Claude Fable 5.2 and Mythos 5.2
TL;DR Highlight
Fable 5.2 generates up to 30% faster than 5.1 at a lower price
Who Should Read
A backend developer running coding agents on Opus-class models who wonders whether moving to a higher-tier model is worth the cost
Core Mechanics
- Cheaper and faster
- Fable 5.2 and Mythos 5.2 generate output up to 30% faster than 5.1.
- Price is $8 per million input tokens and $40 per million output tokens.
- Optional fast mode speeds generation up to 2.5x at $16 input and $80 output.
- Fable and Mythos are the same model
- Fable ships with cyber and biology safeguards so it can be open to all users.
- Mythos is for organizations that apply through the Life Sciences and Cyber Verification Programs.
- Fable 5.2's safeguards block most cybersecurity tasks.
- Lower effort matches Opus 5.5
- On FrontierCode 2.0, Fable 5.2 at High effort scores 45.3%, about level with Opus 5.5 at Max effort (45.5%).
- Cost per task is roughly half.
- At Xhigh effort it hits 49.1%, the highest of any released model Anthropic tested.
- Cost savings vary by benchmark
- On WANDR, Medium effort matches Opus 5.5's top score at about 40% of the cost per task.
- On Humanity's Last Exam, Low effort matches Opus 5.5's top score at under one-tenth the cost.
- On Chartography, Medium effort scores 66.4%, ahead of Opus 5.5 at Max (64.4%), at under one-tenth the cost.
- Start with Opus 5.5 for everyday work
- Anthropic recommends starting with Opus 5.5 and moving up to Fable or Mythos, or down to Sonnet 5.5 or Haiku 5.5, as needed.
- Fable 5.2 can act as an Advisor that smaller models consult on hard problems.
- It can also be an Orchestrator that delegates simpler tasks to smaller models.
- Large long-running agent examples
- Over 1,000 subagents worked eight days to build a bootable prototype operating system.
- An agent team migrated a production service from Python to Rust, and it now starts about 26x faster.
- 1,000 agents formalized Fermat's Last Theorem in 30 hours, while Opus 5.5 with the same setup finished 26% in 18 hours.
Community Reactions
- Experience: no difference on ordinary coding
- One commenter says they can hardly tell Fable 5.2 from Opus 5.5 on everyday coding tasks.
- They say you must push prompt ambition to see a gap, and the slowness and max-output-token overflows from 5.1 were gone in 5.2.
- Counterpoint: it just spends more tokens
- A commenter's SVG benchmark shows at High effort Fable 5.2 wrote 76,158 tokens ($3.05) versus 19,767 ($0.40) for Opus 5.5.
- At Max, Fable 5.2 reportedly used its whole 128K output limit on thinking when streamed and returned nothing.
- They still recommend sticking with Opus 5.5.
- Opinion: the price is hard to justify
- A commenter cites Artificial Analysis: GPT-6.1 Sol costs $1.27 per task versus $26.54 for Fable 5.2, about 21x.
- They argue Fable is not 21 times better.
- Experience: big speedup on legacy code
- A commenter says Fable 5.2 sped up computational materials science work 36x on a 600k-line Fortran codebase in two days.
- They say they open-sourced the results.
How to Apply
- Compare Opus 5.5 and Fable 5.2 on three real tasks
- Run three real tasks on Opus 5.5 and Fable 5.2 at Medium effort with identical prompts.
- Log success, tokens used, and cost in a table.
- Default to Opus 5.5, escalate to Fable when stuck
- Fix your pipeline's default model to Opus 5.5 or Sonnet 5.5.
- Branch to call Fable 5.2 as an Advisor when a step fails twice.
- Record cost per task while varying effort
- Run the same task at Low, Medium, and High and log score and cost per task.
- Pick the effort level where score per dollar flattens as your default.
- Check whether your Fable tasks involve cybersecurity
- Fable 5.2 blocks most security tasks, so consider applying to the Cyber Verification Program if you need them.
Try It
# Paste into Claude Code or Claude (simplified prompt version of the Advisor/Orchestrator setup from the post)
You are an Orchestrator. Split the task below into steps and delegate simple steps to smaller models (subagents).
Only escalate a step to the stronger model as an Advisor if it is stuck or has failed twice.
Task: {{task description}}
Instructions:
1. Show the step list first.
2. For each step, give the model you will use and a one-line reason.
3. When done, summarize tokens used and failure counts per step in a table.
Next to try: rerun the same task with different models and effort levels and compare the tablesTerminology
Related Papers
Gemini 4 Neon
Gemini 4 Neon cuts cost per task 32% versus Carbon, at $3 input and $15 output per million tokens
AI is killing All About Berlin
After Google AI Overviews, a Berlin guide site lost 70-75% of its traffic.
Not Everything Needs an LLM
PostHog used small decision models to skip 98.7% of expensive LLM calls
GPT‑6.2 Astra and Sol
GPT-6.2 Astra and Sol cut costs sharply, but over-engineering and early-stop reports are mixed
Show HN: TrnRT – An open-source LLM inference engine for AWS Trainium
One trn2.48xlarge serves GLM-5.3-Flash through an OpenAI-compatible API using NKI megakernels.
Know When to Zoom: How Small Is Too Small for Vision-Language Models?
Splitting large photos into tiles raised Claude and Gemini accuracy by 11.1–14.5 points
Understanding Positional Embeddings in Modern LLMs
A walkthrough of how LLMs encode token position, from GPT-2 to RoPE
Show HN: DeepSeek V3 in ~200 lines of pure Jax
DeepSeek V3's core design in 190 lines of JAX, trained on a MacBook Air in 2.6 hours