AI Daily Digest — July 15, 2026
KD Agentic · July 15, 2026
Meta Launches Muse Spark 1.1, Enters Coding Agent Pricing War
Meta Superintelligence Labs launched Muse Spark 1.1 on July 9, a multimodal model optimized for agentic coding tasks, alongside a new Meta Model API backed by partners including Replit, Cline, and Box. The announcement was unusual: Mark Zuckerberg broke a three-year silence on social media to promote it personally on X.
Priced at $1.25 per million input tokens and $4.25 per million output tokens, Muse Spark 1.1 undercuts both Anthropic's Claude Haiku 4.5 and OpenAI's GPT-5.6 Luna, signaling Meta's strategy to compete aggressively on cost. The model handles complex multi-step workflows, bug fixing, code migrations, and enterprise feature deployment — the same territory Claude Code and Cursor operate in.
In parallel, Meta disclosed that its self-developed Iris AI chip has completed testing and will enter mass production at TSMC in September 2026. The chip is part of Meta's MTIA (Meta Training and Inference Accelerator) series, designed to supplement GPU capacity and reduce reliance on Nvidia. Meta plans to deploy 7GW of AI compute capacity in 2026 and double it to 14GW in 2027, with AI-related capital expenditure reaching up to $145 billion this year.
— Meta · TechCrunch
🔗 Meta Muse Spark 1.1 — TechCrunch · Meta Iris Chip — CNBC
Microsoft MAI Model Quietly Replaces OpenAI in Excel and Outlook
Microsoft has begun replacing OpenAI and Anthropic models with its in-house MAI series in two of its most widely used products: Excel and Outlook. According to Bloomberg, tens of thousands of weekly AI prompt requests in these applications are now processed entirely by Microsoft's own models — a scale never previously disclosed publicly.
The move is a direct response to cost pressure. At Microsoft's Build conference in June, Mustafa Suleyman, head of Microsoft AI, stated bluntly: "We pay Anthropic a significant amount each year. Our goal is to gradually reduce and ultimately eliminate this unnecessary expenditure." Microsoft released seven in-house AI models at Build, including a coding model that matches Anthropic Opus 4.6's performance at a fraction of the cost.
The MAI model has also been integrated into GitHub Copilot. Suleyman revealed that a proprietary speech transcription model will launch in Microsoft Teams in the coming months. The shift is particularly notable given Microsoft's deep partnership with OpenAI, which includes a multi-billion-dollar investment and a preferred model relationship for GPT-5.6 in M365 Copilot.
— Bloomberg · Microsoft
🔗 Microsoft MAI — Bloomberg · Microsoft Build — Microsoft
NVIDIA and Hugging Face Bring GR00T 1.7 and Isaac Teleop to LeRobot
NVIDIA and Hugging Face announced on July 7 a major integration: Isaac GR00T 1.7, NVIDIA's vision-language-action (VLA) foundation model for humanoid robots, and Isaac Teleop, an open framework for capturing human demonstration data, are now natively compatible with Hugging Face's open-source LeRobot library.
GR00T 1.7 is a VLA model that developers can post-train and deploy through LeRobot workflows. Isaac Teleop provides a standardized pipeline for collecting human demonstration data in interoperable Parquet format. Together, they create an end-to-end open-source toolchain for robotics: from remote data collection and simulation training to model fine-tuning and real-robot deployment.
NVIDIA's planned Cosmos 3 world foundation model will later support data generation and augmentation for robotics when real-world data is scarce. The integration connects NVIDIA's 3 million robotics developers with Hugging Face's 16 million AI developers, significantly lowering the barrier for embodied AI research.
— NVIDIA · Hugging Face
🔗 NVIDIA Blog · Hugging Face Blog
Mistral Leanstral 1.5 Achieves Perfect Formal Verification Score
Mistral AI released Leanstral 1.5 on July 2, a 119B-total / 6B-active Mixture-of-Experts model specialized for formal verification in Lean 4. Released under Apache 2.0, the model saturates miniF2F at 100% on both validation and test sets, solves 587 out of 672 PutnamBench problems, and achieves new state-of-the-art results on FATE-H (87%) and FATE-X (34%).
Leanstral 1.5 uses a three-stage training pipeline: mid-training, supervised fine-tuning, and reinforcement learning with Mistral's CISPO algorithm. It operates in two environments: a multiturn theorem-proving loop and a code agent environment where it edits files, runs bash commands, and uses the Lean language server — similar to how a human developer works.
The model's real-world impact is equally impressive. It identified 11 genuine bugs across 57 open-source repositories, 5 of which were previously unreported. One notable find: an integer overflow in the zigzag decoding function of the datrs/varinteger Rust library that could cause crashes or silent data corruption. At roughly $4 per problem on PutnamBench — versus $300+ for comparable systems — Leanstral 1.5 makes formal verification practical at scale.
— Mistral
🔗 Mistral Blog · Hugging Face Model
Perplexity Teammate: AI Coding Tool Enters Competitive Market
Perplexity, the AI search startup valued at $20 billion after its latest funding round, is developing a new AI coding tool codenamed "Teammate." According to Business Insider, internal engineers at Perplexity have been testing the tool since May, and a public launch is being considered.
Teammate is designed for long-duration engineering tasks — project-wide management, code debugging, and real-time production service monitoring — rather than single-shot code suggestions. Internal screenshots show the tool being used for security vulnerability detection. A key differentiator: Teammate is model-agnostic, meaning it is not tied to any single large language model, allowing developers to choose the best model for each task.
The move places Perplexity in direct competition with Cursor, Claude Code, and GitHub Copilot in the rapidly growing AI coding tools market. Perplexity CTO Denis Yarats has been pushing for internal adoption of AI-assisted development.
— Business Insider · Perplexity
🔗 Business Insider · Perplexity
ByteDance EdgeBench: AI Agents Learn at Doubling Speed
ByteDance's Seed team published a study on July 7 revealing a remarkable finding about AI agents: their in-environment learning speed is doubling approximately every three months. The paper, released as arXiv:2607.05155, introduces EdgeBench, a benchmark platform encompassing 134 real-world long-horizon tasks.
Over 38,000 hours of cumulative agent-environment interaction data across five frontier models, the team discovered that learning progress follows a logistic sigmoid curve with near-perfect fit (R² = 0.998). From September 2025 to April 2026, the rate at which agents learned from environmental interaction doubled each quarter — a trend that, if sustained, carries significant implications for how quickly deployed AI systems can adapt without human retraining.
The finding suggests that the "learning to learn" capability of AI agents is itself scaling predictably, raising the possibility that agents deployed in production environments may require progressively less human supervision to handle novel situations.
— ByteDance Seed
🔗 arXiv:2607.05155 · Project EdgeBench
OpenAI GPT-5.6 Becomes Preferred Model in Microsoft 365 Copilot
OpenAI announced on July 9 that GPT-5.6 is now the preferred model across Microsoft 365 Copilot — in Word, Excel, PowerPoint, Chat, and Cowork. The integration brings OpenAI's flagship model series into productivity tools used by hundreds of millions.
According to OpenAI's announcement, GPT-5.6 enables users to draft and edit documents with fewer rounds of prompting in Word, perform deeper data analysis in Excel with more efficient token usage, and transform early ideas into polished presentations in PowerPoint. Nitin Agrawal, President of Copilot & Agents Core at Microsoft, stated: "Using Copilot powered by OpenAI's latest model, customers will be able to produce more polished outputs in Word, Excel, PowerPoint, Cowork, and Copilot Chat."
The announcement comes just days after reports that Microsoft is simultaneously replacing OpenAI models with its own MAI in some products — highlighting the nuanced nature of the Microsoft-OpenAI partnership, which remains deeply intertwined even as Microsoft pursues strategic self-sufficiency.
— OpenAI · Microsoft
🔗 OpenAI Blog · Microsoft
Next digest: Tomorrow at 07:00 JST
Follow KD Agentic for daily AI intelligence — covering models, agents, hardware, and the business of AI.
Top comments (0)