← Back to Blog AI News EN

AI Monthly Recap April 2026 — Mythos, GPT-5.5, $65B Anthropic, Manus Blocked

Anthropic in 30 days locks Mythos, ships Opus 4.7, secures up to $65B and 10 GW from Google and Amazon, kept out of part of the Pentagon contracts. OpenAI returns with GPT-5.5. Four Chinese frontiers in 14 days. Big Tech 2026 CapEx tracking ~$705–735B.

Wieże CapEx — pieniądze i krzem, kwiecień 2026

April 2026 was the month the model race exited the “who’s first” phase and entered the “who pays for power” phase. Anthropic in 30 days locked Mythos away, shipped Opus 4.7, and secured commitments of up to $65B plus 10 GW of compute from Google and Amazon.

OpenAI returned to the top with GPT-5.5. Four Chinese frontier models in 14 days. Big Tech raised combined 2026 CapEx to roughly $705–735B based on cumulative post-Q1 guidance. AI monthly recap April 2026.

Thread 1 — The Model Race: From Mythos to GPT-5.5

April delivered the densest frontier model release wave in history:

  • April 7 — Anthropic announces Claude Mythos and does NOT release it publicly. The model autonomously discovers thousands of vulnerabilities, including decades-old ones, and converts many into working exploits. It goes to a closed defensive consortium, Project Glasswing (40+ firms: Apple, Amazon, Microsoft, Google, Cisco, Broadcom, Linux Foundation).
  • April 14 — OpenAI ships GPT-5.4-Cyber for defensive cybersecurity work, gated through the TAC (Trusted Access for Cyber) program for verified defenders only.
  • April 14Alibaba Qwen 3.6-35B-A3BMoE (Mixture-of-Experts) model with 35B total parameters and 3B active. Score: 73.4% on SWE-bench Verified.
  • April 16 — Anthropic publicly releases Claude Opus 4.7 and openly admits it doesn’t match Mythos. +13% on coding benchmarks, image input up to 3.75 megapixels.
  • April 17Alibaba Qwen 3.6 Max Preview — early preview of Alibaba’s next flagship, focused on agentic coding.
  • April 20Moonshot Kimi K2.6 — 1T-parameter MoE, 256K context, 300 parallel sub-agents.
  • April 23OpenAI GPT-5.5 “Spud” — first fully retrained base model since GPT-4.5. 60 on the Artificial Analysis Intelligence Index (a record), 78.7% on OSWorld-Verified, omnimodal, 1M-token context.
  • April 24DeepSeek V4-Pro and V4-Flash — open-source. V4-Pro 1.6T params, 80.6% on SWE-bench Verified at $3.48 per million output tokens (roughly 7–9x cheaper than GPT-5.5 and Opus 4.7).

Total: seven frontier models in 18 days, four Chinese. The release tempo shifted from quarterly to weekly. A clearer specialization is emerging: locked (Mythos, GPT-5.4-Cyber for cyber), public (Opus 4.7, GPT-5.5), cheap open-source (DeepSeek V4, Kimi K2.6).

Thread 2 — Money and Power: $65B and 10 GW for Anthropic

The biggest stream of the month was compute, not models. Anthropic in a single week (April 20–24) pulled in commitments totaling up to $65B plus at least 10 GW of compute:

  • April 20Amazon — additional up to $25B (on top of the prior $8B) plus 5 GW on AWS. The other direction: Anthropic commits to spend over $100B on AWS services across the decade.
  • April 24Google — up to $40B ($10B now, $30B milestone-tied) plus 5 GW on Google Cloud. Valuation confirmed: $350B.

Other major capital moves in April:

  • Project Prometheus (Bezos) closes $10B at $38B valuation (April 23, JPMorgan and BlackRock).
  • Cursor in talks for a $2B round at $50B+ valuation.
  • Cognition AI in talks for a round doubling valuation to $25B.
  • Microsoft +$18B in Australia (April 23, digital and AI infrastructure).
  • Q1 2026 globally: $300B venture capital into startups, of which $242B into AI (80% of the market).

By month-end, post Big Tech earnings: combined 2026 CapEx for Magnificent 5 (Microsoft, Amazon, Alphabet, Meta, Apple — excluding NVIDIA and Tesla) stands at roughly $705–735B. Microsoft $190B, Amazon ~$200B, Alphabet $180–190B, Meta $125–145B, Apple ~$10B. Comparable in scale to the entire US defense budget ($824B in 2025).

Thread 3 — Big Tech Earnings: AI Delivers, CapEx Climbs

The two-day Big Tech “earnings superweek” (April 29–30) showed AI is no longer just an investment line:

  • Microsoft fiscal Q3: revenue $82.9B (+18%), Azure +40%, AI business at $37B annual run rate (+123% YoY), 2026 CapEx $190B.
  • Alphabet Q1: $110B revenue (+22%), Google Cloud +63% to $20B quarterly. Stock +7% after-hours. CapEx raised to $180–190B.
  • Meta Q1: revenue +33%, but CapEx raised to $125–145B. Stock −6% after-hours.
  • Amazon Q1: AWS fastest growth in 15 quarters. CapEx held at ~$200B.
  • Apple fiscal Q2 (April 30): $111.2B revenue (+17%), R&D +33% to $11.42B. Apple alone is not scaling AI infrastructure CapEx — holding at roughly $10B and instead, per Bloomberg/Reuters reports, Gemini is set to power the new Siri in iOS 27.

Signal: Apple chose a different model — buy intelligence from Google rather than build its own data centers. The other four are going all in on infrastructure spending.

Thread 4 — Borders and Conflicts: Pentagon, China, OpenAI

Three big confrontations of the month:

Pentagon vs Anthropic (April 29) — The US Department of Defense, also using the Department of War name, signs deals with eight AI firms for classified networks: SpaceX, OpenAI, Google, NVIDIA, Reflection AI, Microsoft, AWS, and Oracle. Anthropic is kept off the list — refused “unrestricted access” to Claude for all lawful military uses. The price of safety policy: no billion-dollar government contracts.

China blocks Meta-Manus (April 27) — China’s NDRC blocks Meta’s $2B Manus acquisition. The first hard intervention in a strategic AI deal by Beijing. Signal: Chinese AI tech is under government control even when the company is formally Singaporean.

OpenAI misses targets (April 28) — WSJ reports: OpenAI didn’t hit its 1B weekly active users target for ChatGPT or revenue goals in 2025. Per WSJ and Reuters, CFO Sarah Friar reportedly signaled tensions over compute contracts. Oracle ($300B OpenAI compute contract), NVIDIA, and AMD shares drop. OpenAI is losing consumer share to Gemini and ground in enterprise/coding to Claude — while planning an IPO.

Thread 5 — Robotaxi and Robotics: Tesla, Waymo, Lucid, Norway

April also accelerated transport autonomy:

  • April 14Lucid + Uber expand robotaxi fleet to 35,000 vehicles plus $750M new capital ($550M Ayar/PIF, $200M Uber).
  • April 14Norway issues its first-ever permit for an autonomous bus without a safety driver (Stavanger).
  • April 18Tesla launches unsupervised robotaxi in Dallas and Houston. Three active cities including Austin.
  • April 22 — Tesla pushes back launch in 5 additional cities to H2 2026.
  • April 22 — Tesla Q1: 2026 CapEx raised to $25B for Optimus, Cybercab, and AI. Cybercab production starts in Austin in late April.
  • April 27 — Tesla Robotaxi app on Android (a year after iOS).
  • May 1 — Meta acquires Assured Robot Intelligence, strengthening robotics and humanoid efforts.

Waymo remains the leader — already in 11 cities plus 20 announced. In Dallas and Houston Waymo arrived two months before Tesla.

Thread 6 — Cybersecurity: CyberStrikeAI, Vercel, MCP RCE

Three big AI security stories in April:

  • CyberStrikeAI — an AI-assisted campaign compromised 600+ FortiGate devices across 55 countries (March reporting, continuing in April).
  • MCP RCE (mid-April) — a vulnerability in Anthropic’s MCP (Model Context Protocol) enables remote code execution (RCE) on 7,000 publicly accessible MCP servers. MCP-related packages have over 150M cumulative downloads.
  • Vercel (April 19) — leak via a hijacked Context.ai OAuth token (a third-party AI analytics tool). Pattern: don’t attack the target directly, attack through a trusted AI-tool dependency.

April 15: IBM unveils a toolkit for detecting and countering agentic attacks. Darktrace report: 92% of security teams are concerned about how AI agents affect organizational security. April 28: OpenAI and Anthropic met in closed-door briefings with the House Homeland Security Committee — Congress is preparing federal cyber-AI regulation.

Thread 7 — Labor Market: AI-Driven Layoffs

April accelerated structural employment restructuring:

  • Per industry trackers (Tom’s Hardware, Layoffs.fyi) — over 150,000 tech-sector layoffs from January through mid-April 2026 across hundreds of companies.
  • Nikkei Asia: 47.9% of Q1 2026 cuts are tied to AI automation.
  • Meta announces an 8,000-person layoff wave for May 20 (April 17).
  • Snap cuts 1,000 roles (April 16).
  • UKG (Blackstone): 950 layoffs (April 21).
  • Meta + Microsoft: 20,000 cuts announced in one week (April 24).
  • The opposite direction: IBM tripled entry-level hiring — because “AI can do a lot, but still needs a human touch.”

Thread 8 — Tools and Platforms

Key non-model April releases:

  • April 17Anthropic Claude Design (research preview) — UI prototypes, slides, and one-pagers from chat, on Pro, Max, Team, and Enterprise subscriptions.
  • April 18xAI Grok Voice APIs (STT — speech-to-text $0.10/hour, TTS — text-to-speech $4.20 per million characters) — direct shot at ElevenLabs and Deepgram.
  • April 21ChatGPT Images 2.0 — major generator upgrade, pressure on Midjourney and Black Forest Labs.
  • April 22–24Google Cloud Next ’26: Gemini Enterprise Agent Platform, Gemini Deep Research Agent, Veo 3.1 with 4K output.
  • May 1Microsoft Agent 365 reaches GA (general availability): observe/govern/secure for an organization’s AI agent fleet, $15 per user per month, registry synchronization with AWS Bedrock and Google Cloud in preview.
  • April 19 — Apple teases the Siri redesign in iOS 27 — per Bloomberg/Reuters reports, expected to be powered by Gemini (premiere at WWDC June 8–12).

Thread 9 — Regulation and Policy

  • US states — over 600 active AI bills in 2026 sessions (Cooley, April 24). Sen. Marsha Blackburn introduces “TRUMP AMERICA AI Act” — a bill to preempt state laws conflicting with federal policy.
  • European Commission — April 21 releases €63.2M for AI in health and online safety. The “Digital Omnibus” may delay select AI Act requirements to 2027–2028.
  • Hard deadline from the AI Act: by August 2, 2026, every EU Member State must have at least one AI regulatory sandbox (Article 57).

Full Weekly Deep-Dives — April 2026

Top Takeaways of the Month

First — the model race has moved to a weekly cycle. Seven frontier models in 18 days, four Chinese. The “GPT-4 reigns for a year” era is over; the “who ships on Wednesday” era has begun. DeepSeek V4 undercuts Western frontier pricing roughly 7–9x at comparable quality. Chip embargoes haven’t slowed the pace of Chinese releases — V4 ships with tight Huawei silicon integration.

Second — AI model valuation equals access to power and silicon. Anthropic pulled up to $65B and 10 GW in a single week. Big Tech raised 2026 CapEx to roughly $705–735B. OpenAI is missing growth targets but is still buying everything Nvidia ships. The 2026 question isn’t “who has the best model,” it’s “who runs out of compute first.”

Third — borders are tightening. Anthropic says no to the Pentagon and loses contracts. China says no to Meta and blocks the Manus acquisition. Apple per reports is choosing Gemini over building its own infrastructure. The Pentagon picks eight firms for classified networks. The era of a single, open AI market is giving way to a bloc-divided world: a US market with Anthropic outside part of the military contract pool, a Chinese market closed to Meta, and a Europe under the AI Act — each with its own trust architecture.

Fourth — AI delivers real money, but also real social cost. Microsoft AI business at $37B annual run rate. Google Cloud +63%. AWS fastest in 15 quarters. The other side: over 150K tech-sector layoffs since January, 47.9% of Q1 cuts tied to AI. Apple chooses a third path — R&D over CapEx. This is shaping up to be one of the more important AI monetization tests in Big Tech.

Źródła / Sources:

MML Studio

Written by

MML Studio

Comments

Leave a comment

Comments are published after admin approval.

← Previous AI News Weekly Summary — April 20–26, 2026 | GPT-5.5, Google+Anthropic $40B, DeepSeek V4
Next → AI News Weekly Summary — April 27 – May 3, 2026 | Big Tech Q1, Manus Blocked, Pentagon