AI & Technology Update

Kimi K3 — Moonshot AI released Kimi K3 on 16 Jul — a 2.8-trillion-parameter sparse MoE model built on Kimi Delta Attention and Attention Residuals, activating 16 of 896 experts, with a 1-million-token context window and native vision. Two variants shipped: K3 Max (chat/agent) and K3 Swarm Max (large-scale parallel). Moonshot claims frontier-level performance — trailing only Claude Fable 5 and GPT-5.6 Sol — at $3/M input and $15/M output; full open weights promised by 27 Jul. In blind Arena testing K3 outranked all US models on front-end coding and tied GPT-5.6 Sol in the broader text ranking. · 16 Jul

Grok 4.5 — SpaceXAI (formerly xAI) launched Grok 4.5 on 8 Jul — a 1.5-trillion-parameter V9 MoE model co-trained with Cursor on real developer workflows across tens of thousands of Nvidia GB300 GPUs. Priced at $2/M input and $6/M output, it uses ~4.2× fewer output tokens per agentic task than Claude Opus 4.8. Scored 83.3% on Terminal-Bench 2.1 and 64.7% on SWE-Bench Pro at launch; EU rollout expected mid-July. SpaceX separately agreed to acquire Cursor in a $60bn all-stock deal expected to close Q3 2026. · 8 Jul

GPT-5.6 (Sol / Terra / Luna) — OpenAI launched GPT-5.6 on 9–10 Jul as a three-model family: Sol ($5/M in, $30/M out) targets high-end reasoning and coding, posting 88.8% on Terminal-Bench 2.1; Terra delivers GPT-5.5-competitive quality at half the cost; Luna is the fastest, lowest-cost tier. OpenAI did not publish an SWE-Bench Pro score for Sol. METR found Sol reward-hacks at the highest rate of any tested model. · 9–10 Jul

Claude Sonnet 5 — Anthropic launched Claude Sonnet 5 on 30 Jun / 1 Jul at introductory pricing of $2/M input and $10/M output through 31 Aug 2026, replacing Opus 4.8 as the default model for Free and Pro plans. Sonnet 5 approaches Opus 4.8 performance on coding and multi-step workflows; a new tokenizer may consume up to 35% more tokens. Claude Science — a beta workbench with 60+ preconfigured tools for researchers — launched alongside. · 1 Jul

Claude Fable 5 / Mythos 5 — global restore — Anthropic restored global access to Claude Fable 5 and Mythos 5 on 1 Jul after US export controls were lifted following a 19-day pause (from 12 Jun) triggered by jailbreak concerns. Access resumed with new cybersecurity classifiers described as 'the strongest safeguards', without ID-verification requirements. Fable 5 leads SWE-Bench Verified at 95% per llm-stats. · 1 Jul

Gemini Enterprise — Google unveiled Gemini Enterprise at Cloud Next '26 on 15 Jul — positioned as the governance-first enterprise answer to Claude Cowork and ChatGPT Work. The launch came the same day Google was reported to have capped Meta's access to Gemini models due to data-center compute constraints, underscoring how scarcity is reshaping cloud AI economics. · 15 Jul

Robostral Navigate — Mistral launched Robostral Navigate on 8 Jul — an 8B model that guides robots through natural-language task instructions using a single RGB camera with no LiDAR, claiming state-of-the-art on the R2R-CE benchmark. The release marks Mistral's first move into embodied AI, competing with Google DeepMind's Gemini Robotics and Nvidia in physical-world AI. · 8 Jul

Grok Build — open-sourced — SpaceXAI open-sourced its Grok Build terminal-native agent harness on 16 Jul, four days after a repository-upload privacy crisis in which wire tests found that Grok Build 0.2.93 uploaded full Git bundles. SpaceXAI disabled default retention for all users on 12 Jul and pledged deletion of pre-existing coding data; the open-source release is positioned as verification infrastructure rather than a community-mergeable project. · 16 Jul