Technology

DeepSeek released V4-Pro on April 24, 2026, a 1.6 trillion-parameter open-weight model that leads all open-weight competitors on math and coding but trails Claude Fable 5 by 43 ELO points on the LMArena leaderboard. Neither DeepSeek nor its parent High-Flyer is currently on the US Bureau of Industry and Security Entity List, though both were flagged internally as security risks; the operative export-control constraint remains chip-level, with V4-Flash confirmed to have used Huawei Ascend hardware for part of its training. A rumored V5 successor has no official documentation as of July 2026, and the legacy API aliases are not retired until July 24, 2026, indicating the V4 transition is still in progress.

On 2 July 2026, the Court of Justice of the European Union dismissed Google and Alphabet's appeal in Case C-738/22 P, making the EUR4.125 billion Android antitrust penalty the largest single EU competition fine ever imposed on one company. The ruling confirms three prohibited practices — tying Play Store licenses to Google Search and Chrome pre-installation, anti-fragmentation agreements blocking unauthorized Android forks, and a revised revenue-share scheme — but introduces no new operational orders, leaving the unbundled paid-licensing model Google adopted in October 2018 structurally unchanged. Separately, the US Department of Justice ad-tech case against Google remains in its remedy phase, with closing arguments completed in November 2025 and a potential divestiture of AdX and DoubleClick for Publishers still pending before Judge Brinkema.

At Build 2026, Microsoft released a first-party family of seven MAI models covering reasoning, coding, image generation, transcription, and speech — all built without distillation from OpenAI or other labs. The clearest capability result is MAI-Code-1-Flash, which scored 51.2% on SWE-Bench Pro versus Claude Haiku 4.5's 35.2% at $0.75 per million input tokens, while MAI-Thinking-1 sits in the competitive middle of reasoning benchmarks rather than at the top of public leaderboards. The strategic core is Frontier Tuning, Microsoft's customization layer, which the company reports delivered GPT-5.5-level quality at one-tenth the cost on a McKinsey deployment — though those figures come from Microsoft's own disclosures and have not been independently audited.