Easy AI Daily | November 24, 2025 AI Industry Update
Model Updates & Performance
Anthropic releases Claude Opus 4.5, positioning it as the best model for coding and agents Anthropic launched its flagship Claude Opus 4.5 at a price 3x lower than Opus 4.1 ($5/$25 per million tokens), adding new API features such as effort control. It scored 80.9% on SWE-Bench Verified, a new SOTA.
- Official release
- Benchmark results
- Claude Opus 4.5 sets multi-benchmark 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 beats Kimi K1.5 on MathVision by 24.8 points, showing stronger visual reasoning.
- Claude Opus 4.5 new API features: effort control, context compaction, and advanced tool use; supported on Bedrock, Vertex, and other cloud platforms. Effort docs
- Windsurf supports Claude Opus 4.5: stable 1.12.35 and preview 1.12.152 releases; Opus 4.5 available at Sonnet price (2x credits) for a limited time. Changelog
- Weaviate v1.32 defaults to 8-bit Rotational Quantization, claiming 98–99% accuracy retention with lower latency and improved write performance.
- Weights & Biases launches Serverless LoRA inference: upload adapters and switch them dynamically at inference time with no cold-start issues.
- Zyphra launches AMD-native MoE model ZAYA1-base: built with AMD and IBM, 8.3B total parameters (760M active), outperforming Llama-3-8B on math and coding tasks.
- DiRL framework optimizes diffusion language models: combines SFT with the diffusion-native RL algorithm DiPO; an 8B model reached 83% on MATH500.
- Sakana AI's Continuous Thought Machines (CTM), a NeurIPS spotlight paper, achieves adaptive computation and emergent sequential reasoning via neuron-level dynamics and synchronization, performing well on maze planning.
- US launches Genesis Mission: a White House initiative to accelerate scientific discovery through AI; Anthropic is partnering with the US Department of Energy.
- Gemini 3 hallucination complaints: users report severe hallucinations and ignoring explicit instructions (e.g., generating a third option when told not to).
- Manus.im users protest Chat Mode removal: forced switch to Agent Mode has upset some users demanding Chat Mode's return.
- Anthropic engineer claims software engineering will be "finished" in the first half of next year, arguing AI-generated code will be as trusted as compiler output; sparked industry discussion. Reddit thread
- LM Studio users request deprecating the system prompt section, citing two years of disuse.
- ArliAI releases GLM-4.5-Air-Derestricted: uses Norm-Preserving Biprojected Abliteration to remove refusal behavior while preserving reasoning; based on the Gemma 3 12B architecture. Hugging Face
- Qwen3-Next supported in llama.cpp: user tests show ~12 tokens/sec on an RTX 5070 Ti for Qwen3-Next-80B-A3B-Instruct. GitHub PR
- BASI Jailbreaking community publishes a Gemini 3.0 jailbreak guide: bypasses safety filters by uploading a Google Docs file, with multilingual prompts (e.g., Croatian).
Google releases Gemini 3 Pro, scoring 76.2% on SWE-Bench Verified Google's Gemini 3 Pro achieved 76.2% on SWE-Bench Verified, but users report hallucination problems and frequent instruction-ignoring behavior.
OpenAI launches GPT-5.1-Codex-Max, scoring 77.9% on SWE-Bench Verified OpenAI's GPT-5.1-Codex-Max briefly held SOTA on SWE-Bench Verified at 77.9%, boosting coding capabilities.
Google releases Gemini 3 Image, topping image benchmarks Gemini 3 Image ranked first on Artificial Analysis image benchmarks, supports 14 input images, and improves photorealism and editing capabilities.
Benchmarks & Reasoning
Tools & Ecosystem
Research & Technical Breakthroughs
Industry & Policy
Community & Product Feedback
Open Source & Local Models
Safety & Jailbreaks
*Source: Easy AI teaching project*