Anthropic gave opposite advice for Fable 5 and Opus 4.8: what to change in prompts

Anthropic has started releasing separate prompting guides for each of its models — and the advice in them sometimes directly contradicts one another. Developers who have spent years accumulating collections of tested prompts will need to review them separately for each new flagship.

Why Fable 5 Behaved Strangely

A telling case from the community: a developer connected Claude Fable 5 immediately after its release. The model objectively outperforms Opus 4.8 on benchmarks, yet in practice it worked worse than expected. In one run, an agent consumed 200,000 tokens over an hour and a half without completing the task. At first, the developer blamed the model itself, until he opened Anthropic's official prompting guide specifically for Fable 5. The problem turned out to be in the prompts. They were written for Opus 4.8 logic and conflicted with the architecture of the new flagship. After reworking them according to Anthropic's recommendations for Fable 5, everything worked.

Key takeaway: Anthropic now has two separate documents on prompting — and the advice in them is sometimes diametrically opposed. Those instructions that had lived in skills and system prompts for years became counterproductive on the new model.

Contradictory Advice for Two Flagships

When comparing Anthropic's official recommendations for Fable 5 and Opus 4.8, the differences prove sharper than you might expect from two models of the same vendor:

• Subagents: Opus 4.8 needs to be explicitly pushed toward delegating tasks — without instructions it prefers to solve everything in a single context. Fable 5 launches subagents aggressively by default, and it needs to be restrained with explicit limits on depth and number of calls.

• Reasoning explanation: The instruction "explain your reasoning" — a proven technique for Opus — began causing task refusals on Fable 5.

• Verbosity: Fable 5 produces long answers even without being asked. Opus 4.8 needs to be explicitly asked for detailed responses.

• Recursion limits: For Fable 5, it's critical to specify in advance constraints on the number of tool calls and depth of nested chains — otherwise you easily end up with exactly those 200K tokens per session.

• Default format: Fable 5 tends toward structured output even without explicit instruction; Opus in a similar situation prefers free-form text.

Anthropic explains the discrepancy through different training regimes: Fable 5 is optimized for extended agent chains with maximum autonomy, Opus 4.8 — for high-quality responses within a single call.

A Third Path from OpenAI

OpenAI has its own prompting philosophy for o3 and GPT-5, which aligns with neither the recommendations for Fable 5 nor those for Opus 4.8. While Anthropic focuses on managing agent autonomy and their tendency to self-expand context, OpenAI emphasizes clear task decomposition, explicit constraints on output format, and mandatory checkpoints in multi-step scenarios.

Community reaction is mixed. Some developers see differentiated guides as a sign of platform maturity: models are constructed differently, and it's more honest to acknowledge this explicitly than to give universal advice that works at half strength for everyone. Others see it as an inconvenience — instead of one verified collection of prompts, you now need to maintain at least three parallel versions.

"Technically this is justified — the models are different.

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

What This Means

The era of universal prompts that work equally well on any flagship is coming to an end. If you're working with multiple models in production, you'll need either a separate configuration for each one or a prompt router that automatically selects the right variant. For teams, this is a specific engineering task that will need to be resolved before the next flagship arrives.