Overview
This is an English translation of the Easy AI Daily digest for November 24, 2025, originally published in Chinese on zhichai.net by the Easy AI teaching project. The edition centers on a wave of frontier model releases and ecosystem updates.
Model Updates & Performance
- Anthropic released Claude Opus 4.5, positioned as its best model for coding and agents. Pricing is 3x lower than Opus 4.1 ($5/$25 per million tokens), with new API features including effort control. It scored 80.9% on SWE-Bench Verified, setting a new SOTA. (official announcement)
- Google released Gemini 3 Pro, scoring 76.2% on SWE-Bench Verified, though users report hallucination issues and frequent instruction-ignoring.
- OpenAI released GPT-5.1-Codex-Max, scoring 77.9% on SWE-Bench Verified, briefly holding SOTA for coding.
- Google released Gemini 3 Image, topping the Artificial Analysis image benchmark, supporting 14 input images with improved photorealism and editing.
- Claude Opus 4.5 set multiple records: SWE-bench Pro 52%, BrowseComp-Plus 85.3%, ARC-AGI-1 80%, ARC-AGI-2 37.64%. (system card)
- Qwen3-VL-32B surpassed Kimi K1.5 on MathVision by 24.8 points, showing stronger visual reasoning.
- Claude Opus 4.5 adds effort control, context compaction, and advanced tool use APIs, available on Bedrock, Vertex, and other cloud platforms.
- Windsurf 1.12.35 (stable) / 1.12.152 (preview) supports Claude Opus 4.5, offered for a limited time at Sonnet pricing (2x credits). (changelog)
- Weaviate v1.32 enables 8-bit Rotational Quantization by default, claiming 98–99% accuracy retention with lower latency and better write performance.
- Weights & Biases launched Serverless LoRA inference: upload adapters and switch them dynamically at inference time with no cold start.
- Zyphra, AMD, and IBM released ZAYA1-base, an AMD-native MoE model (8.3B total params, 760M active) that beats Llama-3-8B on math and coding tasks.
- DiRL framework combines SFT with a diffusion-native RL algorithm (DiPO) to optimize diffusion language models; an 8B model reaches 83% on MATH500.
- Sakana AI's Continuous Thought Machines (CTM), a NeurIPS spotlight paper, achieves adaptive computation and emergent serial reasoning via neuron-level dynamics and synchronization.
- The White House launched Genesis Mission to accelerate AI-for-science; Anthropic is partnering with the US Department of Energy.
- Gemini 3 users report severe hallucinations and frequent instruction-ignoring.
- Manus.im users protest the removal of Chat Mode in favor of mandatory Agent Mode.
- An Anthropic engineer claimed software engineering will be "finished" in the first half of next year, sparking industry debate.
- LM Studio users request deprecating an unused system-prompt section.
- ArliAI released GLM-4.5-Air-Derestricted using Norm-Preserving Biprojected Abliteration, removing refusals while preserving reasoning. (Hugging Face)
- Qwen3-Next models are now supported in llama.cpp; users report ~12 tokens/sec on an RTX 5070 Ti. (GitHub PR)
- The BASI Jailbreaking community published a Gemini 3.0 jailbreak guide, bypassing safety filters via uploaded Google Docs files and multilingual prompts (e.g., Croatian).