AI News Roundup — May 26, 2026
Three explosive stories landed in a 72-hour window: leaked Claude Mythos launch code suggests an imminent public release, DeepSeek made its 75% V4-Pro price cut permanent (escalating the AI price war), and US spy agencies were found still using Claude despite the government's own blacklisting of Anthropic. Microsoft raised its 2026 AI capex to $190B and Amazon's custom silicon hit a $20B run rate.
Leaked Code Suggests Claude Mythos Is Launching Sooner Than Expected
Code preparations for Claude Mythos 1 surfaced in public GitHub repositories on May 22–24, including a reasoning protocol nicknamed 'Strict Write Discipline,' Mythos Router memory schemas, and references to a production Opus 4.7 thinking variant — strongly suggesting the public launch window is much narrower than Anthropic has indicated. The leak is the third Mythos-related disclosure in two months, following a 3,000-file internal CMS leak in March and the Claude Code source map leak that followed days later.
DeepSeek Makes 75% V4-Pro Price Cut Permanent — AI Price War Escalates
DeepSeek announced on May 24 that its temporary 75% discount on V4-Pro tokens is now permanent, making frontier-class Chinese AI reasoning available at roughly one-quarter the cost of equivalent OpenAI or Anthropic tokens. Chinese models already represent 60% of OpenRouter usage. Analysts expect quiet pricing responses from Anthropic, OpenAI, and Google within 30 days. DeepSeek also cut input cache-hit prices across its entire API lineup to one-tenth of the original price.
US Spy Agencies Are Still Using Claude Despite Government's Own Blacklist
US intelligence agencies are continuing to use Anthropic's Claude in active operations — directly contradicting the Pentagon's designation of Anthropic as a national security concern and its ongoing evaluation of Claude replacements. The contradiction reflects classic bureaucratic hedging: regulators warn while operators integrate. The NSA's use of Mythos despite the blacklist was confirmed by Axios and The Next Web.
Claude Code Comes to the Web — No More Terminal Required
Anthropic shipped Claude Code to the browser and iPhone, moving it beyond the terminal for the first time. Agents that previously required a physically open laptop can now run from anywhere. The same week, OpenAI Codex added the ability to control a locked Mac — both moves signal that long-running AI coding agents are becoming untethered from specific machine states.
Microsoft Raises 2026 AI CapEx to $190 Billion
Microsoft now projects $190 billion in 2026 capital expenditure — with $25 billion of the increase driven by surging memory and storage component prices from AI infrastructure demand. Microsoft has spent $97 billion over the last four quarters and generates $37 billion in AI ARR. Wall Street is increasingly focused on the widening gap between AI investment and AI returns.
Amazon Custom Silicon Hits $20 Billion Annual Run Rate
Amazon CEO Andy Jassy disclosed that Amazon's custom silicon business — Graviton, Trainium, and Nitro combined — now exceeds $20 billion in annual run rate, up 100% year-over-year. That makes AWS the third-largest data center chip business in the world behind only Nvidia and AMD, and gives builders a credible alternative to Nvidia for AI inference at scale.
China Tightens Travel Restrictions on Top AI Talent
China is expanding travel curbs targeting its top AI professionals, restricting overseas travel for researchers and engineers at leading AI labs and chipmakers. Bloomberg reports the restrictions are part of a broader effort to prevent knowledge transfer to Western competitors and to retain AI talent inside China's domestic ecosystem — a direct mirror of US export controls on chips.