AI News: July 8, 2026

1. Microsoft Phases Out OpenAI and Anthropic Models in Copilot to Cut Costs

Microsoft. Microsoft is shifting Copilot features in products like Excel and Outlook away from third-party OpenAI and Anthropic models toward its own in-house MAI models to reduce inference costs. The company’s AI lead said the goal is to reduce and ultimately eliminate payments to Anthropic, with proprietary models handling a growing share of requests over time. The move highlights the margin pressure of building consumer products on top of external frontier APIs. Source

2. Report: Gemini 3.5 Pro Delayed to July 17 for an Architecture Rebuild

Google. A report relayed by BigGo says Google DeepMind has pushed Gemini 3.5 Pro’s release to July 17, abandoning the Gemini 2.5 Pro architecture for a ground-up redesign. The rebuilt model reportedly targets a 2-million-token context window and a “Deep Think” reasoning layer, with the delay attributed partly to staff turnover and performance-regression concerns. Google has not confirmed the timeline or details, so treat the specifics as unverified. Source

3. China Weighs Export Curbs on Its Top AI Models

Policy. Beijing is reportedly considering a tiered access system that would require registration for basic open-source tools, security review for advanced models, and domestic-only use for frontier models. Such curbs would threaten European firms that lean on cheap Chinese alternatives like Alibaba’s Qwen and ByteDance’s Doubao, potentially forcing a costlier shift back to US providers. The proposal mirrors the export-control logic the US has applied to chips, now aimed at model weights. Source

4. Chinese AI Models Now Regularly Exceed 30 Percent of OpenRouter Traffic

Open source. Models from DeepSeek and Z.ai have held over 30 percent of OpenRouter traffic every week since February 8, peaking at 46 percent, despite trailing US frontier models by roughly six to nine months on coding, reasoning, and cybersecurity benchmarks. The adoption is driven by a 60 to 90 percent cost advantage over OpenAI and Anthropic offerings. The data points to price, not raw capability, as the decisive factor for a large slice of production traffic. Source

5. DeepSeek Is Designing Its Own AI Inference Chip

DeepSeek. DeepSeek has been quietly hiring chip engineers to develop a proprietary inference chip, aiming to reduce dependence on Nvidia and Huawei hardware amid US export controls. The effort is still early stage and runs alongside a reported $7 billion funding round the company is pursuing. A homegrown accelerator would give DeepSeek more control over cost and supply as it scales serving for its open-weight models. Source

6. OpenAI and Anthropic Hand Out Millions in Free Compute to Lock In Startups

AI industry. Both labs are competing for early-stage developer loyalty through large compute-credit grants, with individual founders reporting offers worth over $3 million and Y Combinator startups receiving $500,000 to $1.5 million each. Combined, the two companies could distribute up to $800 million in credits across YC’s roughly 800 annual companies. The giveaways strain margins ahead of both companies’ planned public offerings. Source

7. Analysis: Open-Source AI Isn’t Hurting Anthropic Yet

AI industry. Using Vercel and OpenRouter data, a TechCrunch analysis argues that frontier and open-source models occupy different lifecycle stages: expensive frontier models prove out new use cases that later migrate to cheaper open alternatives, while fresh use cases keep emerging to sustain frontier demand. Despite open models capturing more raw token volume, Anthropic still accounts for over half of total spending due to far higher per-token pricing. The framing complicates the common narrative that open weights are eroding frontier labs’ revenue. Source

8. Apollo Economist Warns AI Profit Gains Outside Tech Could Lag Wall Street

AI industry. Apollo chief economist Torsten Slok argues that AI-driven productivity gains in regulated, non-tech industries face structural delays from process overhauls and privacy constraints, pushing realized profits well beyond current market timelines. He notes that even where knowledge-work gains occur, they are hard to measure and often get absorbed into operations rather than showing up on balance sheets. The caution lands as valuations increasingly price in near-term, economy-wide AI returns. Source

9. xAI Fully Merges Into SpaceX and Rebrands as SpaceXAI

xAI. xAI has completed its merger into SpaceX, unveiling a new “SpaceXAI” logo and brand and finishing a consolidation first announced in May following SpaceX’s February acquisition of xAI. The combined entity is pursuing orbital data center infrastructure, including an FCC filing for a satellite constellation intended for space-based AI compute. The rebrand formally ties Grok’s development to SpaceX’s launch and hardware capacity. Source

10. Even Realities Raises $150M and Hits Unicorn Status on Camera-Free Smart Glasses

Funding. Shenzhen-based Even Realities raised $150 million in a pre-Series B round led by Meituan with Tencent participating, valuing the camera-free AI smart glasses maker at $1 billion. Founded by an ex-Apple Watch and iPhone engineer, the company plans to fund its next-generation waveguide-optics platform and global expansion. The camera-free design is positioned as a privacy-forward alternative to camera-equipped rivals. Source

11. Bespoke Labs Raises $40M for AI Agent Training and Evaluation Infrastructure

Funding. San Francisco-based Bespoke Labs raised a $40 million Series A led by Wing VC to build simulation and synthetic-data infrastructure for training and evaluating autonomous agents before production deployment. Rather than building foundation models, the company focuses on tooling that helps enterprises measure agent performance and catch failure modes pre-launch. The raise reflects growing investment in the evaluation layer of the agent stack. Source

12. Savi Launches an App to Detect AI Voice-Cloning and Deepfake Scam Calls

Security. Savi Security, founded by former Cisco, Splunk, Apple, and Spotify executives, launched an iOS and Android app with real-time call monitoring to detect AI-generated scam calls, including deepfake “kidnapping” ransom scams, after raising $7 million in seed funding. The founders were motivated after their mother nearly fell victim to a voice-cloning scam. The app costs $8 per month or $63 per year for a family plan. Source

13. Station F Launches a Second Cohort of Its European AI Startup Accelerator

Funding. Paris-based Station F is launching a second F/ai accelerator cohort in September, aiming to help startups reach 1 million euros in revenue within six months by connecting them to partners including Anthropic, Google, Microsoft, and OpenAI. Its first cohort collectively raised $34 million. The program signals a maturing European AI startup pipeline anchored to the major model providers. Source