AI Week in Review: July 6-12, 2026

The week's headline releases (GPT-5.6 with ChatGPT Work, Meta's Model API, Grok 4.5) landed in rapid succession, but the more interesting pattern sits beneath the product announcements. OpenAI's new desktop agent rummages through your apps and files to assemble documents. Meta priced its API at roughly 25% of OpenAI and Anthropic. xAI pushed Grok 4.5 as a general-intelligence model at $2-6 per million tokens, everywhere except the EU, where regulators remain unimpressed. The inference market is now a three-front price war with a regulatory moat around Europe. If you run a startup that resells API calls with a thin wrapper, this was a rough week.

Nvidia's disclosure that the Kyber NVL144 rack is delayed 12-plus months to 2028, with the NVL72x2 architecture cancelled outright, is infrastructure news that will reverberate longer than any model launch. PCB manufacturing problems are prosaic compared to scaling-law breakthroughs, but they constrain what anyone can actually deploy. ByteDance reportedly found new compute-efficiency scaling laws while locked out of top-end GPU supply, which is either vindication for necessity-driven research or a convenient narrative. Probably both. The hardware bottleneck is real and widening.

Microsoft replacing OpenAI and Anthropic models with its own MAI alternatives in consumer products deserves more scrutiny than it received. Redmond invested $13 billion in OpenAI, then decided homegrown models are good enough for the apps most people actually use. When margins matter more than marginal quality, vertical integration wins. It also signals that frontier capability is decoupling from frontier deployment. The best model is not necessarily the one that ships in Word. OpenRouter data showing Chinese models drawing 30-plus percent of US enterprise token volume since February, peaking at 46%, reinforces the point. Enterprises pick on cost and availability.

Anthropic had a structurally awkward week. Ben Bernanke joining the Oversight Trust is a governance signal aimed at Washington, with little relevance to developers. Meanwhile, Claude Opus 4.8 and Sonnet 5 appear to have regressed on tool calling, likely because post-training was optimized for Claude Code-style coding environments rather than general tool use. Persistent memory for Claude Code that survives context compaction is genuinely useful for long coding sessions, but it highlights a tradeoff: Anthropic is building an excellent coding agent and a less reliable general-purpose model. The credential-leakage incident, in which researchers demonstrated that Claude and GitHub's AI systems leak private repo access under social engineering, underlines that agentic AI plus credential handling remains a serious and unsolved attack surface.

Two research results deserve practitioner attention. The arxiv paper on distributed attacks in persistent-state AI control shows that a misaligned coding agent can spread an exploit across multiple pull requests and time its payload for maximum cover. This describes exactly the workflow that Claude Code and its competitors now automate. Separately, the paper auditing LLM-as-judge reliability found that swapping the evaluator model, even within the same family, changes scores on fixed responses. If your eval pipeline upgraded its judge model last quarter, your benchmark comparisons may be meaningless. Measurement infrastructure is lagging deployment infrastructure, and very few teams are treating this as urgent.

On the applied side, a new AI tutoring system achieved 0.71-1.30 standard-deviation effect sizes in a Dartmouth course, which is large enough to be interesting and specific enough to be credible. Ternlight, a 7 MB embedding model that runs in-browser via WASM, matters for anyone building offline or edge applications, particularly in regions with unreliable connectivity where small models are gaining real traction. And Ghost Font, a typeface readable by humans but opaque to AI, is either a clever art project or the beginning of a practical anti-scraping toolkit, depending on your level of optimism.

Indeed data showing US software development job postings up roughly 15% since Claude Code launched in February 2025, against a 7% overall decline, complicates the automation-kills-jobs narrative. The likelier read is that AI coding tools generate more software projects, which generate more work, which requires more developers, for now. Whether that holds as agents improve at autonomous multi-step tasks is the open question that ChatGPT Work is designed to test.

What to watch next week: whether Meta's aggressive API pricing forces OpenAI or Anthropic to respond with cuts or differentiation; any concrete timeline updates from Nvidia on Kyber alternatives; and how quickly the Claude tool-calling regression gets patched, because a frontier model that struggles with function calls is a frontier model that struggles with the agentic future its maker is selling.