Model launchesOpenAI begins limited preview of GPT‑5.6 family: Sol, Terra, and LunaOpenAI has started a limited preview of the GPT‑5.6 model family—Sol (highest capability), Terra (balanced), and Luna (fast and low‑cost)—with phased access for trusted partners coordinated with the U.S.4 min read
Model launchesZ.ai releases GLM-5.2: open-weights 753B model with 1‑million‑token context and competitive agentic performanceZ.ai announced GLM-5.2, an open-weights mixture-of-experts transformer with 753 billion parameters and an input context up to 1 million tokens.4 min read
Model launchesApple unveils AFM 3 Core Advanced — an MoE-inspired model optimized for on-device useApple announced AFM 3 Core Advanced, a 20-billion-parameter foundation model variant optimized to run locally on Apple silicon.4 min read
Model launchesLiquid AI launches 230M‑parameter LFM2.5‑230M for on‑device data extraction and agentic edge workflowsLiquid AI released LFM2.5‑230M, a 230‑million‑parameter foundation model optimized for on‑device data extraction, tool calling and autonomous edge systems.4 min read
Model launchesOpenAI introduces updated version of GPT-5.5 InstantOpenAI released a new version of GPT-5.5 Instant, which arrives with improved intent recognition, more reliable handling of complex constraints, and better shopping and local recommendations; the…1 min read
Model launchesOpenAI quietly releases cyber model rivaling Mythos as Anthropic faces export scrutinyOpenAI introduced GPT-5.5-Cyber with strong performance on an internal vulnerability benchmark and limited, defender-only access, while Anthropic remains subject to export restrictions tied to the Trump administration's concerns about its Fable 5 and Mythos models.2 min read
Model launchesGLM-5.2: an open-weight model that shifts agent capabilitiesZ.ai’s GLM-5.2, released to GLM Coding Plan members on June 13 and with MIT-licensed weights published June 16, has generated strong community praise and benchmark results that place it alongside leading closed models.4 min read
Model launchesEvaluating Claude Fable 5: safeguarded responses, hidden fallbacks and benchmark distortionMultiple independent evaluators found it difficult to measure Anthropic’s Claude Fable 5 because safety classifiers either refused prompts or routed them to the weaker Claude Opus 4.8, and Anthropic requires prompts and outputs to be retained for 30 days.5 min read
Model launchesNvidia releases Nemotron 3 Ultra: a large open-weights model for long-context agentic tasksNvidia has unveiled Nemotron 3 Ultra, a 550-billion-parameter open-weights large language model optimized for long-context, agentic workloads.4 min read
Model launchesGPT-5.5 Instant now matches the most advanced Thinking models on health questionsThe GPT-5.5 Instant in ChatGPT now rivals the most advanced Thinking models in answering health questions, performing better at identifying when urgent care is needed, at requesting context, and at explaining uncertainty.1 min read
Model launchesCursor releases Composer 2.5 — agentic coding model optimized for speed and costCursor unveiled Composer 2.5, an agentic software-engineering model built from Moonshot’s open-weights Kimi K2.5 and further pretrained and fine-tuned for coding agents.4 min read
