← Home

Blog

Articles about AI search, verification, and platform updates

AI News: The Day's Highlights

May 4, 2026 · 4 min read

Perplexity

DeepSeek Releases New Model Optimized for Long Dialogues (May 2, 2026)

Chinese company DeepSeek has introduced an updated version of its language model that addresses a critical problem with current LLMs – the sharp increase in computational costs when processing long dialogues [6]. The new model demonstrates results comparable to the previous version, DeepSeek-V3.1-Terminus, while maintaining high accuracy and the ability to efficiently work with long text sequences [6].

Significantly, DeepSeek has released the source components under an MIT license and provided open model weights [6]. This allows other researchers to verify solutions and develop them further – an approach that contrasts with the closed models of market leaders. The release confirms a trend identified by experts: there is no longer a monopoly on AI progress, and innovations are coming not only from large Western laboratories [4].

Sources:

  • SecurityLab: Overtook ChatGPT, cost pennies, and surprised again. DeepSeek
  • Habr: AI Model Race in 2026: Real Progress, Marketing

Gemini

In mid-April 2026, Anthropic introduced Claude Opus 4.7, a new flagship model that demonstrated leading results on most agent and coding benchmarks among publicly available solutions. The model boasts an input context window of 1 million tokens and an output of up to 128,000 tokens, along with improved reasoning capabilities and increased vision resolution up to 3.75 MP. Claude Opus 4.7 achieved top scores on SWE-bench Verified (87.6%) and GPQA Diamond (94.2%), and also secured leading positions in the Terminal-Bench 2.0 (69.4%) and OSWorld (78.0%) benchmarks.

Another significant event was the emergence of OpenAI's GPT-5.5, announced on April 22, 2026. This model is fully omnimodal, capable of processing text, images, audio, and video within a unified architecture. GPT-5.5 leads the Artificial Analysis Intelligence Index with a score of 60, surpassing Claude Opus 4.7 (57). The model offers a context window of up to 1 million tokens and is available in three versions: GPT-5.5, GPT-5.5 Thinking, and GPT-5.5 Pro.

Also noteworthy is the release of DeepSeek V4 on April 27, 2026, which includes the V4-Pro model with 1.6 trillion parameters and a context window of 1 million tokens. This model is optimized for operation on Huawei Ascend chips and is positioned as a competitor to closed models, outperforming open-source counterparts in agent programming and reasoning tasks.

Sources:

  • Claude Opus 4.7, GPT-5.5, DeepSeek V4: Top LLM Releases of April 2026
  • LLM Leaderboard 2026 — Compare Top AI Models
  • AI Ranking | Comparison of Language Models and Neural Networks | Best AI 2026

ChatGPT

On April 16, 2026, Anthropic introduced the updated language model Claude Opus 4.7. This model achieved 64.3% on the SWE-Bench Pro benchmark, making it the leader among public models. Additionally, it showed strong results on other benchmarks, including Terminal-Bench 2.0 (69.4%) and GPQA Diamond (94.2%). Claude Opus 4.7 supports an input context of up to 1 million tokens and an output context of up to 128 thousand tokens, as well as improved vision resolution up to 2576 pixels. The model is available at a price of $5 per million tokens, with additional rates for contexts exceeding 200,000 tokens.

Sources:

Grok

On April 24, 2026, Chinese company DeepSeek released a preview version of its highly anticipated large language model V4 – fully open-source, with "pro" and "flash" versions. The model is optimized for agentic tasks, knowledge processing, and inference, showing excellent results in relevant benchmarks against competitors, while requiring significantly fewer computational resources. It is compatible with local Huawei Ascend chips, strengthening China's position in the AI race amidst restrictions on Nvidia.

This is not just an upgrade: V4 reduces reliance on imported hardware, allows developers to freely modify code, and runs locally, fueling competition in the open-source segment.

Sources:

  • CNBC: China's DeepSeek releases preview of long-awaited V4 model
  • LLM Stats: AI Trends (May 2026)

Claude

Found an interesting discovery – it's worth delving into the topic of 1-bit models, it's a real breakthrough in April 2026. Let me clarify the details. Great, now I have enough information. This is a truly significant breakthrough, a completely different angle than political news and benchmarks.

March 31, 2026. Architectural Breakthrough: 1-Bit Language Models

PrismML, founded by researchers from Caltech, has emerged from stealth mode with a