01 OpenAI arms Codex with desktop control, image tools and memory to challenge Anthropic
OpenAI rolled out a major update to Codex that expands the agent’s capabilities: it can use a user’s computer, generate images, and retain memories from past interactions. The changes package a set of agentic features intended to make Codex more autonomous and useful for development workflows.
The timing highlights an intensifying product rivalry with Anthropic. Coverage frames the update as a direct response to Anthropic’s gains with Claude Code: OpenAI has been shifting resources toward stronger developer tools and agent features to keep parity in coding and automation use cases.
Practical implications are straightforward: developers and teams that adopt Codex can expect broader local integration (desktop access), multimodal outputs (image generation), and persistent context (memory), which together change how agents can assist with longer-running coding or debugging tasks.
- Codex now has desktop access, image generation, and memory features.
- OpenAI’s update is widely read as a competitive move against Anthropic’s developer tooling.
- The new capabilities aim to make Codex more autonomous for extended coding and automation workflows.
02 InsightFinder raises $15M to pinpoint where AI agents break down in production
InsightFinder announced a $15 million funding round to build tooling that helps companies pinpoint where AI agents and models go wrong. The startup’s pitch centers on monitoring and diagnosing problems not just in models but across the full tech stack now that AI is embedded in workflows.
According to the company’s CEO, the challenge is broader than model-level errors: teams need visibility into integrations, data flows, and runtime behavior of agents. The funding will support product development aimed at runtime observability and root-cause analysis for agent-driven systems.
For enterprises deploying agentic AI, the takeaway is clear: there’s growing demand and investor interest in post-deployment tools that can trace failures, tool abuse, or miscoordination across services—not just traditional model evaluation metrics.
- InsightFinder raised $15 million to expand monitoring and diagnostics for AI agents.
- The company stresses diagnosing the whole tech stack, not only model errors.
- Investors see observability for agentic systems as a practical enterprise need.
03 Gemini’s Personal Intelligence pulls Google Photos to generate context-aware images
Google updated Gemini’s Personal Intelligence to let the model pull data from apps such as Google Photos and combine that context with its Nano Banana 2 image model to generate personalized images. The feature lets users give prompts that reflect their own content—examples include asking Gemini to design a dream house or create imagery tied to personal photos.
The change ties Google’s app-level data integrations to multimodal generation, making image outputs more personally contextualized than generic image models. That raises practical questions about data access and user controls, since the feature leverages a user’s own stored photos to inform generated results.
For users, the update is a step toward more customized creative tooling inside Google’s ecosystem. For developers and product managers, it signals how cloud-backed personal data and multimodal models are being combined to create user-specific experiences.
- Gemini can access Google Photos to inform image generation via Nano Banana 2.
- The update couples personal-data integrations with multimodal image capabilities.
- Expect more app-level, personalized creative features inside Google’s ecosystem.
What moved around the edges
Anthropic releases Claude Opus 4.7 for tougher engineering tasks
Anthropic announced Claude Opus 4.7 as its most powerful generally available model to date, positioned as an improvement over Opus 4.6 for complex coding tasks and image analysis.
The Verge AIFactory closes $150M at a $1.5B valuation to sell AI coding to enterprises
Three-year-old startup Factory closed a $150 million funding round led by Khosla Ventures, valuing the company at $1.5 billion as it scales AI coding tools for enterprise customers.
TechCrunch AIGoogle publishes personalized-image examples for Gemini’s Personal Intelligence
Google’s AI blog published example images and use cases showing how Personal Intelligence plus Nano Banana can create images tailored to a user’s photos and prompts inside the Gemini app.
Google AI BlogOpenAI introduces GPT‑Rosalind, a biology‑tuned LLM available in closed access
OpenAI began offering GPT‑Rosalind, an LLM trained on biology workflows; the model is available in closed access for specialists rather than general public rollout.
Ars Technica AIRunway CEO says AI could shift studios from one big blockbuster to many smaller films
Runway’s CEO suggested that AI could let Hollywood make dozens of lower-cost films rather than betting on a single $100M blockbuster, arguing volume could raise the odds of hits for studios.
Hacker News AI