xAI released Grok 4.5 on July 8, 2026, positioning it as the company’s first model designed specifically for software development and agentic workloads (Axios; xAI). Built in partnership with Cursor, the model carries an API price of $2 per million input tokens and $6 per million output tokens - making it the lowest-cost entry among models placing in the top tier of the Artificial Analysis Intelligence Index, where Grok 4.5 ranks fourth. The pricing gap matters for practitioners running sustained agentic coding pipelines, where per-session token budgets compound quickly against higher-priced alternatives. European availability is withheld at launch; xAI is targeting mid-July 2026 for EU access pending mandatory model evaluations, adversarial testing, and cybersecurity assessments required under the EU AI Act’s framework for general-purpose AI systems at the systemic-risk compute threshold (Heise Online).
Reports published July 7-8 by SiliconAngle and Android Headlines revealed that DeepSeek is developing a custom chip aimed at AI inference rather than training, with the goal of reducing its operational reliance on hardware from Nvidia and Huawei (SiliconAngle; Android Headlines). The move follows a pattern seen at several AI labs of bringing inference silicon in-house to reduce serving costs and improve latency control at scale. For DeepSeek specifically, the path is constrained: U.S. export controls restrict Chinese chip designers’ access to the most advanced overseas foundries and to high-bandwidth memory components central to inference workloads, which will shape both the manufacturing options and the memory architecture of any resulting chip. The effort remains early-stage - the company has expanded its internal chip-design team and opened discussions with foundry partners, but has not disclosed a named manufacturer, prototype specifications, or a production timeline.
Mistral AI confirmed that a new open-weight mixture-of-experts model family is entering partner early access in July 2026, with initial availability limited to research institutions, government customers, and select industry partners (TechTimes). CEO Arthur Mensch described the architecture as “fat but sparse” - a design that keeps total parameter count large while routing each token through only a small fraction of active sub-networks, holding per-token compute well below what the aggregate model size would imply and making it viable to self-host at a lower active-memory footprint than comparable dense models. No benchmark results, parameter counts, or license terms have been officially published ahead of the planned broader open-weight release, expected later this summer.