This brief covers AI news from 2026-09-22 UTC.
xAI’s Grok 4.7 lands with a price that undercuts the frontier, Alibaba goes big on parameters and silicon, and SoftBank taps the bond market for OpenAI.
xAI ships Grok 4.7 at $2 per million input tokens
What happened: xAI released Grok 4.7, a 2.1 trillion parameter model, after at least five delays since late July. It is live in the Grok app, Cursor, Grok Build and the xAI API at $2 per million input tokens and $6 per million output tokens, unchanged from Grok 4.6.
Why it matters: Developers get a frontier-adjacent coding model at roughly a quarter of Claude Fable 5.1’s output price, though independent scores trail the leaders.
Alibaba plans a 5 to 10 trillion parameter model
What happened: Alibaba CEO Eddie Wu said the company plans to train a new AI model with 5 trillion to 10 trillion parameters, and unveiled the Zhenwu V900 AI chip, which it says is the most powerful in China with three times the performance of its predecessor.
Why it matters: A model that size would dwarf today’s frontier releases, and the in-house chip signals Alibaba intends to train it without depending on export-restricted accelerators.
SoftBank launches $11B bond sale to fund OpenAI
What happened: SoftBank Group launched an $11 billion-equivalent bond sale, split into $10 billion of dollar notes and 1 billion euros of euro notes, to fund its third $10 billion OpenAI tranche due October 1. The bonds are expected to price September 24 and settle September 29.
Why it matters: The OpenAI investment is now financed with long-term debt rather than bridge loans, so the compute and hiring it pays for carries a fixed repayment schedule.
StepFun previews a 600B MoE model with 1M context
What happened: StepFun released Step 5 Preview, a 600 billion parameter mixture-of-experts model that activates 27 billion parameters per token and supports a 1 million token context window. It is available via API at $1.00 per million input tokens and $2.70 per million output tokens, with open weights due October 15.
Why it matters: Long-horizon agentic work gets another cheap hosted option, and the October open-weight release would let teams self-host instead of paying per token.
Google open-sources AX, an agent fleet orchestrator
What happened: Google published AX, a declarative orchestrator for running large fleets of autonomous agent workloads in a cluster, configured through ax.io/v1alpha1 YAML and a kubectl-shaped CLI. The launch drew roughly 626 points and 285 comments on Hacker News, much of it about the Kubernetes cluster its quickstart requires.
Why it matters: Teams already on Kubernetes can declare agent tasks as YAML instead of building their own scheduler, but the setup cost is real for anyone without a cluster.
Cognition adds cloud VM and SSH access to Devin
What happened: Cognition launched command-line controls and full SSH access for Devin Cloud, letting developers start a session with devin —cloud, switch a local session with /cloud, resume by ID or URL, and enter the remote VM directly to inspect files and run commands.
Why it matters: Coding agent sessions can now move between a laptop and a persistent cloud VM, so long-running work survives a closed terminal and stays inspectable.
MIND raises $72M for AI data loss prevention
What happened: MIND, an Israeli-founded AI-native data loss prevention company, raised $72 million in a Series B led by Crosspoint Capital Partners, with YL Ventures and Paladin Capital Group participating, bringing total funding to $112 million. It says it was the first data security company accepted into Anthropic’s Cyber Verification Program.
Why it matters: As agents touch more enterprise data, buyers get another vendor selling controls specifically for data moving through GenAI and agentic systems.