AI News Digest: April 2026 — the week's top stories

AI Summary: Over the week of April 2026, GPT-5.5 ($5/$30 per million tokens, 1M context), DeepSeek V4 (1.6T parameters, -73% compute), Kimi K2.6, and Qwen3.6 were released. Anthropic reached a $1 trillion valuation with investments from Google ($40 billion) + Amazon ($13 billion). Important legal precedent: AI chats are not protected by attorney-client privilege. AutomationBench (Zapier) showed that no model scored above 10% on real automation tasks.

April 2026 will go down in AI industry history. In a single week, three new flagships were released, Anthropic reached a $1 trillion valuation, and Google announced $40 billion in investments. At the same time, there were $150 billion lawsuits, a Vercel breach, and a landmark court ruling on AI chats. Let's go through everything in order.

Contents

1. New models: GPT-5.5, DeepSeek V4, and others
2. Money: Anthropic on the way to $1 trillion
3. Products and tools of the week
4. Scandals and legal precedents
5. Research and analysis
6. Summary table of models
7. Key takeaways of the week

New models: GPT-5.5, DeepSeek V4, and other releases

Key Takeaway: Over one week, four significant new models were released — GPT-5.5, DeepSeek V4 (Pro + Flash), Kimi K2.6, and Qwen3.6. The pace of releases in the second half of April 2026 is unprecedented.

GPT-5.5 — OpenAI's new flagship for agents

[Fact]: GPT-5.5 costs $5 per million input tokens and $30 per million output tokens, and supports a 1 million token context.

OpenAI released GPT-5.5 — a flagship model focused on agentic scenarios. GPT-5.5 is already rolling out to users on paid ChatGPT and Codex plans. The main emphasis is on "agency": the model does not just answer questions, but plans actions and checks results automatically.

Key features of GPT-5.5:

• Price: $5 / $30 per million tokens (input/output)
• Context: 1 million tokens
• Focus: autonomous agents, complex multi-step tasks
• High Terminal-Bench scores, improved token efficiency

Practical value: For developers of AI agents, this is OpenAI's most capable tool to date. At $5 per million tokens, it is more expensive than DeepSeek, but cheaper than OpenAI's previous flagships.

DeepSeek V4 — 1.6 trillion parameters at one-tenth the price

[Fact]: DeepSeek V4 uses 27% of the compute and 10% of the KV-cache memory compared with the previous V3.2 version at comparable quality.

DeepSeek unveiled V4 in two variants: Pro and Flash. This is the most technically significant release of the week in terms of efficiency.

Comparison of DeepSeek V4 versions:

DeepSeek V4 Pro benchmarks:

• Terminal-Bench 2.0: 67.9% (Claude Opus 4.6 — 65.4%)
• LiveCodeBench: 93.5% (Claude Opus 4.6 — 88.8%)
• SWE-bench Verified: 80.6% (Claude Opus 4.6 — 80.8%)

Limitations: no multimodality (text only), higher hallucination rate on factual questions, preview status. The old deepseek-chat and deepseek-reasoner APIs will be shut down on July 24, 2026.

GPT-Image-2 — image generation with a thinking mode

OpenAI introduced GPT-Image-2 — a new image flagship. Improvements: better text rendering in images, layout, multilingual support, and a thinking mode before generation.

Kimi K2.6 — an unexpected leader on SWE-bench

[Fact]: Kimi K2.6 from Moonshot outperforms GPT-5.4 and Claude Opus 4.6 on SWE-bench Multilingual and BrowseComp benchmarks.

Moonshot released Kimi K2.6 — a family of four models. For tasks involving coding in multiple programming languages and browser agents, it is worth testing.

Qwen — two new models in one week

Alibaba released:

• Qwen3.6-27B — the best dense model (not MoE) in its class
• Qwen3.6-Max-Preview — focused on agentic coding and instruction following

Grok Voice — voice agents in 25 languages

[Fact]: Grok Voice supports 25 languages including Russian, 5 voices, and tool calling with web/x/file_search.

xAI launched Grok Voice — a voice model for agents. xAI's first serious step toward voice interfaces with tool-calling support.

Money: Anthropic on the way to $1 trillion

Key Takeaway: In one week, Anthropic raised $53 billion in investments (Google + Amazon) and reached a $1 trillion valuation on the secondary market. This is the largest round in AI history.

[Fact]: Google is investing up to $40 billion in Anthropic, and Amazon added $13 billion. Anthropic's valuation on the secondary market (Forge) surpassed $1 trillion — higher than OpenAI's.

• Anthropic shares are trading at a $1T valuation
• Google is investing up to $40 billion
• Amazon added $13 billion

At the same time Jeff Bezos is raising ~ $10 billion for Project Prometheus — a startup focused on "Physical AI".

What this means: Anthropic's $1T valuation is a market signal of Claude platform's long-term competitiveness. For enterprise customers, it means investment stability in the platform over a 5+ year horizon.

Products and tools of the week

Key Takeaway: OpenAI, Google, and Anthropic simultaneously released enterprise products — organizational agents, corporate platforms, and integrations with work tools.

OpenAI Workspace Agents

Workspace Agents — an evolution of GPTs based on Codex in the cloud with sharing inside an organization.

ChatGPT for Clinicians and Excel

• ChatGPT for Clinicians — a specialized version for medical professionals
• ChatGPT for Excel and Google Sheets — a beta release of add-ons

Google: three major announcements

• AI Studio without an API key with Google AI Pro/Ultra subscriptions
• Workspace Intelligence — a semantic layer for Gemini agents over email, chats, and files
• Enterprise Agent Platform — enterprise agents
• Deep Research Max and open-source DESIGN.md

Inside Google — "a strike team" in DeepMind for coding: the trigger was Anthropic's releases.

Claude Cowork on third-party infrastructure

Claude Cowork became available on AWS Bedrock, MS Foundry, and Google Vertex AI. Non-Anthropic models can be connected.

Sakana Fugu and OpenAI Privacy Filter

• Fugu — Sakana AI's first commercial product: an LLM for orchestrating other LLMs
• A method for increasing creativity through a random string at the beginning of the prompt
• OpenAI Privacy Filter — an open model for PII detection

AutomationBench: the harsh reality of agents

[Fact]: According to AutomationBench (Zapier), none of the tested models passed the 10% mark on real-world automation tasks.

Zapier showed AutomationBench — the results ground the hype: automating real business processes is still extremely difficult.

Scandals and legal precedents

Key Takeaway: The AI industry faced three parallel crises: corporate, legal, and infrastructure-related.

WSJ: investors are discussing replacing OpenAI's CEO

According to the WSJ, a number of OpenAI investors are in talks about the possible removal of Sam Altman.

Musk's lawsuit against OpenAI for $150 billion

Elon Musk filed a lawsuit against OpenAI and Microsoft for $150 billion. At the same time launched XChat — an end-to-end encrypted messenger.

AI chats are not protected by attorney-client privilege

[Fact]: A U.S. federal judge ordered documents from a chat with Claude to be produced. More than 12 major law firms issued warnings to clients.

Important precedent: AI chats are not protected by attorney-client privilege. Corporate correspondence with AI on legal and business matters can become evidence in court.

Vercel hacked

Vercel confirmed the breach: hackers got access to environment variables. The dump is being sold for $2 million. All users — immediate secret rotation.

Research and Analytics

Key Takeaway: Three notable publications cast doubt on the pace of AI adoption: productivity is still not rising, agents failed AutomationBench, AI threatens democracy.

• Science: AI threatens elections and democracy
• Fortune: survey of 6,000 executives — AI is still not visible in productivity statistics (the Solow paradox)
• Anthropic postmortem — how compute savings reduced the quality of Claude Code
• Cloudflare: internal AI development stack with details and figures

Weekly model summary table

Main takeaways of the week

5 key points for April 2026:

1. The model war is accelerating: GPT-5.5, DeepSeek V4, Kimi K2.6, and Qwen3.6 were released within days of each other — an unprecedented pace.
2. DeepSeek V4 — the best deal: 10x less memory, 4x less compute with GPT-5.5-level quality on coding tasks.
3. Anthropic reached $1 trillion: Google + Amazon investment = $53 billion. The most heavily funded AI startup in history.
4. Agents still do not work in practice: AutomationBench — no model surpassed 10%. The hype is outpacing reality.
5. AI legal risks are growing: AI chats are not protected by attorney-client privilege, Vercel has been hacked. AI infrastructure security is a critical issue.