GPT-5.6 Thursday ⭐️, Claude Cowork mobile 📱, Gemini API agents 🤖

OpenAI chose to publicly release the three GPT-5.6 models on July 9 (Thursday): Sol, Terra, and Luna. Sol is the flagship model, Terra is described as a balanced option for everyday work, and Luna focuses on speed and low cost. Unlike the previous "one model fits all" strategy, this tiered pricing clearly targets differentiated needs across user scenarios.

Notable data: Terra competes with GPT-5.5 in performance but at half the cost; Luna offers strong capabilities at the lowest cost. This suggests OpenAI is trying to provide more flexible choices between cost and performance while capturing the low-cost market with Luna. However, these performance comparisons come from official statements and have not been independently verified.

Microsoft is starting to replace OpenAI and Anthropic models with its own in products like Excel and Outlook. The decision is driven by cost considerations: Microsoft consumes massive amounts of tokens, and although it previously received discounts, those agreements are expiring. Avoiding being "held hostage" by leading AI labs on pricing is Microsoft's real motivation.

This move echoes DeepSeek's plan to develop its own chips. Due to US export controls, DeepSeek cannot freely access advanced chips, so it is focusing on designing data center inference chips and has been in talks with potential partners for a year. Both companies are trying to vertically integrate the AI value chain—from models to chips or application layers—to reduce dependence on external suppliers.

The release of MiniMax M3 demonstrates the potential of sparse attention in long-duration tasks. By reading only part of the context at each step, it keeps costs predictable even as the context grows indefinitely. This makes long-running agents (e.g., continuous monitoring or codebase analysis) economically viable. M3 has a context window of 1 million tokens and is open-source.

Anthropic's Claude Cowork addresses the "persistence" problem of agents—users can start a task on mobile or web and close the device, and the task continues running in the background. This shifts agents from "conversations" to "background workers" at the product level.

Google's Gemini managed agents also add background execution, remote MCP integration, and other features, turning agents into asynchronous workers with all computation in isolated sandboxes. Additionally, research has introduced "anti-doom training" targeting repetitive degradation (doom loops), which almost completely eliminates loop collapse through final token preference optimization. These technologies together point to a goal: enabling AI agents to run reliably and cost-effectively in real-world environments.

Google released the Gemma 4 technical report, introducing dense and MoE architectures, parameter ranges from 2.3B to 31B, and supporting thinking mode and long context. These models compete with larger models on multiple benchmarks and are licensed under Apache 2.0. This means the open-source camp is narrowing the gap with closed-source models.

MiniMax M3 is also released as open-source, and its sparse attention mechanism may become the standard for future long-context agents. Meta's Muse Image is also freely available, integrated into Instagram and WhatsApp. These developments show that competition in open-source and free models is intensifying even in image generation and multimodal domains.