Daily AI RoundupsThursday, 9 July 2026

The 09 Jul 2026 roundup leads with SpaceXAI introducing Grok 4.5 for coding and agentic work, Zhipu AI raising $4B as the GLM compute race intensifies, and OpenAI launching GPT-Live for full-duplex voice. It also covers BofA extending a $520M loan to OpenAI, Mistral moving into robotics with Robostral Navigate, Prime Intellect raising $130M for open superintelligence infrastructure, NVIDIA and Databricks publishing agent benchmark and cost signals, Claude Code adding /checkup, UCL/Gatsby research on RL post-training, OpenAI retracting its SWE-Bench Pro recommendation, China warning on Claude Code, Marble open-sourcing a curriculum graph, and a Gemini CLI model-string diff.

frontier models
voice models
coding agents
AI funding
robotics
AI benchmarks
China AI
developer tools

Stories in this roundup

SpaceXAI introduces Grok 4.5 for coding and agentic work, with availability through Grok Build, Cursor and the SpaceXAI consolexAI
Zhipu AI, maker of the GLM model family, raises $4B in a Hong Kong share sale as Chinese frontier labs keep funding the compute raceReuters
OpenAI launches GPT-Live, a full-duplex voice model family that can listen and speak at the same timeOpenAI
BofA extends first $520M loan to OpenAI ahead of IPOReuters
Mistral launches Robostral Navigate, its first robotics model, using a single RGB camera for embodied navigationMistral
Prime Intellect announces a $130M Series A to build the Open Superintelligence Stack, led by Radical Ventures with NVIDIA, Intel Capital and Dell CapitalPrime Intellect / X
NVIDIA says Nemotron 3 Ultra achieves benchmark-leading performance at lower cost than top closed models on LangChain’s agent orchestration platformNVIDIA
Databricks benchmarks coding agents on its multi-million-line codebase and finds harness choice can cut costs by about 2x, with GLM 5.2 performing stronglyDatabricks
Boris Cherny says Claude Code added /checkup, a command for cleaning unused skills, MCPs and plugins, deduping CLAUDE.md files, turning off slow hooks and pre-approving read-only commandsBoris Cherny / X
UCL/Gatsby researchers Azwar Abdulsalam, Nishil Patel and Andrew M. Saxe argue that RL post-training builds reusable compositional reasoning strategies rather than only amplifying latent skillsarXiv
OpenAI retracts its recommendation for SWE-Bench Pro after estimating that roughly 30% of the coding benchmark’s tasks are brokenOpenAI
China issues a backdoor security alert over older Claude Code versions, while Anthropic says the mechanism was an anti-abuse measureReuters
Marble open-sources a primary-school curriculum graph with 1,590 concepts and 3,221 prerequisite links across eight subjectsMarble / X
Gemini CLI v0.50.0 npm diff reveals new model strings gemini-apozrzef and o7yzxvzwnpm

Back to all roundups