AI News Weekly Summary — April 20–26, 2026 | GPT-5.5, Google+Anthropic $40B, DeepSeek V4

OpenAI returns to the top with GPT-5.5 — its first fully retrained foundation model since GPT-4.5. Google commits up to $40B to Anthropic and 5 GW of compute. DeepSeek V4 cuts frontier inference cost six-fold. AI news weekly recap for April 20–26, 2026.

1. OpenAI GPT-5.5 “Spud” — First New Foundation Model Since GPT-4.5

On April 23, OpenAI shipped GPT-5.5 — its first fully retrained base model since GPT-4.5. Every model in between was an incremental update on the same architectural foundation. Internal codename: “Spud”.

GPT-5.5 scores 60 on the Artificial Analysis Intelligence Index — a record, returning OpenAI to the top spot. On OSWorld-Verified (autonomous OS navigation and UI clicking) it hits 78.7%, on Terminal-Bench 2.0 — 82.7%. The model is natively omnimodal (text, image, audio, video in one architecture), with a 1M-token context window and 74% on MRCR v2 at 512K–1M lengths (GPT-5.4: 36.6%).

API pricing: $5 per million input tokens, $30 per million output. GPT-5.5 Pro: $30 / $180. Available in ChatGPT from April 23, in the API from April 24.

OpenAI is back on top after six weeks — GPT-5.4 shipped in March, GPT-5.5 is a new foundation. A week after Anthropic’s public concession (Opus 4.7 trails Mythos — see our April 13–19 recap), OpenAI counters with a new generation. The “new frontier model every quarter” cycle holds.

2. Google Commits Up to $40B to Anthropic, Plus 5 GW of Compute

On April 24, Google announced its largest investment in Anthropic to date: $10B now plus up to $30B more tied to milestone achievement. Anthropic’s valuation confirmed at the February level — $350B (pre-fresh-rounds). Part cash, part Google Cloud compute — an additional 5 GW of compute spread over five years, starting in 2027.

Anthropic’s statement explicitly cites “inevitable strain” on infrastructure from enterprise and developer demand. Competitive context: Claude is outselling Gemini in the enterprise segment — exactly where Google needs growth most.

Google goes all-in on a “two-engine” strategy — its own Gemini plus a stake in Anthropic. Classic Cloud Wars move: when you don’t know who’ll win, back both. For Anthropic — relief from a compute ceiling that was starting to throttle deployment velocity in Q1.

3. Amazon Adds Up to $25B to Anthropic — and $100B From Anthropic Into AWS

On April 20, four days before Google, Amazon and Anthropic announced an expanded partnership: Amazon invests up to $25B more on top of the existing $8B (part upfront, the rest milestone-tied). The other direction: Anthropic commits to spend over $100B on AWS services across the decade and reserve 5 GW of new capacity for training and serving Claude.

Combined with Google’s announcement (point 2) — Anthropic now has commitments for at least 10 GW of additional compute over five years. Comparable scale to all of OpenAI.

The lesson of the week: in 2026, AI model valuation equals access to power and silicon. Anthropic plays both clouds, OpenAI plays Microsoft and Stargate, DeepSeek plays Huawei. Whoever runs out of GPUs first loses.

4. DeepSeek V4 — Frontier-Tier at One-Sixth the Price of GPT-5.5 and Opus 4.7

On April 24, DeepSeek released previews of two new models: V4-Pro (1.6T total parameters, 49B active, trained on 33T tokens) and V4-Flash (284B total, 13B active, 32T tokens). Both open-source, with native 1M-token context windows.

V4-Pro hits 80.6% on SWE-bench Verified — within 0.2 points of Claude Opus 4.6. On Terminal-Bench 2.0 it leads Claude (67.9% vs 65.4%), and on LiveCodeBench (93.5% vs 88.8%). Pricing: $3.48 per million output tokens for V4-Pro (vs GPT-5.5’s $30, Opus 4.7’s $25). V4-Flash is even cheaper: $0.28 per million output. Tight integration with Huawei chips — DeepSeek is showing that non-Nvidia compute works for frontier models.

Six-fold price difference. If V4-Pro were 5–10% weaker, you could ignore it. But it’s at frontier-tier. This isn’t a problem for Anthropic and OpenAI in the premium segment, but for 80% of cost-driven production workloads — DeepSeek becomes the default. Chip embargoes failed again.

5. Tesla Q1 2026 — Beat on Earnings, CapEx Jumps to $25B

On April 22, Tesla reported Q1 2026: EPS $0.41 vs $0.37 expected. Revenue $22.4B (+16% YoY) — below consensus. Stock initially +4% after-hours, then sold off when management announced 2026 CapEx rising to $25B (prior guide: $20B). Reason: aggressive investment in Optimus humanoids, Cybercab, and AI.

In the background: April 18 saw Tesla launch unsupervised robotaxi in Dallas and Houston. Three active cities (Austin, Dallas, Houston), Cybercab production starting at the Austin plant in late April. Dan Ives (Wedbush): “Tesla is now more an AI company than a car company.”

Bear case: $25B CapEx, auto sales stalling, margins compressing. Bull case: Cybercab and Optimus are a trillion-dollar platform. Investors are giving back this year’s 14% drawdown — and TSLA still trades at a higher P/E than the entire Magnificent 7 combined.

6. Google Cloud Next ’26 — Gemini Enterprise Agent Platform, Veo 3.1 to 4K

Google Cloud Next ’26 (April 22–24, Las Vegas) brought the Gemini Enterprise Agent Platform — a unified umbrella for Vertex AI plus tools for building, scaling, and monitoring agents. Includes Gemini Deep Research Agent in preview: autonomous planning, execution, and synthesis for multi-step research tasks.

Veo 3.1 received an update with 4K output, native vertical format for Ingredients-to-Video, and upscaling to 1080p and 4K. Accenture and Google Cloud announced the Gemini Enterprise Acceleration Program — an accelerated rollout track for large enterprises.

7. Moonshot Kimi K2.6 — 256K Context, 300 Parallel Sub-Agents, Open-Source SOTA

April 20–21: Chinese Moonshot AI released Kimi K2.6 — a 1T-parameter MoE (32B active, 384 experts), with native multimodality and INT4 quantization. Day-0 support in vLLM, OpenRouter, Cloudflare Workers AI, MLX, and OpenCode.

SOTA among open-source models: HLE w/ tools 54.0, SWE-Bench Pro 58.6, SWE-bench Multilingual 76.7, BrowseComp 83.2. Key feature: long-horizon execution — 4,000+ tool calls, 12+ hours of continuous operation, 300 parallel sub-agents.

8. Alibaba Qwen 3.6 Max Preview — Flagship for Agentic Coding

In parallel with Kimi K2.6, Alibaba shipped Qwen 3.6 Max Preview — an early preview of the next flagship focused on agentic coding and knowledge reliability. The model solved AIME 2026 #15 after roughly 30 minutes of thinking. Qwen 3.6 Plus (smaller variant) reached #7 on Code Arena.

Together with Kimi K2.6, DeepSeek V4, and last week’s Qwen 3.6-35B-A3B — that’s the fourth Chinese frontier model in 14 days.

9. Bezos’s Project Prometheus Closes $10B at $38B Valuation

On April 23, Bloomberg confirmed: Project Prometheus — Jeff Bezos’s AI lab led by Vik Bajaj (ex-Google) — closed a $10B round at a $38B valuation. Among participants: JPMorgan Chase and BlackRock.

Prometheus remains tight-lipped about product, but signals focus on “physical AI” — models for design in the physical world (absent from the current LLM wave). Bezos co-leads the lab personally.

10. ChatGPT Images 2.0 — Major Image Generator Upgrade

On April 21, OpenAI released ChatGPT Images 2.0 — the next generation of its image generation tool. Key improvements: composition control precision, multilingual text rendering inside images, visual fidelity, and flexible aspect ratios (panorama, portrait, square) without quality loss.

Direct pressure on Midjourney and Black Forest Labs (FLUX). Free tier in ChatGPT with limits, separate API endpoint.

11. Apple WWDC 2026 Teaser — Heralds a Siri Overhaul in iOS 27

April 19–20: Apple published its WWDC 2026 promo graphic (event runs June 8–12). The glow around the “26” numerals is read by Mark Gurman (Bloomberg) as a direct preview of the redesigned Siri interface in iOS 27.

Expected features: Siri integrated with Dynamic Island, an “Ask Siri” option in third-party app context menus, “Write with Siri” in the keyboard, Spotlight replaced by Siri as the primary iPhone search interface. Underlying: an expanded Google partnership — next-gen Apple Foundation Models built on Gemini, with on-device and Private Cloud Compute processing.

12. Vercel — Data Leak Through a Third-Party AI Tool’s OAuth

On April 19, Vercel disclosed an incident: an attacker compromised the Google Workspace OAuth application of Context.ai (a third-party AI analytics and productivity tool) used by Vercel teams. Through the hijacked OAuth token, the attacker reached environment variables and API keys in internal systems.

Continuation of the AI supply chain thread from previous weeks — earlier Mercor (via LiteLLM), now Vercel (via Context.ai). Pattern: don’t attack the target directly, attack through a trusted AI-tool dependency.

13. UKG, Snap, Meta and Microsoft — AI-Driven Layoffs Accelerate

The full week brought a wave of explicitly AI-motivated layoff announcements: UKG (Blackstone) cuts 950 jobs on April 21, Snap cuts 1,000 roles (after an activist investor flagged “over-hiring”) on April 16, Meta and Microsoft flag a combined 20,000 cuts over the year. Cumulative tech-sector cuts since January through mid-April 2026: over 150,000 jobs across 500+ companies.

According to Nikkei Asia, 47.9% of Q1 2026 cuts are tied to AI automation. IBM moves the other way: 3x growth in entry-level hiring — because “AI can do a lot, but still needs a human touch.”

Conflicting signals. Some firms cut directly under automation cover, others raise junior headcount to oversee model output. The consensus: this stretch of 2026 is to the labor market what 2008–2010 was to banking — a structural shift, not cyclical.

14. Cursor and Cognition — Valuations Doubling on Monthly Cycles

On April 19, CNBC: Cursor (AI coding assistant) is in talks for a $2B round at over $50B valuation. On April 23, Bloomberg: Cognition AI (the Devin agent maker) is negotiating a round that would double its valuation to $25B.

Q1 2026 globally: $300B in venture capital into startups, of which $242B into AI (80% of the entire market). Four of the five largest rounds in history closed in Q1 2026: OpenAI $122B, Anthropic $30B, xAI $20B, Waymo $16B.

15. European Commission — €63.2M for AI in Health and Online Safety

On April 21, the European Commission released €63.2M in support programs for AI innovation in health and online safety. In parallel: the “Digital Omnibus” proposed in November 2025 is still in the legislative process — it may delay application of select AI Act requirements to 2027–2028.

Hard deadline unchanged: by August 2, 2026, every Member State must have at least one AI regulatory sandbox (AI Act, Article 57).

16. Microsoft Commits $18B to AI Infrastructure in Australia

On April 23, Microsoft announced a A$25B (~$18B) investment in Australia’s digital infrastructure. Scope: new AI data centers, cybersecurity programs, and Azure expansion for AI services across Asia-Pacific.

The same day, Meta signed a deal with Amazon for millions of AWS Graviton CPUs for its AI data centers — another move to diversify away from Nvidia. Signal of the week: hyperscalers are no longer just buying GPUs, they’re also building their own CPU clusters for inference workloads.

17. In Brief

• GPT-5.5 Pro in the API — available to developers from April 24 alongside standard GPT-5.5 ($30 / $180 per million tokens).
• Anthropic — RCE flaw in MCP — The Hacker News disclosed a Model Context Protocol vulnerability allowing remote code execution on servers running vulnerable implementations. Scope: 7,000 publicly accessible servers; packages totaling over 150M downloads.
• Cooley US state AI laws update — On April 24, Cooley published its review of state-level AI legislation: over 600 active bills. Concentrated in California, Texas, and New York.
• Fitch Ratings — In an April 20 report, the agency warns that fast AI rollouts on the cybersecurity defense side may open new vulnerabilities in the short term, despite long-term benefits.
• Tesla Cybercab — production starts at the Austin plant in late April. First series-produced Tesla without a steering wheel or pedals, two seats.

Weekly Takeaway

The week of April 20–26, 2026 was defined by three moves. First — the model race is back on the front page: GPT-5.5 retakes the crown, DeepSeek V4 undercuts the entire Western frontier on price, Moonshot Kimi K2.6 and Qwen 3.6 Max push the open-source ceiling. The “new frontier every quarter” cycle is holding.

Second — money and power. Anthropic in a single week pulled in commitments totaling up to $65B (Google up to $40B + Amazon up to $25B) and at least 10 GW of compute. Bezos’s Project Prometheus closed $10B. Cursor and Cognition are negotiating rounds that double valuations on monthly cycles. Q1 2026 globally: $242B into AI startups. Microsoft adds $18B in Australia, Meta orders millions of Graviton CPUs from Amazon. This is the hardware-and-power round of the game, not the software round.

Third — the social cost is becoming visible. UKG, Snap, Meta, and Microsoft cut a combined 22,000+ jobs in one week, with 47.9% of Q1 cuts directly tied to AI. Vercel demonstrates that every external AI tool is a potential supply-chain vulnerability. Apple is positioning a chatbot-style Siri rebuild for June, powered by Gemini under the hood — desktops and phones are no longer app platforms; they’re agent surfaces.

Źródła / Sources: OpenAI Blog, Anthropic Blog, Google Blog, Bloomberg, CNBC, TechCrunch, The Information, VentureBeat, Reuters, Fortune, Axios, Tom’s Hardware, MarkTechPost, Artificial Analysis, DeepSeek API Docs, Nikkei Asia, Crunchbase, Wedbush Securities, Electrek, Microsoft Blog, Data Center Dynamics, MacRumors, Glass Almanac, Boston Dynamics, Stat News, Cybernewscentre, European Commission, Cooley LLP