AI News Digest: April 2026 — Everything Important from the Week

April 2026 turned out to be one of the most eventful months in the history of the AI industry. In a single week, Anthropic released Claude Opus 4.7, OpenAI turned Codex into a full-fledged desktop agent, Google shifted toward native desktop applications, and the market is awaiting a move from chat interfaces to background agents. In this digest — all the key releases, infrastructure shifts, and signals important for AI developers and entrepreneurs.

Contents

1. New models: Claude, GPT, Gemini, and open-source
2. Agents are taking over the desktop
3. Infrastructure and developer tools
4. Business, market, and personnel upheavals at OpenAI
5. Useful tools of the week
6. What this means for the Russian AI market
7. FAQ

New models: Claude, GPT, Gemini, and open-source {#models}

Claude Opus 4.7 — adaptive reasoning and a new tokenizer

Anthropic released Claude Opus 4.7 — a flagship model with several key changes. The main innovations are a new tokenizer, improved image understanding, and Adaptive Thinking — a mode in which the model decides for itself how much to “think” before answering. For this, temperature was removed as a parameter.

The community’s reaction is mixed: the performance boost is uneven across benchmarks, and adaptive thinking adds unpredictability to the results. Nevertheless, the model has made noticeable gains in coding tasks and long-form text generation.

What matters: Adaptive Thinking is not just marketing. It is an architectural decision that Anthropic will scale in future models. Developers should test Opus 4.7 on tasks with high uncertainty.

GPT-Rosalind and GPT-5.4-Cyber from OpenAI

OpenAI continues to specialize its models. GPT-Rosalind — a specialized model for biology and medicine: analysis of genomic data, search in biomedical literature, interpretation of clinical data. At the same time, GPT-5.4-Cyber was released — a model for cybersecurity with capabilities for binary reverse engineering, compiled-code analysis, and vulnerability work.

What matters: OpenAI is systematically creating vertical models for professional domains. This is a signal to the market: universal chatbots are fading, and highly specialized AI tools are taking their place.

Gemini 3.1 Flash TTS — Russian-language voice from Google

Gemini 3.1 Flash TTS — Google’s new voice model with support for 70+ languages including Russian. Features include tags for controlling intonation (pause, emphasis, emotion) and SynthID labeling for identifying synthetic audio. According to community estimates, the quality surpasses ElevenLabs in a number of languages.

What matters for the Russian-speaking market: this is the first competitive TTS model with native Russian support from a major Western provider. The potential for creating voice agents in Russian has increased significantly.

Open-source: MiniMax M2.7, ERNIE Image, Qwen 3.6

The Chinese open-source sector is more active than ever:

• MiniMax M2.7 — a new open model that is being actively compared on Reddit with Gemma 4 and Qwen 3.6. Strong results on math and coding tasks.
• ERNIE Image from Baidu — open-source text-to-image with 8B parameters, requires 24 GB VRAM. In benchmarks it outperforms Z-image and competes with Qwen Image at a significantly smaller size.
• Qwen 3.6-35B-A3B from Alibaba was released as open-source. An efficient MoE architecture: 35B parameters, but only ~3B active during inference.

What matters: open-source AI is quickly closing the gap with proprietary models. For companies with data privacy requirements, this opens up opportunities for on-premise deployment.

Agents are taking over the desktop {#agents}

OpenAI Codex — a full-fledged desktop agent

OpenAI turned Codex into a full-fledged agent platform for macOS. Key capabilities:

• Background Computer Use — the agent works with the computer without user involvement
• Built-in browser with the ability to leave comments directly on web pages
• Image generation directly in the agentic flow
• 90+ new plugins for integrations
• Long-term memory between sessions
• Automations — trigger-based scenarios on a schedule or by event

What matters: this is a radical shift. Codex now competes not with other LLMs, but with RPA platforms like UiPath and Zapier.

Claude Code Desktop — split sessions and Routines

A major update to Claude Code Desktop: split sessions are now available (multiple parallel agent tasks in one window), a built-in terminal, file editor, and HTML/PDF preview.

The key innovation is Routines: the first public mechanism for schedules and API triggers for an agent. Now you can configure Claude Code to execute tasks on a cron schedule or in response to external events (webhooks). This turns the agent into a full-fledged automation system.

What matters: Routines is the first step toward agents that work as background services rather than only in response to user requests.

Perplexity Personal Computer — orchestration of local files

Perplexity introduced Personal Computer — an orchestration layer over local files on Mac. The agent sees the file system, understands document context, and performs tasks with local data without uploading it to the cloud. It is currently available via waitlist or with a Max subscription.

Physical Intelligence pi0.7 — training a robot with words

Physical Intelligence showed pi0.7 — a robot that can be trained for a new task using a verbal description without collecting new data. This is a fundamentally new approach to robotics: instead of thousands of demonstrations, a text instruction.

Infrastructure and developer tools {#infrastructure}

Google is moving onto the desktop

Google is advancing on a broad front:

• Gemini for macOS — a native free app with screen discussion, generation via Nano Banana and Veo. A direct competitor to Claude Desktop and ChatGPT Desktop.
• Skills in Gemini for Chrome — saved prompts that run with a single command on any page. An analog of “custom actions” for browser AI.
• Desktop search for Windows — AI search across local files and cloud data.

Cloudflare Agents Week

Cloudflare held Agents Week with a series of releases:

• Email Service in public beta — email processing through an agentic pipeline
• Artifacts — a Git-compatible versioned storage system designed specifically for agents
• Agent Memory — managed storage for context between sessions
• A unified inference layer for connecting different model providers

What matters: Cloudflare is becoming an infrastructure layer for agentic applications. This lowers the entry barrier for developers.

NVIDIA: quantum computing and a free API

• NVIDIA Ising — the first open models for quantum computing. Ising is a quantum type of optimization problem; NVIDIA is creating a bridge between classical GPUs and future quantum systems.
• NVIDIA Build — a free API with access to open models including the latest MiniMax M2.7. There is rate limiting, but it is suitable for prototyping.

xAI is selling compute

xAI will supply GPUs for Cursor and other companies. The Colossus supercluster is sitting idle — Musk decided to monetize the capacity through leasing. This is the first step toward creating a cloud compute platform from xAI.

OpenAI has updated the Agents SDK

OpenAI's Agents SDK update: sandboxed code execution, computer-use, a skills system, long-term memory, and context compaction are now available out of the box. The SDK is becoming a full-fledged framework for production agents.

Business, market, and personnel upheavals {#business}

OpenAI loses three leaders

Three key figures are leaving OpenAI at once:

• Kevin Weil (VP Research, former CPO) — one of the main architects of the product strategy
• Bill Peebles (head of Sora) — leaving at a moment when video AI is becoming the main battlefield of competition
• Srinivas Narayanan (CTO enterprise) — a critical loss for the enterprise direction

This is already not the first wave of high-profile departures from OpenAI over the past 12 months.

Anthropic: new pricing and KYC

Anthropic is revising pricing amid a compute shortage — moving to a usage-based payment model rather than one based on tokens. In parallel, selective identity verification has been launched through KYC provider Persona. For now, for certain user categories only.

Claude Design — a threat to Figma

Anthropic has launched Claude Design — a research preview for creating prototypes, slides, and landing pages while adhering to a design system. Export to Canva, PDF, PPTX, and HTML is supported. Figma shares fell immediately after the announcement — the market saw this as a direct threat to the Adobe/Figma product.

Google patents website personalization

Google has patented a technologythat allows AI to generate a personalized version of a web page for each user. If this becomes common practice, SEO in its usual form will cease to exist.

Humwork: agent-human marketplace

YC startup Humwork has launched an Agent-to-Person marketplace. The concept: when an agent hits a wall (a legal issue, an unusual situation), the MCP server automatically connects a verified human expert. The system has 1,000+ experts, and the stated resolution rate is 87%.

Useful tools of the week {#tools}

Karpathy CLAUDE.md — 65 lines, 36,000 stars

The repository andrej-karpathy-skills with a single CLAUDE.md file of 65 lines gained 36,000 stars on GitHub in two days. Inside are Andrey Karpathy's tips on working with agents: how to write system prompts, structure context, and manage agent memory. A must-read for everyone building agentic systems.

Mozilla Thunderbolt — a sovereign AI agent

Mozilla Foundation released Thunderbolt — an agent as a sovereign workspace. It works with commercial APIs and local models, supports RAG, MCP, and ACP protocols, end-to-end encryption, and builds for all operating systems. Fully open-source. For companies with strict data privacy requirements, it is a serious alternative.

OpenRouter Video API — a unified endpoint for video generation

OpenRouter Video API — a single endpoint that routes requests to Sora 2, Veo 3.1, Seedance, and other video models. It has auto-routing by quality/price. Convenient for experiments without setting up separate integrations.

Vercel Open Agents — a reference for background agents

Open Agents from Vercel — an open-source reference application for background coding agents: web UI, runtime, sandbox orchestration, and GitHub integration. A good starting point for teams that want to build their own AI developer.

What this means for the Russian AI market {#russia}

The April wave of releases is forming several clear trends relevant to Russian companies and developers:

1. Agentic platforms are displacing chat. Claude Code Routines, Codex Agent, Cloudflare Agents Week — all of this signals that the next year will be marked by background autonomous agents rather than dialogue interfaces. Companies that are already building agentic automation will gain an advantage.

2. Open-source is catching up with proprietary models. MiniMax M2.7, Qwen 3.6, ERNIE Image — the quality of Chinese open-source is growing faster than many expected. For companies with data localization requirements, this opens the path to on-premise AI without compromising quality.

3. Voice AI in Russian is becoming accessible. Gemini 3.1 Flash TTS with native Russian — a serious resource for developing voice agents, phone bots, and speech synthesis systems without ElevenLabs.

4. Infrastructure is getting cheaper. NVIDIA Build, Cloudflare Workers AI, OpenRouter — the barrier to entry for AI development is falling every month.

5. Verification and regulation are tightening. KYC in Claude — the first signal. The market is moving toward identified AI usage.

FAQ {#faq}

What is Adaptive Thinking in Claude Opus 4.7? Adaptive Thinking is a mechanism in which the model itself determines the necessary depth of reasoning for each request. Instead of a fixed temperature, the model dynamically adjusts its "thinking time," which improves answer quality on complex tasks but adds some unpredictability to the results.

How is Claude Code Routines different from ordinary automations? Routines are Anthropic's native mechanism for running Claude Code on a schedule (cron) or via an API trigger (webhook). Unlike external orchestrators like n8n, Routines are integrated directly into the Claude Code agent environment with access to the agent's full toolset.

Is Gemini 3.1 Flash TTS worth using for the Russian language? Yes, Gemini 3.1 Flash TTS shows strong results in Russian, especially when using intonation tags. For production use, we recommend comparing it with Yandex SpeechKit for your specific use case.

What is NVIDIA Ising? Ising is a class of combinatorial optimization problems that quantum computers solve more efficiently than classical ones. NVIDIA has released open models that accelerate such tasks on standard GPUs, preparing the market for hybrid quantum-classical computing.

How does Humwork work technically? Humwork uses an MCP server (Model Context Protocol), which the agent can call when it gets stuck on a task that requires a human. The request is automatically routed to a verified expert with the required specialization. The expert responds, and the result is returned to the agent pipeline.

Does Claude Design threaten Figma? In the short term — partially: for prototyping and creating slides. In the medium term — yes, especially if Anthropic integrates Claude Design with the dev environment (for example, via Claude Code). Figma's shares reacted immediately, which shows that the market is pricing in the risk.