Claude Haiku 4.5

Anthropic identifies Claude 4.5 Haiku as a 'hybrid reasoning model' and a 'dense transformer,' but provides no specific architectural details such as layer counts, attention mechanisms (beyond general multi-head attention mentions), or specific modifications from previous generations. While the system card mentions 'context-aware' training to mitigate agentic laziness, the underlying pre-training methodology and specific architectural provenance remain proprietary and largely undisclosed.

Information regarding the training data is extremely limited. The system card vaguely refers to 'internet data' and 'AI feedback' (RLHF/RLAIF) without providing any breakdown of data sources, proportions (e.g., code vs. text), or specific filtering and cleaning methodologies. There is no public disclosure of the specific datasets used or sample data for verification, falling into the 'proprietary dataset' red flag category.

While a public tokenizer tool exists for the Claude 4.5 family (claudetokenizer.com) and the API provides token counting functionality, the technical documentation for the tokenizer itself—such as vocabulary size, training data alignment, and specific normalization techniques—is not comprehensively documented in a public technical paper. It is functional for users but opaque in its construction.

Anthropic does not disclose the parameter count for Claude 4.5 Haiku. While it is marketed as a 'small' or 'lightweight' model within the 4.5 family, no specific numbers (total or active) are provided. This lack of transparency makes it impossible to verify efficiency claims or compare density against competitors.

There is zero public information regarding the compute resources used to train Claude 4.5 Haiku. No GPU/TPU hours, hardware specifications, training duration, or carbon footprint data have been disclosed. Anthropic cites competitive reasons for withholding these metrics, which is a direct red flag under the scoring guidelines.

Anthropic provides evaluation results for standard benchmarks like SWE-bench Verified (73.3%) and internal 'computer use' tests. However, while they mention some methodology (e.g., averaging over 50 trials), the exact evaluation code, full prompt sets, and detailed reproduction instructions are not publicly available. Third-party verification is limited to high-level leaderboard positions rather than granular reproduction.

The model consistently identifies itself as Claude 4.5 Haiku and maintains a clear versioning identity (e.g., claude-haiku-4-5-20251001). It is transparent about its role as a 'sub-agent' or 'executor' within the broader Claude 4.5 ecosystem and accurately reflects its 200k context window and reasoning capabilities in its system prompts.

The model is governed by a strictly proprietary license. While the terms for API usage and commercial use are clearly stated in Anthropic's Commercial Terms of Service, it is not open source or open weights. The score is low because the 'license' is a service agreement rather than a model license that allows for inspection or derivative works.

As a closed-source API-only model, there is no documentation regarding the actual hardware requirements (VRAM, compute) to run the model. While Anthropic provides pricing and latency expectations, the underlying hardware footprint and quantization tradeoffs are entirely hidden from the public.

Anthropic uses dated versioning (e.g., 20251001) and maintains a basic changelog for its API and tools (like Claude Code). However, there is no public documentation of model drift over time or specific 'alignment tax' impacts. While version pinning is possible, the transparency regarding what changes between silent 'point' updates is minimal.