AI Industry Daily News
A roundup of the AI industry's day, centered on Codex Windows support, grok-build-0.1, Claude Opus 4.8, Command A+, and Rosalind Biodefense.
Today's highlights
Key topics and reactions
SpaceX Files SEC 8-K on Acquisition of Cursor Maker Anysphere at About $60B
SpaceX filed an 8-K with the U.S. Securities and Exchange Commission regarding its acquisition of Anysphere, the company behind the Cursor AI coding tool. The deal carries an effective valuation of approximately $60 billion.
Pending regulatory approval, the transaction is expected to close in the third quarter of 2026.
The acquisition is being watched as a major consolidation in the AI development-tools market.
GLM-5.2 (Max) Ranks First in Design Arena and Second in Code Arena Frontend
GLM-5.2 (Max) placed first in Design Arena, second in Code Arena: Frontend, and topped open models in the new Agent Arena. The model runs at around 300 tokens per second, and users cited Claude-level performance at roughly one-tenth the cost, with MIT-licensed weights planned.
Some users reported scores surpassing Claude Opus 4.7 on complex frontend, React, and HTML tasks, and one-shot working apps using a 1M-token context for large-scale debugging and agentic workflows.
Other testers were more critical, noting a loss to Claude Opus 4.6 in practical use and a 41.5 vs. 60.1 gap on Bridge Bench. Reported issues included an average 156-second response time, heavy quota consumption, and reduced steerability versus prior versions.
MiniMax M3 Gains Attention for Native Tool Calling and Self-Built Agent Workflows
MiniMax M3's native tool calling was singled out for letting users build agents immediately, including via llama.cpp. A workflow in which M3 builds and self-edits its own mini-agent ("Minion") circulated widely among users.
The model was reported to hold up across long agent loops with a 1M-token context and to recover well from bash errors.
Context consumption remained a challenge: one case consumed 9,600 requests within a five-hour quota, prompting calls for latency optimization.
xAI's Grok 4.3 Becomes Available on Amazon Bedrock and PowerPoint Add-Ins
Grok 4.3 became available on Amazon Bedrock and was integrated into PowerPoint add-ins and live screen sharing. Users reported it could turn a paper into a near-complete Beamer slide deck and generate slides directly from video analysis in one shot.
The Bedrock version supports reasoning-effort tuning and agentic workflows, drawing interest for enterprise deployment, though some cases required fine adjustments to mathematical details.
Grok 4.3 is also being integrated into Warp.
Category highlights
Foundation Models: StepFun, Gemma 4 E2B, Together AI, Riverflow Pro 2.5
StepFun released Step 3.7 Flash, an efficient agent-focused model. Gemma 4 E2B runs fast on Intel AI PCs with LiteRT NPU support. Together AI reported a large prefill-throughput improvement from redesigning GLM 5.1's indexer kernel. In images, Sourceful's Riverflow Pro 2.5 ranked first in Image Arena, Graphic Design Arena, and Image Edit, averaging 10 Elo above GPT Image 2.
Video: Runway Aleph 2.0, Seedance 2.0 Pro, VEED Fabric, Amazon Series
Runway Aleph 2.0 added frame editing, Seedance 2.0 Pro held the top spot, and VEED Fabric introduced hyper-realistic video up to five minutes. Seedance 2.0 cut prices and improved CapCut integration. Amazon greenlit three AI animation series, and fal and Sequoia announced a 72-hour video hackathon backed by Google DeepMind, xAI, and Kling, the latter used as the engine for the animated series Pequena Máquina.
Audio: Gemini 3.5 Live Translation, Gladia STT, Inworld Price Cut, Vapi Index
Updates included Gemini 3.5 live translation, a new Gladia STT, and TTS releases such as Miso One and Resemble AI Chatterbox v3, alongside a 50% Inworld API price cut and Retell AI's Launch Week. ZONOS2 went live on fal with voice cloning from short samples in 10+ languages. Vapi published a "Humanness Index" for AI voices, and Together AI cut per-turn voice-agent costs roughly sixfold in a Decagon case.
OpenAI Speeds Up Photo Capture and Attachment in ChatGPT for iOS
OpenAI revamped the photo capture and upload experience in the iOS ChatGPT app, smoothing the transition from the attachment menu to the camera and inserting captured images directly into the composer. A design engineer described the flow as a single continuous action. The change builds on a July 2025 update that consolidated photo and file attachment, though no specific timing benchmarks were published.
Microsoft Publishes Video on Its MAI Foundation Model Development Team
Microsoft AI released a short video introducing the team behind its MAI foundation models, emphasizing its "Humanist Superintelligence" philosophy about two weeks after Build 2026, where seven new MAI models were unveiled. The flagship MAI-Thinking-1 uses an MoE architecture with 35B active of about 1T total parameters and a 256K context window, scoring 52.8% on SWE-Bench Pro, 97.0% on AIME 2025, and 87.7% on LiveCodeBench v6.
Key trends
Codex, NotebookLM and Anthropic Expand Agent and Production Tooling
OpenAI Codex rolled out Computer use, a Chrome extension, and memory in the EEA, UK, and Switzerland. An enhanced NotebookLM completed rollout to all Google AI Ultra users. Anthropic published a blog on how teams deploy agents to production, reflecting an industry focus on credentials, observability, and cost management for production agents.
AWS Summit NYC, Databricks Summit, NVIDIA Robotics and Runway Festival
AWS Summit NYC and the Databricks Data + AI Summit took place this week. NVIDIA opened a Humanoid Robot Pavilion at Automate 2026 with live embodied-AI demos, and its Blackwell platform led MLPerf Training 6.0. Runway's fourth AI Festival is scheduled for July 30 in Tokyo.