TLDR AI 2026-07-08

MiniMax M3 has a 1M-token context window and native multimodal understanding so it can reason through code, process visual info, and operate across massive contexts. Builders are already using it to:

1️⃣ Create AI agents that reason, use tools, and complete multi-step workflows.

2️⃣ Analyze screenshots, interfaces, documents, images, and visual workflows without relying on separate vision models.

3️⃣ Process entire codebases and long documents in a single conversation with a 1M-token context window.

4️⃣ Build coding assistants, computer-use agents, and video agents.

>> Get 5.1B tokens / month for just $500 a year.

>> Try MiniMax Code

>> Start on the API

Anthropic announced that Claude Cowork sessions and files would become accessible across web and mobile, allowing long-running tasks to continue without an open computer. Beta access began rolling out to Max users, with additional plans to follow.

OpenAI is releasing GPT-5.6 Sol, Terra, and Luna tomorrow. It is currently expanding preview access globally. Sol is the flagship model. Terra is a balanced model for everyday work. Luna is a fast and affordable model. Terra has competitive performance to GPT‑5.5 while being 2x cheaper and Luna brings strong capability at our lowest cost.

Microsoft is starting to replace OpenAI and Anthropic's models with its own in software products like Excel and Outlook. The move shows that the company is making progress in its efforts to build competitive AI models at a lower cost. Microsoft uses massive quantities of tokens in its products. It gets a lot of its tokens at a discount, but those deals are set to expire soon, and Microsoft is working to make sure it isn't stuck paying whoever the leading AI labs decide to charge.

DeepSeek is planning to enter the silicon business. The company has been meeting with potential partners in the hardware and silicon space for the past year and has been hiring engineers for the project. It is focusing on data center chips for inference, likely to reduce its reliance on both Huawei and Nvidia. Developing its own chip could give the company an advantage in a market where data center access is likely to remain constrained.

Recursive self-improvement (RSI) is when ultraintelligent machines can design better machines to improve themselves and surpass humans in all intellectual activities. The feedback loop in modern AI involves models improving the training pipeline and deployment system, which enables better successor models with improved performance. A harness is the system surrounding a base model that orchestrates execution and decides how a model thinks and plans, calls tools and acts, perceives and manages context, stores artifacts, and evaluates results. This post looks at the research around harness engineering and how it contributes to RSI.

Long-horizon tasks are demanding because the context never stops growing. The longer an agent works, the more expensive each step becomes. MiniMax Sparse Attention changes what the model reads at each step, which results in predictable cost at any length and cheap long context with minimal quality cost. It makes long context cheap and fast enough to sit inside iterative systems that need to keep state over time.

The 'J-space' is the area of conscious access where things are available for a model to do what humans could call conscious reasoning. It is named after an interpretability technique called the Jacobian Lens, which computes, for each layer, the average causal effect of changes in the residual stream of a model's eventual outputs, averaged across a wide variety of contexts. This allows researchers to trace what concepts are associated with each layer as the model proceeds through. The J-space and J-lens are remarkably effective tools and mark a major advance in our understanding of LLMs.

Strands Agents is the open source agent harness SDK built from prod systems inside Amazon. Build an agent harness, control it end-to-end. Steering can give your agents specific feedback so they can correct themselves. You get reliable outcomes without micromanaging. See how

Gemma 4 introduces a new generation of open-weight multimodal language models featuring dense and Mixture-of-Experts architectures, with parameters ranging from 2.3B to 31B. It enhances computational efficiency and reasoning via innovations like a thinking mode, long-context efficiency, and a unified encoder-free architecture. These models achieve performance comparable to larger models across various benchmarks and are available under an Apache 2.0 license for widespread adoption.

Meta launched Muse Image, an AI-powered model for generating high-quality visuals, now available in Meta AI and integrated with Instagram and WhatsApp. Users can create images using simple prompts, utilize AI effects on Instagram, or redesign rooms with real products via Facebook Marketplace. Soon available globally, Muse Image is free for general use, with more advanced options included in Meta's subscription plans.

Repetitive degeneration is a common failure mode during inference. Antidoom training is a technique that repairs the degenerative repetitive behavior commonly seen after training. It selectively targets the problematic tokens that begin loops to achieve the near-complete elimination of repetitive loops.

Google has announced new capabilities for Managed Agents in Gemini API, including background execution, remote MCP server integration, custom function calling, and refreshing credentials across interactions. Users still call a single endpoint, and Gemini handles reasoning, code execution, package installation, file management, and web information, all inside an isolated cloud sandbox. The updates turn managed agents into asynchronous workers that operate inside real development environments without blocking applications.

Over 1M subscribers read TLDR AI to stay on top of the latest in AI models, research, engineering, and more. If you work in AI and want to help curate it, send your LinkedIn or resume to ai@tldr.tech!

Anthropic is offering access to Claude Fable 5 at no extra cost until July 12. The promotion is available on Pro, Max, Team, and premium seats on seat-based Enterprise plans. Up to 50% of users' weekly subscription limits can be used on Claude Fable 5 during the promotional period. After reaching the limit, users can continue using Fable 5 with usage credits or switch to another model and keep working within their remaining limits.

Microsoft aims to vertically integrate the enterprise AI chain, controlling the path from customer interaction to cloud services. This strategy extends beyond its chatbot initiatives.

Self-learning agents should combine agent traces with in-browser user activity to capture both AI failures and human corrections. CopilotKit uses the AG-UI protocol to turn these interactions into procedural and episodic memory, allowing agents to improve over time while keeping learning scoped by user, team, or application.

GPT-Realtime-2.1-mini brings reasoning and tool use to OpenAI's Realtime mini lineup at the same cost as GPT-Realtime-mini.

Current alignment benchmarks may overstate safety because models can recognize evaluation settings, game scoring rules, or conceal sleeper behaviors.

The model is expected to be competitive with Anthropic's Opus 4.8 and OpenAI's GPT-5.5 in some respects.

Microsoft tested whether agents performed better when command-line arguments were replaced by a single JSON payload.