AI NewsSeptember 22, 2026

OpenAI rolled out a credit score tracker for US ChatGPT Plus and Pro subscribers, letting them securely link an Experian report to view their VantageScore 3.0 and get personalized financial advice inside the chat.
OpenAI rolled out a credit score tracker for eligible US ChatGPT Plus and Pro subscribers on September 21. Users securely connect an Experian credit report to view their VantageScore 3.0, key factors like payment history, and personalized advice such as card trade-offs.
Reports update monthly, and OpenAI says the credit data stays within its own secure widget, not exposed to or used by the underlying model. Disconnecting removes synced account data within 30 days, though related chat history isn't automatically deleted.
Key Capabilities:

• Secure linkage: Experian data displays in a dedicated widget, kept separate from the AI model itself.
• Actionable advice: Get personalized suggestions like which card to pay down or apply for.
• Monthly updates: Credit reports and alerts refresh automatically as your score changes.

Belgian security firm Aikido released Altar-1, an open-weight model compressed from Z.ai's GLM-5.3, letting companies run AI-powered threat detection entirely on their own infrastructure.
Aikido Security released Altar-1 on September 21, an open-weight cybersecurity model compressed from Z.ai's GLM-5.3. Aikido pruned 88 of its 256 routed experts, shrinking it from 1.5 TB at full precision to 328 GB, small enough to run on four NVIDIA H200 GPUs, even air-gapped.
It lets companies run AI-assisted vulnerability analysis without sending source code, architecture docs, or findings to an outside inference provider, a priority for regulated customers like Belgian bank Belfius. Aikido is already deploying it inside its own pentesting platform, Aikido Machine.
Key Capabilities:

• Massive compression: Shrunk from 1.5 TB at full precision to 328 GB while keeping most reasoning quality.
• Runs air-gapped: Deployable on four NVIDIA H200 GPUs with no external network connection needed.
• Already in production: Powers Aikido's own on-premises pentesting platform for customers like Belfius bank.

Moonshot AI launched Kimi Code Desktop for macOS and Windows, bringing its terminal-based coding agent into a graphical interface for managing multiple agents and long-horizon programming tasks.
Moonshot AI launched Kimi Code Desktop on September 21 for macOS and Windows, carrying its Kimi Code CLI's agent core into a graphical interface. It includes a built-in terminal, browser, and Git status viewer for running, debugging, and reviewing projects in one window.
It manages multiple agents in one focused space, runs tasks in parallel, and stays in sync on long-horizon work. Three permission modes, Always Ask, Ask When Needed, or Never Ask, control how much autonomy the agent gets over sensitive operations.
Key Capabilities:

• Multi-agent management: Run and monitor several coding agents in parallel from one interface.
• Built-in tools: Terminal, browser, and Git status viewer without switching apps.
• Granular permissions: Choose Always Ask, Ask When Needed, or Never Ask for sensitive operations.

Google opened preorders for Googlebook, a new premium laptop category running Android with Gemini Intelligence at its core, featuring 2.8K OLED screens and up to 14 hours of battery life.
Google opened preorders on September 21 for Googlebook, a new laptop category starting at $899 built around Android and Gemini Intelligence. Initial models from Acer, Asus, Dell, HP, and Lenovo feature up to 2.8K OLED touchscreens and up to 14 hours of battery life.
A feature called Magic Pointer brings Gemini to whatever's on screen with a wiggle of the cursor, syncing seamlessly with an Android phone. Devices ship October 4 in the US and include 12 months of Google AI Pro, worth 5TB of storage plus Gemini Advanced.
Key Capabilities:

• AI-first cursor: Magic Pointer summons Gemini to help with anything visible on screen.
• Phone sync: Pairs seamlessly with an Android phone to pick up tasks across devices.
• Bundled subscription: Includes 12 months of Google AI Pro, worth 5TB of storage and Gemini Advanced.

ByteDance launched Dramagic, an enterprise AI platform that takes a script through analysis, asset setup, storyboarding, and video preview, aimed at China's booming AI-generated short drama industry.
ByteDance launched Dramagic on BytePlus, an enterprise AI platform covering the full pipeline for short dramas: script analysis, asset setup, storyboard, and video preview. It includes built-in consistency checks for characters and scenes plus multi-user collaboration for production teams.
The launch rides a massive wave: about 128,000 short dramas were published in China in Q1 2026 alone, three times the prior year's total, with 95% AI-generated. One researcher estimates AI video cuts production costs to about a tenth of traditional filming.
Key Capabilities:

• End-to-end pipeline: Script analysis through storyboard to video preview in one connected workflow.
• Built for scale: Multi-user collaboration and consistency checks for characters and scenes.
• Rides a boom: China's AI-generated short drama output tripled year-over-year in Q1 2026 alone.

Xiaomi open-sourced MiMo-V2.6 Pro and Flash, two omnimodal models that perform on par with Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks, with Pro topping the Artificial Analysis Intelligence Index among open-source models.
Xiaomi released MiMo-V2.6 Pro and Flash as open weights on September 21. Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks and scores 46.32 on the Artificial Analysis Intelligence Index, the highest of any open-source model.
Both models are natively multimodal with a 1M-token context, MIT licensed for commercial use. Xiaomi also released the technical report, RL training environments, and training code alongside the weights, letting other teams reproduce and build on the release.
Key Capabilities:

• Top open-source score: 46.32 on the Artificial Analysis Intelligence Index, ahead of Kimi K3 and Qwen3.8 Max.
• Fully open: MIT licensed, with technical report, RL environments, and training code released.
• 1M-token context: Native multimodal support for text, image, video, and audio.

SpaceXAI launched Grok 4.7, a notable jump over Grok 4.6 in coding and reasoning benchmarks, at unchanged pricing of $2 per million input tokens and $6 per million output tokens.
SpaceXAI launched Grok 4.7 on September 21, built on a larger base model with a longer reinforcement learning run on harder, multi-hour tasks. It scores 46.3% on CursorBench 4.0, up from 40.4%, at the same $2/$6 per million token pricing as before.
It also improves on professional benchmarks: 19.6% on the Harvey Legal Agent test versus 15.8% before, and 56.7% on clinical reasoning versus 48.5%. A faster, pricier variant offers twice the output speed, and it's live now in Cursor, Grok Build, and the Grok API.
Key Capabilities:

• Coding gains: Scores 46.3% on CursorBench 4.0, up from Grok 4.6's 40.4%.
• Professional work: 19.6% on Harvey's legal benchmark and 56.7% on clinical reasoning, both improved.
• Same pricing: $2 per million input tokens, $6 per million output, unchanged from Grok 4.6.