Claude Fable 5.2 scores 65 on Artificial Analysis, highest so far
TL;DR Highlight
Fable 5.2 (max) tops the Intelligence Index at 65 but costs $26.54 and 46.4 minutes per task
Who Should Read
A backend developer weighing whether to switch to a new LLM based on cost and latency
Core Mechanics
- Tops the Intelligence Index at 65
- It scores 65 on Artificial Analysis Intelligence Index v5.0.1, ranking first of 734 models.
- The median for comparable models is 40.
- $26.54 per task, last in cost
- Running one Intelligence Index task costs $26.54 on average.
- It ranks 73rd of 73 on cost.
- Priced at 4x the median
- Input is $8.00 per 1M tokens versus a $2.00 median.
- Output is $40.00 per 1M tokens versus a $10.00 median.
- 46.4 minutes per task, 215.95s to first answer token
- The median for comparable models is 12.6 minutes per task.
- The median time to first answer token is 11.84s.
- Extremely verbose
- It generated 275M output tokens during evaluation versus a 72.8M median.
- The cache discount is 98%.
- 1M context, text and image input
- It is a reasoning model with text-only output.
- It was released November 9, 2026 and is available from 6 API providers.
Community Reactions
- Max effort is not how people really use it, some say
- One commenter noted the site's charts default to max effort, which shows model limits but is rarely used in practice.
- They said Fable 5.2 medium looks attractive versus Opus 5.5 high/xhigh.
- Token explosion doubles cost, a commenter points out
- A comment says 5.1 used 118k tokens per task (46k answer, 71k reasoning) while 5.2 uses 274k (81k answer, 192k reasoning).
- They say cost per task is double 5.1 even after the price cut.
- Real-world use disappointed some
- One user said six turns and about 10 minutes of thinking produced shallow, circular logic and they would go back to 5.5.
- Another said it lied several times during feature design analysis and Opus 5.5 did better.
- Doubts about the benchmark itself
- One commenter said Artificial Analysis is paid by AI labs and doesn't publish its full method, so the scores are meaningless to them.
- Another said long execution time suggests brute-forcing with a huge compute budget.
- Worth it as planner or reviewer, some argue
- One commenter sees Fable as a planner or plan reviewer, where you want the best model.
- Another said the high setting scores 62 at 3x lower cost.
How to Apply
- Build a cost and latency comparison table
- Today, pick 10 representative tasks and record token counts and latency for your current model and Fable 5.2 side by side.
- Compare cost per task and time before looking at intelligence scores.
- Test medium and high instead of max effort
- Run the same prompts at max, high, and medium and measure quality and cost differences.
- If the gap is small, fix the lower setting as your default.
- Use Fable only for planning and review
- Route only the planning and result-review steps of your agent pipeline to Fable.
- Leave execution steps to a cheaper model to cut call volume.
Try It
# Prompt to paste into Claude or ChatGPT (cost calculator using the article's numbers)
Below are model prices and our task token usage.
- Fable 5.2: input $8.00/1M, output $40.00/1M
- Comparison model: input {{input price}}/1M, output {{output price}}/1M
- Average per task: input {{input tokens}}, output {{output tokens}} (including reasoning tokens)
- Monthly calls: {{monthly calls}}
Calculate the monthly cost of both models in a table, and also show how cost changes if output tokens double.
Next step: add expected output tokens at a lower effort setting to the same table and recalculate.Terminology
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