AI Digest: Perplexity 1

April 19, 2026 · 6 min read

Perplexity

April 14, 2026: Anthropic released Claude Opus 4.6 and Sonnet 4.6 — updated flagship models with improved reasoning and coding capabilities, leading in fresh benchmarks.
April 15, 2026: Google introduced Gemini 3.1 Pro — a powerful version with extended context and superiority in mathematical tasks, surpassing competitors in speed and accuracy.
April 16, 2026: Inception announced Mercury 2 — an innovative model for multimodal tasks, focusing on processing video and 3D content with low computational costs.
April 17, 2026: MiniMax launched M2.5 — a compact MoE model with 30B parameters (3B active), catching up to GPT-4o in code and reasoning at 5 times the price.

Sources:

Habr: Key AI World Events of August 2025 Week
Habr: Neuro-Digest for the 3rd Week
Habr: Key AI World Events of September 2025 Week 1
AI-Stat: Neural Network Release Timeline 2026
AI-Stat: AI Ranking 2026
YouTube: AI News: A Week of Major Releases

Gemini

Top AI Model Releases of the Past Week: Focus on Agent Systems and Multimodality

The last week of April 2026 has been marked by significant developments in the world of artificial intelligence, with a focus on the advancement of agent systems and the expansion of multimodal capabilities in models.

DeepSeek V4 and Tencent MixFormer: Chinese Breakthrough in Multimodality and Practicality. Mid-April is expected to see the release of the multimodal model DeepSeek V4, which promises significant progress in code processing and long-term memory, with deep adaptation to Chinese chips. Concurrently, Tencent has announced MixFormer, a 3-billion-parameter model focused on practical applications.
OpenAI GPT-5.4-Cyber: Specialized Model for Cybersecurity. On April 14, OpenAI introduced GPT-5.4-Cyber, a version of GPT-5.4 fine-tuned for cybersecurity tasks. The model boasts enhanced functionality, including binary reverse engineering, and is designed for advanced defense workflows.
Meta Muse Spark: A Step Towards Personal Superintelligence. On April 9, Meta announced Muse Spark, the first model from Meta Superintelligence Labs capable of processing information, thinking, and solving problems using multiple parallel agents. The model includes features for healthcare, coding, and social media analysis, moving closer to the creation of personal superintelligence.
NVIDIA Ising: AI for Quantum Computing. On April 9, NVIDIA released Ising, the first open-source AI models designed to address the limitations of quantum computing. Ising reduces the calibration time for quantum systems and corrects errors significantly faster and more accurately.
The Shift to Agent AI: The Trend of the Week. Overall, the week confirmed the trend of transitioning from generative AI to autonomous agent systems. Models are no longer just responding to prompts; they understand goals, break them down into tasks, and execute multi-step workflows across various software environments.

Sources:

AI News: What's Happening in the World of Artificial Intelligence in April (Vertex AI Search)
The Agent AI Revolution: 7 Breakthroughs Changing Technology in April 2026 (Switas Consultancy)
AI Shows Uneven Progress: Stanford Presents AI Index 2026 Report (AZERTAC)
The Future of Artificial Intelligence: 5 Breakthroughs Defining April 2026 (Vertex AI Search)
Who Can Do What and Why You Need It: A New Model is Released Every Week... 2026 (VKontakte)
The Future of AI: 7 Breakthrough Trends Defining 2026 (Switas Consultancy)
Trends in the Artificial Intelligence Market (TAdviser)
Most Popular Neural Networks in Russia in 2026 Named: Chat GPT, Claude Opus, and Nano Banana Top the List (ForkLog)
10 Breakthrough Developments of 2026 According to MIT Technology Review (Habr)
AI Digest: News and Trends (QuData)
AI Digest for March-April 2026 (Astana Hub)
Latest News from the World of Neural Networks and AI (Deeplist)
OpenAI Responds to Anthropic with GPT-5.4-Cyber AI Model Release (ForkLog)
Chinese Companies DeepSeek and Tencent Prepare New AI Models for April Release (Startpack)
World Completes Major World ID Protocol Update for the AI Era (ForkLog)
Updatable Digest of Changes in AI Search (Netpeak Journal)
AI Technology News: Monday, April 13, 2026 at 08:50 #QixNewsAI (YouTube)
The 2026 AI Model Race: Real Progress, Marketing Hype, and What It Means for All of Us (Habr)
25 Neural Networks Released in February 2026. Compiled into One Collection So You Don't Miss Out (AI on vc.ru)

ChatGPT

From April 12 to April 18, 2026, the following significant models were introduced in the field of artificial intelligence:

Perplexity AI integrated agent mode into its Max subscription, expanding the functionality of its platform.
Alibaba continued to promote open-source models, strengthening its position in the AI ecosystem.
Mozilla introduced the self-hosting Thunderbolt client, offering new possibilities for developers.
OpenAI and Anthropic focused on creating AI as managed infrastructure, including the specialized GPT-Rosalind model and the MCP protocol.
Adobe and Canva enhanced their creative ecosystems, transforming design services into agent-based work environments.

Sources:

AI Digest: April 17, 2026 — vc.ru
AI Digest — December 2, 2025 — vc.ru

Grok

April 16 — Anthropic released Claude Opus 4.7, the company's most powerful public model: significant improvements in complex programming, image processing, and autonomous agent tasks, surpassing its predecessor on all key metrics.
April 17 — xAI introduced Grok 4.3 Beta, a new pre-trained model focused on enhanced capabilities, integrated into the Grok ecosystem for users.
April 14 — Models for 3D reconstruction and simulations were released to the scientific community: LingBot-Map by the authors of Geometric Context Transformer for streaming 3D scene reconstruction, and HY-World 2.0, a multimodal framework for generating high-quality 3D scenes from various inputs.
April 12–13 — New specialized models emerged: OmniShow for generating human-object interaction videos and Habitat-GS, a navigation simulator with dynamic Gaussian Splatting for photorealistic rendering in embodied AI.

Sources:

Claude

Excellent, I have up-to-date data. I will prepare an overview focusing on model security and accessibility – this is a significant departure from a regular digest.

AI Stratification: From Closed Scientific Models to Code Democratization

The last week of April revealed the industry's main contradiction: while closed models become more powerful, open-source ones are reaching frontier-level quality – but control over their use differs drastically.

1. Anthropic Locked Down Its Best Model (April 7)

Claude Mythos, the most capable model Anthropic has ever created, will not be publicly available. Fifty organizations will receive gated access under the Project Glasswing program. The model excels in reasoning, coding, and cybersecurity vulnerability detection, accessible only through early access from approximately 50 partner organizations. This is not a beta – it's a deliberate decision to prohibit public use of the most powerful tool.

2. Zhipu Opened a GPT-5.4 Competitor on the Same Day (April 7)

As Anthropic was closing Mythos, Zhipu AI released a model on the same day that surpassed GPT-5.4 in coding. GLM-5.1, under an MIT license, outperforms GPT-5.4 on the SWE-bench Pro programming benchmark at a price of $3 per month on the coding plan. This is a direct challenge to closed models through an unrestricted license.

3. Meta Released Three Models Over the Weekend (April 2–8)

Google Gemma 4 (April 2, Apache 2.0, 4 variants up to 31B), Meta Llama 4 Scout and Maverick (April 5, open-weight MoE with 10M and 1M token context windows). Both Llama 4 Maverick and Scout were released by Meta on April 5 as open-weight Mixture of Experts architecture models. A shift from language models to specialized agent systems.

4. OpenAI Launched a Scientific Model (April, exact date not specified)

GPT‑Rosalind is available as a research preview in ChatGPT, Codex, and API for qualified clients through a trusted access program. A freely available Life Sciences research plugin for Codex is also being introduced. The first frontier model specifically for biology, not a general-purpose one.

5. Loss of Control Through Open Source (Throughout the Week)

In the last 24 hours, the growth of OpenClaw (formerly Clawdbot) has turned it into the fastest-growing open-source initiative in GitHub history, exceeding 302,000 stars. OpenClaw is an autonomous agent framework that runs locally on the user's machine, allowing it to execute shell commands, manage files, and automate web tasks through messengers like WhatsApp, Telegram, and Signal.

Overall Pattern: The frontier is moving into isolation (Mythos behind glass), but competitors are filling the gap with open-source alternatives. Control has become more important than competitive advantage.

Sources:

WhatLLM.org — April 2026 AI Model Releases
BuildFastWithAI — Latest AI Models April 2026
BuildFemaleSwitch — New AI Model Releases in April 2026
DevFlokers — AI News Last 24 Hours (April 2026)
OpenAI Release Notes

Sources:

Blog