OpenAI, xAI, and Anthropic compressed a quarter’s worth of launches into a single week, but the more consequential story is Microsoft, Meta, and DeepSeek all deciding to build instead of rent.
The Frontier Sprint: Three Labs Ship in the Same Week
OpenAI, xAI, and Anthropic all moved at once, competing on price, scale, and reach rather than a single headline benchmark.
- Claude Cowork Goes Live on Web and Mobile for Max Users. Anthropic’s agentic work sessions now follow users across web and mobile, running scheduled tasks in the background even when no device is online, a shift from sessions tied to a single browser tab.
- OpenAI Sets GPT-5.6 Public Launch for Thursday, Undercuts 5.5 on Price. Sol, Terra, and Luna launch globally on July 9, with OpenAI claiming Terra matches GPT-5.5’s performance at half the inference cost.
- SpaceXAI to Launch Grok 4.5 Thursday, Musk Claims Opus-Class Power. Musk says the 1.5 trillion parameter model beats Anthropic’s Opus on speed and cost, a claim no independent benchmark has yet confirmed.
Who Controls the Stack: Microsoft, Meta, and DeepSeek Cut Out the Middleman
Three separate stories point to the same shift: companies that once rented frontier AI are building the pieces themselves.
- Microsoft Shifts Excel and Outlook AI to Its Own MAI Models. Tens of thousands of weekly Excel and Outlook prompts now route to Microsoft’s in-house MAI models, timed to when the discounted token deals that made outside labs cheap are set to expire.
- DeepSeek Moves to Design Its Own Inference Chips. After a year of quiet hiring and partner meetings, the Hangzhou lab is building custom silicon to loosen its dependence on both Nvidia and Huawei.
- Meta’s Muse Image skips the API, lands in WhatsApp and IG instead. Meta’s first Superintelligence Labs image model ships straight into apps used by billions of people, bypassing the developer-first rollout OpenAI and Google both chose.
- Microsoft’s Real AI Bet Is the Stack, Not the Chatbot. Copilot is the storefront. Azure, MAI, and multi-model sourcing are the actual fight for control of the enterprise AI stack, positioning Microsoft to win regardless of which chatbot wins the headlines.
Agents Learn to Work Unsupervised
Google and CopilotKit both shipped changes this week that push agents further from human oversight, not closer to it.
- Google Turns Gemini’s Managed Agents Into Async Background Workers. New background execution, remote MCP support, and credential refresh let Gemini agents run unattended without holding a connection open, closing a gap with Anthropic’s own background push.
- When Users Fix Your Agent’s Mistakes, That Fix Is Data. CopilotKit is turning the corrections users make to agent errors into structured, scoped memory, a bet that fixing agents in production beats retraining them from scratch.
Under the Hood: The Research Nobody Skimmed This Week
Behind the launches, three papers and essays make the more durable argument about where AI actually improves next.
- AI’s self-improvement bottleneck isn’t the model. It’s the harness.. Lilian Weng argues near-term recursive self-improvement runs through better agent scaffolding, not models rewriting their own weights, a distinction that changes where labs should be investing.
- Gemma 4’s technical report bets on ditching AI encoders. Google DeepMind’s paper details the MoE routing, thinking mode, and encoder-free design behind Gemma 4, filling in the architecture choices the phone-ready weights release left unexplained.
- Liquid AI’s Antidoom Nearly Eliminates Reasoning-Model Repetition Loops. A narrow post-training fix targets only the tokens that trigger repetition loops, avoiding the quality tradeoffs that come with blanket repetition penalties.
Today’s Quick Hits
- OpenAI’s New Realtime Mini Model Adds Reasoning, Tool Use, Same Price. GPT-Realtime-2.1-mini brings reasoning and function calling to OpenAI’s budget voice tier without raising the price.
- Anthropic Offers Free Claude Fable 5 Access Through July 12. Pro, Max, and Team subscribers can spend half their weekly limit on the new model before usage credits take over.
- Alignment eval pass rates hide an untested variable: sensitivity. A LessWrong analysis argues a 97 percent pass rate says nothing about the misalignment an eval could miss, and proposes four ways to measure that gap.
- Microsoft finds coding agents do worse with JSON than plain CLI args. A five-model test found JSON CLI payloads never beat conventional flags on correctness and cost up to 11 times more, largely due to shell escaping failures.