In this December 2025 AI coding roundup, the presenter reviews various AI coding agents and models, focusing on their performance, tool-calling methods, and practical usability. They emphasize the complexity of tool calling, distinguishing between native tool calling—where the model directly calls functions with structured parameters—and virtual tool calling, which embeds tool call data within the prompt response for parsing by the harness. The presenter notes a trend between agents using many specific tools versus those relying on fewer, more generic tools that execute terminal commands or code, highlighting the trade-offs in context pollution and execution accuracy.
The evaluation centers on instruction-following ability, measured through unit tests designed to score 100% if instructions are perfectly followed. Models like Gemini 3.0 Pro show variability depending on the harness used, sometimes failing to call tools consistently, which impacts their scores. GPT 5.1 exhibits high variance, occasionally performing well but often struggling with improper tool calls or execution errors. Claude 4.5 and Opus 4.5 stand out for their consistent and high performance, with Opus 4.5 noted for token efficiency despite a higher price point. The presenter also discusses the challenges with some agents like Augment Code and GitHub Copilot, which had significant issues completing tests with certain models.
Among the agents tested, ZAI with Claude 4.5 leads overall, followed closely by Root Code and Crush, with Cursor and Warp also performing well. Open Code impresses particularly when paired with GPT 5.1, showing a surprising performance boost. However, many agents struggle with GPT 5.1 except for a few top performers. The presenter highlights that the Codeex Max variant of GPT 5.1 is notably better than the base model, especially within the Codeex CLI environment. Despite the hype around GPT 5.1, it does not outperform the Anthropic models in this evaluation, which maintain a significant lead.
The presenter shares personal preferences for December 2025, favoring Open Code for its open-source nature and solid performance across multiple models, including Sonic 4.5, Opus, and Gemini 3 Pro. Claw Code is praised for its fixed $100/month pricing and usage limits, making it a cost-effective choice despite not being open source. Cursor is highlighted as an excellent but more expensive option at $200/month, offering advanced workflow features like multi-creation plans and seamless agent switching within the IDE. The presenter appreciates Cursor’s improvements despite some frustrations with its credit system and past secrecy around model tuning.
In closing, the presenter reflects on the evolving landscape of AI coding agents, noting the convergence in capabilities and the importance of harness engineering in model performance. They mention their recent experience integrating AI coding tools in a large engineering company, which has deepened their understanding of practical deployment challenges. The presenter plans to continue producing detailed evaluations, albeit at a reduced frequency due to increased professional commitments. They invite viewers to share feedback and express enthusiasm for advancing testing methodologies to better measure AI coding agent capabilities in the future.
