Anthropic issues opposite prompting advice for Fable 5 and Opus 4.8: what to change

Anthropic has started releasing a separate prompting guide for each of its models, and the advice in them sometimes flatly contradicts each other. Developers who spent years building collections of proven prompts will have to revisit them for each new flagship individually.

Why Fable 5 behaved strangely

A telling case from the community: a developer plugged in Claude Fable 5 right after release. The model objectively beats Opus 4.8 on benchmarks, yet in practice it performed worse than expected. In one run, the agent used 200 thousand tokens in an hour and a half and did not finish the task.

At first the developer blamed the model itself, until he opened Anthropic's official prompting guide for Fable 5 specifically. The problem turned out to be the prompts. They were written for Opus 4.8's logic and conflicted with the new flagship's architecture. After he reworked them following Anthropic's recommendations for Fable 5, everything worked.

The key takeaway: Anthropic now has two separate prompting documents, and their advice is in places diametrically opposed. Instructions that lived for years in skills and system prompts have become counterproductive on the new model.

Opposite advice for two flagships

Comparing Anthropic's official recommendations for Fable 5 and Opus 4.8, the differences are sharper than you would expect from two models from the same vendor:

• Subagents: Opus 4.8 must be explicitly nudged to delegate tasks; without instructions it prefers to solve everything in a single context. Fable 5 launches subagents aggressively by default and needs to be restrained with explicit limits on depth and number of calls.
• Explaining reasoning: The instruction "explain your reasoning," a proven technique for Opus, has started to make Fable 5 refuse to perform the task.
• Verbosity: Fable 5 produces long answers even when not asked. Opus 4.8 must be explicitly asked for detail.
• Recursion limits: For Fable 5 it is critical to set limits up front on the number of tool calls and the depth of nested chains, otherwise it is easy to end up with exactly that 200K tokens per session.
• Default format: Fable 5 tends toward structured output even without being told to; in a similar situation Opus prefers free-form text.

Anthropic attributes the divergence to different training regimes: Fable 5 is optimized for long agentic chains with maximum autonomy, while Opus 4.8 is optimized for high-quality answers within a single call.

A third path from OpenAI

OpenAI has its own prompting philosophy for o3 and GPT-5, which matches neither the recommendations for Fable 5 nor those for Opus 4.8. Where Anthropic focuses on managing agents' autonomy and their tendency to expand their own context, OpenAI emphasizes clear task decomposition, explicit constraints on output format and mandatory checkpoints in multistep scenarios.

The community's reaction is mixed. Some developers see differentiated guides as a sign of platform maturity: models are built differently, and it is more honest to say so than to offer universal advice that works at half strength everywhere. Others see an inconvenience: instead of one well-tuned prompt collection, you now have to maintain at least three parallel versions.

"Technically it's justified, the models are different.

But in practice it means several versions of the same prompts in production," from a discussion in a developer community.

What this means

The era of universal prompts that work equally well on any flagship is coming to an end. If you run several models in production, you will need either a separate configuration for each or a prompt router that picks the right variant automatically. For teams, this is a concrete engineering task that has to be done before the next flagship ships.