Anthropic's Fable 5.2 doubles hardware design scores

A leap in engineering capability

Anthropic recently released Fable 5.2, its newest flagship AI model designed for complex, open-ended projects. The company reported that the model achieved a score of 83.5 on a specific hardware-design benchmark. This represents a massive jump from the 44.4 score recorded by its predecessor, Fable 5.1.

The model was made generally available on November 9. It is positioned as a high-effort tool for tasks that require more reasoning than standard chat applications. While Anthropic still recommends Opus 5.5 for everyday coding, Fable 5.2 is intended for the most difficult technical challenges.

The performance gains are particularly visible in hardware engineering. The model can now handle designs for electronics, enclosures, and firmware. These capabilities allow it to move beyond simple text generation into the realm of physical product development.

From prompts to prototypes

To test these capabilities, Anthropic tasked the model with designing three different physical items. These included a voice-assistant hub, a USB speakerphone, and a handheld virtual pet. The model worked from short descriptions to generate the necessary electronics and firmware for each device.

The company collaborated with Diode, a firm that manufactures circuit boards, to turn these designs into reality. According to Anthropic, each circuit board was ready for manufacturing in less than one day. Diode's engineers assembled the boards and ran the firmware under the direct guidance of the AI model.

The resulting prototypes worked without needing any electrical modifications. However, these devices are still in the early stages of development. They still require antenna tuning, thermal testing, and official certification before they can be used in a commercial setting.

Benchmark and cost efficiency

The new model also offers a more affordable price point for heavy users. It is priced 20% lower per token than Fable 5.1, specifically $8 per million input tokens and $40 per million output tokens. These costs are designed to make high-effort reasoning more accessible to developers and researchers.

Independent testing by Artificial Analysis shows that Fable 5.2 leads the Intelligence Index at its highest settings. At the Max effort setting, it scored 65, beating Google’s Gemini 4 Carbon, which scored 60. The model also outperformed rivals in the FrontierCode 2.0 benchmark, which measures if human maintainers would accept AI-generated code changes.

In robot control simulations, the improvements were even more dramatic. On the RoboDojo Sim, which tests how a two-arm robot moves based on camera images, Fable 5.2 scored 64.9%. This is more than double the 27.6% score achieved by Fable 5.1 and significantly higher than the 29.8% scored by GPT-6 Astra.

Safety and specialized versions

Anthropic has implemented specific safeguards to manage the model's capabilities. Fable 5.2 is programmed to block most cybersecurity tasks and a large portion of research biology. These restrictions are intended to prevent the model from being used for harmful activities in sensitive fields.

For organizations that require access to these capabilities for legitimate purposes, Anthropic offers a different version. Mythos 5.2 uses the same underlying model as Fable 5.2 but is available through specific verification programs. These programs are tailored for cybersecurity and life sciences research.

The release of Fable 5.2 signals a shift toward specialized AI agents. By allowing the model to delegate simpler tasks to smaller models, Anthropic aims to make complex engineering workflows more efficient. This multi-model approach helps users complete large projects while managing costs and computational resources.

Comparative performance metrics

The model demonstrates consistent performance across different intensity levels. At the High and Xhigh settings, Fable 5.2 scored 62, with costs of $9.08 and $14.83 per task, respectively. These results show that the model maintains high intelligence even when not running at its maximum effort capacity.

In terms of 3D modeling, the BenchCAD 2.1 public benchmark evaluates how models rebuild assemblies and circuit boards from images. While Fable 5.2 shows strong hardware capabilities, it is noted that Opus 5.5 scored 65.3 on this specific test. This helps users understand which model is best suited for different types of spatial and engineering reconstruction tasks.

Context and industry trends

Anthropic's progress in physical design follows a trend set by other major labs. This work follows a recent account from OpenAI regarding GPT-6 Astra, which also designed a talking Codex pet. However, that project required several rounds of manufacturing reviews and iterations, whereas Fable 5.2 produced ready-for-manufacturing boards in under a day.

The broader landscape of high-end models continues to evolve with specialized releases. For example, Google announced Carbon on October 28, which was initially made available only to selected cybersecurity defenders. These developments highlight a growing trend of releasing powerful models with specific access controls for sensitive industries like security and life sciences.