Easy AI Daily | November 24, 2025

English translation of the November 24, 2025 AI industry digest from the Easy AI teaching project.

Model Updates & Performance

• Anthropic released Claude Opus 4.5, positioned as the best model for coding and agents. Pricing is 3x lower than Opus 4.1 at $5/$25 per million tokens, with new API features including effort control. It scored 80.9% on SWE-Bench Verified, a new SOTA. (official release | Twitter announcement)
• Google released Gemini 3 Pro, scoring 76.2% on SWE-Bench Verified. Users report hallucination issues and frequent instruction-ignoring. (benchmark results)
• OpenAI released GPT-5.1-Codex-Max, scoring 77.9% on SWE-Bench Verified, briefly holding SOTA at time of release.
• Google released Gemini 3 Image, topping the Artificial Analysis image benchmark. Supports 14 input images with improved photorealism and editing. (benchmark results)

Benchmarks & Reasoning

• Claude Opus 4.5 set multiple records: SWE-Bench Verified 80.9%, SWE-bench Pro 52%, BrowseComp-Plus 85.3%, ARC-AGI-1 80%, ARC-AGI-2 37.64%. (system card)
• Qwen3-VL-32B beat Kimi K1.5 on the MathVision benchmark by 24.8 points, showing stronger visual reasoning.

Tools & Ecosystem

• Claude Opus 4.5 API features: effort control (reasoning intensity), context compaction, and advanced tool use; supported on Bedrock and Vertex. (effort control docs)
• Windsurf (1.12.35 stable / 1.12.152 preview) added Claude Opus 4.5 support, offered for a limited time at Sonnet pricing (2x credits). (changelog)
• Weaviate v1.32 enabled 8-bit Rotational Quantization by default, claiming 98–99% accuracy retained with lower latency and better write performance.
• Weights & Biases launched Serverless LoRA inference: upload adapters and switch dynamically at inference time, with no cold starts.

Research & Technical Breakthroughs

• Zyphra (with AMD and IBM) released ZAYA1-base, an AMD-native MoE model (8.3B total / 760M active parameters) that beats Llama-3-8B on math and coding tasks.
• DiRL framework combines SFT with a diffusion-native RL algorithm (DiPO) for diffusion language models; an 8B model reached 83% on MATH500.
• Sakana AI's Continuous Thought Machines (CTM) — a NeurIPS spotlight — achieves adaptive computation and emergent sequential reasoning via neuron-level dynamics and synchronization, performing well on maze planning.

Industry & Policy

• The White House launched Genesis Mission to accelerate scientific discovery with AI; Anthropic is partnering with the US Department of Energy.

Community & Product Feedback

• Gemini 3 users report serious hallucinations and frequent failure to follow explicit instructions.
• Manus.im removed Chat Mode, forcing Agent Mode — a change some users are protesting.
• An Anthropic engineer claimed software engineering will be "finished" in the first half of next year, with AI-generated code as trustworthy as compiler output. (Reddit discussion)
• LM Studio users requested removal of the long-unused system prompt section.

Open Source & Local Models

• ArliAI released GLM-4.5-Air-Derestricted, using Norm-Preserving Biprojected Abliteration to remove refusals while preserving reasoning. (Hugging Face)
• Qwen3-Next models (e.g., Qwen3-Next-80B-A3B-Instruct) are now supported in llama.cpp; a user tested 12 tokens/sec on an RTX 5070 Ti. (GitHub PR)

Safety & Jailbreaks

• The BASI Jailbreaking community published a Gemini 3.0 jailbreak method using uploaded Google Docs files to bypass safety filters, with multilingual prompts.

*Source: Easy AI teaching project.*