Fable 5.1 Prompting 攻略
💡Fable 5.1 can cut agent costs sharply, but changed tool behavior may break assumptions in your prompts.
⚡ 30-Second TL;DR
What Changed
Fable 5.1 introduces low, medium, high, xhigh, and max effort levels, with high as the API default.
Why It Matters
Agent developers may achieve lower inference costs by moving routine tasks to low or medium effort, but they must account for extra tool-call rounds and reduced user-visible progress messages. Existing system prompts should be simplified and re-evaluated rather than copied unchanged from Fable 5.
What To Do Next
Run your agent eval suite across Fable 5.1 at low, medium, and high effort, while tracking quality, tool-call rounds, latency, and cached-token cost.
Key Points
- •Fable 5.1 introduces low, medium, high, xhigh, and max effort levels, with high as the API default.
- •Low effort reportedly matches Fable 5 high effort on CursorBench at about one-third of the cost.
- •Prompt-cache read pricing falls from $1.00 to $0.25 per million tokens.
- •The model may use fewer parallel tool calls, provide fewer progress updates, and rewrite entire files for small code changes.
- •Anthropic recommends rerunning evaluations because effort levels are not behaviorally equivalent across model versions.
🧠 Deep Insight
Background and context from public sources — not the original article. 13 sources cited.
🔑 Enhanced Key Takeaways
- •Fable 5.1 was released alongside a specialized security-focused variant, Mythos 5.1, designed for high-stakes sectors like life sciences and cybersecurity.
- •The model demonstrates significant performance gains in long-horizon agent benchmarks, specifically doubling scores in Terminal-Bench-Science and nearly doubling them in AutomationBench compared to Fable 5.
- •Anthropic has implemented new restrictions on the Messages API regarding 'thinking blocks' to prevent unauthorized model distillation by limiting developer access to pre-thought system prompts.
- •The model exhibits a tendency toward 'mannered prose' (excessive metaphors), necessitating new 'anti-pattern' definitions in system prompts to enforce concise, direct communication.
- •The release is strategically positioned as a final major model update prior to Anthropic's anticipated IPO, emphasizing commercial viability through aggressive cost-reduction in long-context processing.
📊 Competitor Analysis▸ Show
| Feature | Fable 5.1 | OpenAI o3-series | Google Gemini 1.5 Pro |
|---|---|---|---|
| Agentic Focus | High (Multi-step/Long-horizon) | High (Reasoning-heavy) | Medium (Multimodal/Context) |
| Cache Read Price | $0.25/1M tokens | Varies (Higher) | $0.375/1M tokens |
| Effort Control | 5-level granular control | Limited (Reasoning budget) | None (Fixed latency) |
🛠️ Technical Deep Dive
- Effort levels (low to max) dynamically adjust the model's internal compute allocation per token, effectively acting as a variable inference budget.
- The reduction in parallel tool calls is an architectural trade-off to improve sequential reasoning stability in complex agentic workflows.
- The API now enforces strict immutability on system prompts preceding 'thinking blocks' to mitigate distillation risks.
- Cache read pricing optimization is achieved through a refined KV-cache management layer that reduces memory overhead for long-context retrieval.
🔮 Future ImplicationsAI analysis grounded in cited sources
⏳ Timeline
📎 Sources (13)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.