Anthropic releases Claude Fable 5.2 with 1M context window

Lead

Anthropic has released Claude Fable 5.2, a proprietary reasoning model that positions itself as a top-tier performer in intelligence while carrying a significant price and speed penalty. According to the benchmark data from Artificial Analysis, the model was released on November 9, 2026, and is currently available as the "Max, Default Fallback" variant. The model supports text and image inputs, outputs text, and features a 1M token context window, a specification that places it among the most capable models for long-context retrieval and reasoning tasks.

The most striking metric from the initial evaluation is the trade-off between capability and efficiency. Claude Fable 5.2 scores 65 on the Artificial Analysis Intelligence Index, a figure that sits well above the median of 40 for comparable models in its price tier. However, this intelligence comes at a steep cost: the model generates 275M tokens during the evaluation, which is nearly four times the median of 72.8M tokens. This extreme verbosity, combined with a latency of 44.6 minutes per task, suggests that Anthropic has prioritized depth of reasoning over speed and token economy.

The details

The pricing structure for Claude Fable 5.2 reflects its premium positioning. Input tokens are priced at $8.00 per million, which is four times the median input price of $2.00 for comparable models. Output tokens are priced at $40.00 per million, also four times the median of $10.00. When aggregated into a weighted average cost per Intelligence Index task, the model costs $26.54, a figure that is significantly higher than the typical cost for models in the same class.

Performance metrics indicate that the model is notably slow. At 44.6 minutes per task, it is more than three times slower than the median of 12.7 minutes for comparable models. The time to first answer token is 221.21 seconds, which is at the higher end of the spectrum for this tier. This latency is likely driven by the model's reasoning process, as the page notes that this is the "reasoning version" of the model, implying a non-reasoning variant may also exist. The high token count of 275M per task further supports the idea that the model is engaging in extensive internal reasoning before producing a final output.

The model's capabilities are evaluated across a wide range of benchmarks, including AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, and FrontierCode 2.0. These benchmarks cover agentic knowledge work, scientific research workflows, software engineering, and professional document reasoning. The inclusion of "Terminal-Bench-Science" and "FrontierCode" suggests that the model is particularly tuned for technical and scientific applications, where long-context reasoning and code generation are critical.

Context

Anthropic has historically positioned its Claude models as high-performance options for enterprise and developer use cases, often competing with OpenAI's GPT series and Google's Gemini models. The release of Claude Fable 5.2 follows a trend of increasing context windows and reasoning capabilities in frontier models. The 1M token context window is a significant upgrade, allowing the model to process entire codebases, large documents, or extensive conversation histories in a single pass. This is particularly relevant for Retrieval Augmented Generation (RAG) workflows, where the ability to reason over large amounts of retrieved data is essential.

The model's pricing and speed profile differentiate it from competitors that may offer similar intelligence at lower costs or higher speeds. For example, models with a median intelligence score of 40 and a median cost of $2.00 per million input tokens offer a more cost-effective option for high-volume, lower-complexity tasks. Claude Fable 5.2, by contrast, is designed for high-stakes, complex tasks where the cost of a wrong answer or insufficient reasoning outweighs the cost of the model itself. This positioning is consistent with Anthropic's focus on safety and reliability, as the model's verbosity and slow speed may be byproducts of a more conservative reasoning process that prioritizes accuracy over speed.

What it means for developers

For developers, Claude Fable 5.2 is a tool for specific, high-value use cases. It is not a general-purpose replacement for faster, cheaper models. Instead, it is best suited for tasks that require deep reasoning, such as complex code refactoring, scientific hypothesis generation, or legal document analysis. The 1M token context window makes it ideal for projects that involve large codebases or extensive documentation, where the ability to maintain context over long sequences is critical.

However, developers should be aware of the cost and latency implications. The $26.54 per task cost is significant, and the 44.6-minute latency may be prohibitive for real-time applications or high-volume batch processing. For these use cases, developers may prefer to use a non-reasoning variant of the model, if available, or to use a different model altogether. The model's verbosity also means that developers should carefully manage their token budgets and consider using prompt engineering techniques to reduce the number of tokens generated.

What to watch

The release of Claude Fable 5.2 raises several questions that will need to be answered in the coming weeks. First, will Anthropic release a non-reasoning variant of the model, and how will its performance and pricing compare? Second, how will the model perform in real-world applications, and will its high cost and latency be justified by its superior intelligence? Third, how will competitors respond to the release of a model with a 1M token context window, and will they follow suit with similar specifications? Finally, how will the model's performance on specific benchmarks, such as Terminal-Bench and FrontierCode, translate into practical improvements for developers and enterprises?