusal and a timeout — explicitly in code.

Start with the two that will bite first: refactor your most prescriptive skill using the audit guide, and make one production agent fallback-aware. Then tune your effort routing. You can browse the whole run of this series at the Claude Fable 5 hub, and find installable skills, agents, and loops to build on across /browse and /loops.