Tech • AI • Robotics • Game
Claude Fable 5.2: Complete Analysis
9.4/10TL;DR
Anthropic’s new Claude Fable 5.2 tops several coding and long-horizon AI rankings, but its strongest results often come with very high cost, long runtimes and uneven performance across tasks.
KEY POINTS
Two new models, one shared base
On 9 November 2026, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, described as variants of the same underlying model. Fable is broadly available, while Mythos is restricted to verified organizations. The model offers a 1 million-token context window, up to 128,000 output tokens, and adjustable reasoning effort from low to max.
Anthropic does not position Fable as a universal default
The company explicitly signals that Fable 5.2 is not the right choice for every task and recommends Opus 5 in some cases. That framing is unusual for a flagship release and sets expectations: the model is designed for specific hard problems rather than routine use. The trade-off is higher compute use as reasoning effort increases.
Big gains in software engineering benchmarks
On Frontier Code 2.0, which tests modifications to real open-source projects, Fable 5.2 scored 49.1%, up from 32.4% for Fable 5.1. That result used the XI reasoning setting. At high effort, it reached 45.3%, roughly matching Opus 5 while costing about $6.72 versus nearly $13 for a similar score.
A standout result on long-duration work
One of the most striking claims comes from the Remote Labor Index, which measures performance on longer professional tasks. Fable 5.2 posted 41.3%, compared with 21.3% for Opus 5. The result suggests a major improvement in endurance, coherence and completion over long-running agent-style assignments, a key challenge for practical AI systems.
Independent rankings also place Fable near the top
On Artificial Analysis, Fable 5.2 leads the platform’s intelligence index with a score around 65 at max effort, ahead of Gemini 4 and Opus 5. On Epoch AI, it also ranks first with an ECI score of 174, narrowly ahead of GPT-6 Astra at 169. Methodologies differ, so these numbers are not directly comparable, and the margins remain small enough to warrant caution.
High performance often comes with punishing cost and latency
On Artificial Analysis, average cost per task for Fable 5.2 at max effort is listed at $26.54, with an average runtime of 44.6 minutes. Official pricing is $8 per million input tokens and $40 per million output tokens, about twice the price of Opus 5 at equal token volumes. A faster mode can cut generation time, but doubles pricing again.
It does not dominate every benchmark
Fable 5.2 trails rivals in several specialized evaluations. Gemini leads on Automation Bench, Opus 5 is ahead on AR Briefcase, and Fable ranks only 27th of 84 on Tax Agent Bench and 20th of 37 in general law. The pattern is consistent: top-tier overall capability, but no universal supremacy.
Concrete web-generation tests show quality leadership, but at a steep premium
In blind-ranked comparisons building a 3D Central Park scene and recreating Shopify Editions Winter 2026, Fable 5.2 finished first, ahead of Opus 5, Fable 5.1, GPT-6 Astra and Gemini 4 Argon. But the price gap was enormous: the 3D scene reportedly cost $273 with Fable 5.2 versus $13 with Gemini 4 Argon, and the Shopify-style page cost $382 versus $8. The quality difference was visible, but so was the economic penalty.
Other creative and office tasks tell a mixed story
In a stained-glass HTML animation test, Fable 5.2 matched Opus 5 on completion checks, yet some evaluators preferred Opus 5 visually while Fable cost nearly twice as much. In presentation generation on Deck Bench, GPT-6 Astra and GPT-6 outperformed Fable 5.2 by a significant margin. For some front-end and slide workflows, rivals appear more practical.
More reasoning can reduce usefulness
A test by Simon Willison highlights a core weakness: overthinking. Asked to produce an SVG drawing of a pelican on a bicycle at five reasoning levels, Fable 5.2 delivered a convincing result at XI but failed at max, exhausting the full 128,000-token budget on reasoning without returning the final drawing. The failure illustrates how additional inference can become wasteful rather than helpful.
Long autonomous runs can drift into bureaucracy
In another experiment, agents using the model to build virtual worlds gradually created elaborate approval processes instead of producing output efficiently. After 12 days, the system had generated an 87,000-word rulebook that slowed its own work. The episode underlines a broader operational risk: advanced models can be highly capable on technical subtasks while still making poor organizational decisions without human oversight.
CONCLUSION
Claude Fable 5.2 appears to be one of the strongest AI models available for coding and long-horizon agent work, but its best performance is often slow, expensive and inconsistent across domains. The release strengthens Anthropic’s position at the top end of the market while reinforcing that model selection now depends as much on cost and task fit as on raw benchmark scores.
Ask a question