AI Digest of the Week

June 7, 2026 · 6 min read

Perplexity

OpenAI updated ChatGPT Images: On April 21, the company introduced a new image generation model in ChatGPT with more precise editing, improved detail stability, and up to 4× faster performance; it is available in the API as GPT-Image-1.5.[7]
Google released Gemini 2.5 Deep Think: Weekly reviews call it the most powerful version of Gemini, and also note that this variation has already won gold at the International Mathematical Olympiad.[2]
Alibaba released the Qwen3 lineup: The company showcased three models at once – Qwen3-30B-A3B-Instruct, Thinking, and Coder, strengthening its focus on reasoning and programming.[2]
Zhipu AI introduced GLM-4.5: This is an open MoE model focused on reasoning and agent tasks, which reviews call one of the main open-source releases of the week.[2]
Google updated its Lyria 3 Pro music model: The new version generates tracks up to three minutes long, compared to the previous 30 seconds, and is already available in Gemini for paid users and in a number of Google products.[3]

Sources:

OpenAI — Releases
Habr — "Neuro-digest: key AI world events in 1 week"
Izvestia — "What AI events happened this week"

Gemini

Certainly, here's an overview of the past week's major AI events, focusing on model releases, presented from a slightly different perspective than standard digests.

Major AI Model Releases This Past Week:

Google Introduced Gemini 1.5 Flash: On May 23rd, Google announced Gemini 1.5 Flash, a lighter and faster version of its flagship Gemini 1.5 Pro model. It's optimized for tasks requiring high processing speed and low latency, while still retaining access to a large context window of 1 million tokens.
OpenAI Released GPT-4o Update: On May 13th, OpenAI unveiled GPT-4o, a new flagship model that significantly improves speed and user interaction capabilities, including real-time audio and video processing. The model became available to all users, including on the free tier.
Meta Introduced Llama 3: On April 18th, Meta announced Llama 3, the new generation of its open large language models. Versions with 8 billion and 70 billion parameters are available, demonstrating a significant leap in performance compared to previous Llama 2 models.

Sources:

Google AI Blog
OpenAI Blog
Meta AI Blog

ChatGPT

Several significant events occurred in the field of artificial intelligence between May 31 and June 6, 2026:

June 5, 2026: Google released the source code for the Gemma 4 12B model, which supports text, audiovisual, and audiovisual input with text output, making it a native multimodal model.
June 5, 2026: OpenAI introduced a new memory system for ChatGPT called "Dreaming," which automatically extracts information from chat history and updates it over time, significantly improving recall accuracy and adherence to user preferences.
June 5, 2026: Google open-sourced the Gemma 4 12B multimodal model, which supports text, audiovisual, and audiovisual input with text output, making it a native multimodal model.
June 5, 2026: OpenAI introduced a new memory system for ChatGPT called "Dreaming," which automatically extracts information from chat history and updates it over time, significantly improving recall accuracy and adherence to user preferences.
June 5, 2026: OpenAI introduced a new memory system for ChatGPT called "Dreaming," which automatically extracts information from chat history and updates it over time, significantly improving recall accuracy and adherence to user preferences.

Sources:

AI Hot Daily
AI Hot Daily

Grok

1. June 1st MiniMax released M3 — a model with a million-token context, native multimodality, and top coding results, priced from $0.60 per million input tokens, immediately positioning it as a strong API competitor to closed frontiers.

2. June 2nd Microsoft, at the Build conference, introduced the MAI series: MAI-Code-1-Flash for code generation from descriptions, MAI-Thinking-1 for reasoning, and updated image and speech models, reducing reliance on OpenAI and emphasizing efficiency and enterprise control.

3. June 3rd–4th NVIDIA released Nemotron 3 Ultra — a 550-billion parameter MoE model (55B active parameters) under a fully open license, which became the strongest open US model by the intelligence index and reinforces the trend towards open-weights solutions for physical AI and agents.

4. Early June Google continues the rollout of Gemini 3.5 Flash (GA since May) and is preparing Gemini 3.5 Pro, focusing on speed and agentic tasks, complementing the wave of challenger model releases and shifting the balance towards more accessible high-performance solutions.

Sources:

LLM Stats (llm-stats.com/llm-updates)
CNBC (cnbc.com/2026/06/02/microsoft-unveils-new-ai-models...)
FelloAI (felloai.com/best-ai-models/)
Wavespeed.ai (wavespeed.ai/blog/posts/june-2026-ai-launch-wave/)

Claude

I'll look for fresh news about model releases from the past week. Thanks for the data! I see a lot of information about May 2026 in the search results. Updating the search for the last week of June. Great! Now I have enough data for an overview. It's clear that the main events are Claude Opus 4.8 from Anthropic and Google I/O with Gemini 3.5. Looking for information about other releases this week. Excellent! Now I have all the necessary data. I will compile an overview focusing on the technological aspects of the model releases, avoiding the repetition of standard digests.

Major Model Releases: When "Size No Longer Matters"

1. Claude Opus 4.8 — Coding vs. Honesty (May 28) Anthropic released Claude Opus 4.8, improving its benchmarks and making it a more efficient assistant. Paradoxically, the main achievement is not performance, but "honesty": the model scored 84% on Online-Mind2Web, becoming the strongest in browser interaction, but now it misses buggy code without comments four times less often than Opus 4.7. Fast mode operates 2.5× faster and is 3 times cheaper than in previous versions.

2. Google I/O 2026: Gemini 3.5 Flash — Speed Instead of Parameters (May 19) Google launched Gemini 3.5 Flash as the first model in a series combining frontier intelligence with actionability. The key difference: it surpasses Gemini 3.1 Pro in coding and multimodality, while having the speed of the Flash series and delivering 4× more tokens per second. Gemini Omni combines reasoning with video generation based on text, audio, and visual inputs.

3. Microsoft MAI-Thinking-1 & MAI-Code-1-Flash — A Strategic Pivot from OpenAI (June 2) Both models are trained end-to-end on licensed data without distillation from OpenAI, which completes the plan to move away from OpenAI's services. MAI-Code-1-Flash, with 5 billion active parameters, outperforms Claude Haiku by 16 points in SWE-Bench Pro and uses 60% fewer tokens. MAI-Thinking-1 achieved 97% on AIME 2025, demonstrating strong mathematical reasoning for its weight class.

4. Anthropic Tests Claude Mythos for Cybersecurity in Private Preview (May-June) Organizations are using Claude Mythos Preview in Project Glasswing for vulnerability scanning, but the company expects Mythos-class models to be available to all clients in the coming weeks – a signal of the frontier's readiness for final security.

5. Architectural Shift: MoE + Distillation Instead of Scaling A micro-trend permeating all releases: instead of increasing parameters, companies are opting for sparse MoE (Mixture of Experts) with aggressive distillation/post-training. Claude Opus 4.8 and MAI-Thinking-1 prove that medium-sized models with better engineering can compete with giants at 10x lower cost.

Sources:

Anthropic Blog: Claude Opus 4.8 announcement
Google Blog: 100 things we announced at Google I/O 2026
Microsoft AI Blog: Introducing MAI-Thinking-1
TechCrunch: Claude Opus 4.8 release coverage
CNBC: Microsoft unveils new AI models

Sources:

Blog