According to Ethan Mollick on X, Meta's Muse Spark Thinking underperforms compared with the current Big Three models, exhibiting odd tone and occasional factual looseness, and falls short on a neo-gothic shader coding task in twigl compared with leading models (source: Ethan Mollick on X, Apr 9, 2026). As reported by Mollick, earlier benchmarks he shared showed GPT 5.2 Pro generating a single-shot shader for an infinite neo-gothic city partially submerged in a stormy ocean, suggesting stronger code synthesis and visual reasoning than Muse Spark Thinking on the same prompt (source: Ethan Mollick on X). According to Mollick, these results indicate practical implications for developers: teams needing reliable shader generation, graphics prototyping, or complex code synthesis may achieve higher productivity with top-tier models while monitoring Muse Spark Thinking for improvements in factuality and stylistic control (source: Ethan Mollick on X).

According to @emollick referencing @JustinWaugh’s release, the Pencil Puzzle Bench tests 51 LLMs on 62k unique pencil puzzles across 94 types with an evaluation set of 300 puzzles over 20 types, showing modern reasoner models dramatically outperform early non‑reasoner LLMs. As reported by @JustinWaugh, the best score is 56% by GPT 5.2 at xhigh settings, and roughly half the puzzles remain unsolved, highlighting significant headroom for tool‑supported reasoning and verification‑driven training. According to the X thread by @JustinWaugh, the benchmark emphasizes multi‑step logical reasoning with step‑verifiable solutions, providing a clearer signal for chain‑of‑thought robustness and planning. As noted by @emollick, performance gains appear logistic due to a 100‑point ceiling, suggesting maturing returns and the need for targeted data curricula, planner‑solver architectures, and self‑verification loops for enterprise use cases like operations optimization, scheduling, and compliance workflows.

According to Ethan Mollick on X, DnD puzzle creation remains an unsolved benchmark for state-of-the-art models, with Gemini 3.1 Deep Think producing an engaging scenario rather than a true puzzle, while GPT 5.2 Pro and Opus 4.6 overcomplicate designs and generate unworkable mechanics (as reported by Ethan Mollick). According to Mollick, the task—creating a compelling, choice-rich, solvable DnD puzzle—demands long-horizon planning, constraint satisfaction, and playability testing that current models fail to reliably integrate, highlighting a gap in model-based planning and iterative validation for game design workflows (according to Ethan Mollick). For AI product teams, this underscores opportunities in tool-augmented reasoning, domain-specific validators, and human-in-the-loop puzzle editors to operationalize content quality and ensure puzzle solvability (as reported by Ethan Mollick).

According to Abacus.AI (@abacusai), the Abacus AI Desktop platform now provides access to leading coding AI models, including Sonnet, Opus, GPT 5.2, and Gemini. The default model, Sonnet 4.5, is highlighted for delivering an optimal balance between cost and performance. The company reports achieving the top position on terminal bench #1 and aims to lead terminal bench #2 soon. This multi-model integration offers businesses and developers a unified interface for leveraging the latest coding AI technologies, streamlining workflow automation and enabling more efficient software development processes. Such a platform positions Abacus AI Desktop as a competitive solution in the rapidly evolving AI coding assistant market, with direct business benefits for enterprises seeking scalable, high-performance code generation tools (source: @abacusai).