XNews

Recent model releases and benchmark news

Recent model releases and benchmark news

Major model releases and benchmark updates mark a competitive wave across vendors.

Key insights

OpenAI releases GPT‑5.2 with large benchmark gains

OpenAI published GPT‑5.2 Thinking results showing significant gains over GPT‑5.1: SWE‑Bench software engineering up to 55.6%, GPQA Diamond at 92.4%, CharXiv scientific reasoning at 82.1%, FrontierMath overall 40.3%, AIME 2025 perfect score 100%, and large jumps on ARC‑AGI benchmarks and GDPval knowledge tasks. The release was promoted as a major step in reasoning performance.

OpenAI also released GPT‑5.2‑Codex for agentic programming

OpenAI released GPT‑5.2‑Codex focused on agentic programming with upgrades for long‑term tasks, defensive cybersecurity capabilities, and improved visual UI understanding. Benchmarks: SWE‑Bench Pro 56.4% and Terminal‑Bench 2.0 at 64.0%, with added deployment safety measures for powerful coding actions.

Anthropic launches Opus 4.5 with computational and pricing improvements

Anthropic released Claude Opus 4.5 (claude‑opus‑4‑5‑20251101) with SWE‑Bench Verified >80 and OSWorld agent scores of 66.3%. Opus 4.5 reduces token usage by up to 65% in long tasks and price per million tokens dropped to $5 input / $25 output versus prior much higher rates. Opus 4.5 is available via multiple cloud platforms and supports mixed inference control via an 'effort parameter'.

Qwen3 and Qwen3‑Omni expand open ecosystems with FP8 and omni‑modal variants

Alibaba Qwen announced Qwen3‑Next‑80B FP8 families for fast inference and Qwen3‑Omni unifying text, image, audio and video with SOTA on many audio/AV benchmarks. Several Qwen3 models were open‑sourced (30B variants) and integrated with Hugging Face and ModelScope.

Meta released open research Code World Model (CWM) 32B

Meta FAIR published Code World Model (CWM), a 32B research model with pass@1 metrics: 65.8% SWE‑bench Verified, 68.6% LiveCodeBench, 96.6% Math‑500, and 76.0% on AIME 2024. Meta encouraged the research community to use the open weights.

Google releases FunctionGemma and expands Google AI Plus

Google introduced FunctionGemma, a small on‑device Gemma 3 variant for converting natural language into executable device commands for mobile and IoT, and expanded Google AI Plus to 36 more countries, increasing access to Nano Banana image generation and other multimodal services.

Source posts