Z.ai releases GLM-5.3: same base model as GLM-5.2, with all gains from post-training. Terminal Bench 3.0 jumps from 4.6 to 28.3, CyberGym hits a best-overall 84.5%, ExploitBench more than doubles vs GLM-5.2, and weights open within two weeks. A digest of official benchmarks, the environment synthesis pipeline, the security disclosure ledger, and API changes.
Anthropic released Claude Opus 5: a perfect-score gold at IMO 2026, roughly 4× the previous best on ARC-AGI-3, 93.6% on multi-agent BrowseComp, and Anthropic's strongest computer-use model to date. Prompt injection robustness improves dramatically; alignment scores are the highest ever, though a new “overconfidence” pattern appears. It crosses CB-1 but not CB-2, so ASL-3 safeguards remain. All key figures from the official System Card.
Moonshot AI has launched Kimi K3: a 2.8-trillion-parameter MoE built on KDA hybrid linear attention and Attention Residuals, with native visual understanding and a 1M-token context window — the world's first open-source model in the 3-trillion class. This post rounds up the official benchmarks, long-horizon coding cases, architecture notes, access channels, and pricing.
Thinking Machines has released Inkling, an open-weight model trained from scratch: a 975B-total / 41B-active MoE multimodal foundation model with native text, image, and audio input, up to 1M-token context, and controllable thinking effort. Full weights are on Hugging Face, and a 12B-active Inkling-Small was previewed alongside.
On July 9, 2026, OpenAI released the GPT-5.6 series (Sol/Terra/Luna), rated High in cybersecurity and biological-chemical capabilities, with unprecedented safety investment.
On 2026-07-08, OpenAI released GPT-Live (Live-1 / Live-1 mini), replacing Advanced Voice Mode with a full-duplex architecture; complex tasks are delegated to GPT-5.5 in the background. Covers architecture, benchmarks, availability, known limitations, and community feedback.
Meta Superintelligence Labs releases Muse Spark 1.1, a multimodal reasoning model built for agentic tasks with a 1M-token context window and major gains in tool use, coding, and multimodal understanding. A new OpenAI-SDK-compatible Meta Model API launches alongside at $1.25 / 1M input tokens.
SpaceXAI releases Grok 4.5: the first model trained specifically for coding and agentic tasks. ~1.5T params, 80 TPS. SWE Marathon 29% leads Opus 4.8; API $2/$6 with markedly higher token efficiency. Covers Cursor co-training, contamination notes, and community field tests.
Tencent Hunyuan open-sources Hy3: a 295B MoE model with 21B activated parameters and 256K context, under Apache 2.0 with no regional restrictions. Reasoning and agentic capabilities approach leading flagship models, with the hallucination rate dropping from 12.5% to 5.4%.
