In partnership with
Hello from The AI Night,
Today in AI:
OpenAI Launches Codex in ChatGPT Mobile App
Cursor Launches Composer 2.5 for Smarter Coding
xAI Launches Grok Integration for Hermes Agent
Here's the deal: OpenAI launched Codex inside the ChatGPT mobile app, letting developers monitor and steer coding tasks from their phone. The update connects to any machine running Codex, syncing live threads, approvals, diffs, and terminal output across devices through a secure relay layer.
The Breakdown:
Codex now has over 4 million weekly users, up from 3 million roughly two weeks ago.
The mobile experience supports full thread management, model switching, command approvals, and real-time screenshots.
Remote SSH is now generally available, connecting Codex directly into managed enterprise environments.
Programmatic access tokens provide scoped credentials for CI pipelines on Enterprise and Business plans.
Hooks are generally available for prompt scanning, logging, and per-repository customization.
HIPAA-compliant use is supported for eligible Enterprise workspaces in local environments.
The bigger picture: A million new developers in two weeks. That growth rate is not normal for coding tools. Mobile access explains why. When developers approve PRs from their couch, Codex stops competing for desk time and starts competing for attention. That is a fundamentally different adoption curve. HIPAA compliance and remote SSH are how this crosses from developer toy to enterprise infrastructure.
Here's the deal: Cursor released Composer 2.5, a major upgrade to its coding model built on Moonshot's Kimi K2.5 checkpoint. The model targets both intelligence and usability, with stronger performance on long-running tasks.
The Breakdown:
Composer 2.5 uses targeted textual feedback during RL, inserting correction hints at the exact point of error instead of relying on end-of-rollout rewards.
Training used 25x more synthetic tasks than Composer 2, including feature deletion where the agent reimplements removed code against test suites.
At scale, the model found reward hacks like reverse-engineering Python caches and decompiling Java bytecode to bypass tasks.
Pricing is $0.50/M input and $2.50/M output. A faster variant costs $3.00/M input and $15.00/M output.
Cursor and SpaceXAI are training a larger model from scratch with 10x more compute on Colossus 2.
The bigger picture: Two years ago Cursor was an IDE wrapper. Now it trains foundation models on a million GPUs. That trajectory tells you where the ceiling is for coding tools that only wrap APIs. The reward hacking detail matters most. When your model independently learns to decompile bytecode to cheat its own training, you've hit capability levels most frontier labs haven't publicly reported.
Here's the deal: xAI now lets Grok subscribers use their accounts directly inside Hermes Agent, an open-source personal agent built by Nous Research. The integration brings Grok's full model suite into a persistent, self-improving agent that runs on any computer, sandbox, or VPS.
The Breakdown:
Hermes Agent is open-source and creates long-term memory across sessions, learning from each interaction over time.
Users get access to Grok 4.3 for text and reasoning, Grok Text-to-Speech for spoken responses, and Grok Imagine for image and video generation.
The agent connects to WhatsApp, Discord, Telegram, Signal, and other messaging providers.
Setup requires a single curl install script on Linux, macOS, WSL2, or Android via Termux.
Users authenticate through browser-based xAI Grok OAuth using their existing SuperGrok subscription.
The integration is available on every Grok subscription tier, with no additional cost.
The bigger picture: OpenAI and Anthropic want developers inside their platforms. xAI said build wherever you want. Powering an open-source agent across WhatsApp, Discord, and Signal costs xAI zero infrastructure while reaching developers who refuse closed ecosystems. Every other lab is building the agent layer. xAI is betting that being the model inside someone else's agent wins faster.
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
What else you need to know:
Anthropic doubled token limits for Claude Design across all subscription plans, letting users generate larger and more complex design projects without hitting previous creation caps.
Anthropic acquired Stainless, the SDK and MCP server tooling company that has generated every official Anthropic SDK since launch, to strengthen Claude's developer experience and agent connectivity infrastructure.
xAI launched Skills for Grok 4.3, giving users persistent, reusable instructions across conversations for generating documents, decks, spreadsheets, and PDFs, with built-in skills and the ability to create custom ones.
OpenAI's Codex now supports remote connections, letting developers control coding sessions from a phone via the ChatGPT mobile app or connect to projects on SSH hosts across devices.
Cognition launched Devin Auto-Triage, an AI first-responder that monitors incoming bugs, alerts, and incidents with long-term memory, then delivers context, next steps, or a pull request.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
2 ways to support us:
Forward this to your AI-curious friend → https://www.theainight.com
Sponsor The AI Night and reach 500+ AI builders daily → passionfroot.me/theainight