Claude Sonnet 4.5

Anthropic identifies Claude 4.5 Sonnet as a 'dense transformer-based' model utilizing a 'hybrid reasoning system' with standard and extended thinking modes. However, no technical paper or detailed architectural documentation has been released. Specifics regarding layer counts, attention mechanisms (beyond a mention of multi-head attention), or the exact nature of the hybrid reasoning implementation remain proprietary and undisclosed.

The model's training data is described vaguely as a 'proprietary mix' of public internet data (up to July 2025), non-public third-party data, and user-provided data. While the System Card mentions general cleaning methods like deduplication, it provides no specific breakdown of dataset proportions (e.g., code vs. web), naming of specific sources, or verifiable details on the filtering criteria used.

While a 'Claude Tokenizer' is publicly accessible via web tools and APIs for token counting, official technical documentation detailing the vocabulary size, specific tokenization algorithm (e.g., BPE), or training data alignment for the 4.5 series is absent. Users can verify token counts through the API, but the underlying technical specifications are not fully transparent.

Anthropic does not disclose the parameter count for Claude 4.5 Sonnet. While the model is described as 'dense' to distinguish it from sparse or MoE architectures, there is no verifiable information regarding total or active parameters, nor any architectural breakdown of parameter distribution across model components.

No information is provided regarding the compute resources used to train the model. There are no public disclosures of GPU/TPU hours, hardware specifications, training duration, or the total carbon footprint associated with the training phase. Environmental data is limited to third-party inference estimates rather than official training reports.

Anthropic provides scores for several public benchmarks (SWE-bench Verified, OSWorld, GPQA) and some details on evaluation settings (e.g., 100 max steps for OSWorld). However, the full evaluation code, exact prompts, and few-shot examples required for independent reproduction are not publicly available, and some results rely on 'internal benchmarks' or specific 'prompt addendums' that are not fully disclosed.

The model demonstrates high identity consistency, correctly identifying itself as Claude 4.5 Sonnet and maintaining awareness of its versioning. It provides clear information about its capabilities, such as the extended thinking mode and context window limits, and does not exhibit confusion with competitor models in official documentation or API responses.

The model is governed by a clear but strictly proprietary license. Commercial terms are defined for API and enterprise users, while consumer terms apply to Pro/Max users. While the terms are accessible, the lack of an open-source or open-weights option and the presence of restrictive usage caps on 'flat-rate' plans create some complexity for users regarding derivative works and commercial scaling.

As a closed-source API-based model, there is no documentation regarding the hardware required to run the model locally (VRAM, quantization tradeoffs, etc.). Guidance is limited to API-side constraints like context window limits (200k/1M) and output token maximums (64k), which do not provide transparency into the model's actual computational requirements.

Anthropic uses specific model strings (e.g., claude-sonnet-4-5-20250929) and maintains a changelog for associated tools like Claude Code. However, the model weights themselves are subject to silent updates and 'behavioral improvements' (e.g., alignment tuning) that are not always accompanied by new version numbers, making it difficult for developers to track or mitigate performance drift over time.