New Releases
xAI launched Grok Build, a CLI-based coding agent in early beta, available today exclusively to SuperGrok Heavy subscribers at $300 per month.
- Agent Client Protocol support lets Grok Build operate inside third-party applications, and parallel subagent execution targets complex multi-step engineering workflows that single-agent tools cannot handle.
- xAI president Michael Nicolls has set matching Claude Code performance as an explicit internal target, signaling this is a strategic product rather than a side experiment.
Microsoft released a GitHub Copilot desktop app in technical preview today, built around an agentic workflow that starts from a GitHub issue and ends with an auto-merged pull request.
- Agent Merge closes the loop autonomously: it responds to review comments, fixes failing CI checks, and merges PRs once conditions are met, removing human handoffs from routine code changes.
- The launch coincides with a reported internal Microsoft memo winding down Claude Code use in favor of Copilot CLI, suggesting the desktop app is designed to absorb the workflows Claude Code had captured inside Microsoft.
Google DeepMind released Gemma 4 today, a family of open-source models up to 31B dense and 26B MoE, licensed under Apache 2.0 and built on the same research stack as Gemini 3.
- Apache 2.0 licensing is a deliberate departure from prior Gemma restrictions, allowing commercial deployment without Google's usage terms and directly challenging Meta's Llama 4 on open-weights enterprise use.
- Google claims the models outperform competitors up to 20 times their size on benchmarks, which matters most for enterprises self-hosting to control data residency and inference costs.
Google shipped Genkit Middleware today, a system for intercepting, customizing, and hardening agentic AI applications built on any model including Gemini, OpenAI, and Anthropic.
- Middleware hooks let developers enforce guardrails, logging, and fallback behavior at the framework layer rather than inside each agent, addressing the production reliability gap that has blocked enterprise agentic deployments.
- Support spans TypeScript, Go, Dart, and Python, making this a cross-stack infrastructure bet rather than a Gemini lock-in play.
Funding
Anthropic and PwC today deepened their alliance to roll out Claude Code and Cowork across PwC's global workforce, targeting enterprise transformation engagements it estimates represent a $2 trillion drag on pre-AI operating models, which means Anthropic is using a Big Four firm as a distribution channel at scale rather than through direct sales.
Mistral AI is in active discussions with European banks about a cybersecurity-focused model designed to identify weaknesses at speed and scale, currently in limited partner testing, signaling that Mistral is pursuing regulated-industry verticalization as its path to enterprise revenue rather than competing on general-purpose model benchmarks.
Case Studies
An internal memo from Microsoft EVP Rajesh Jha, reported today, directs engineering teams to transition from Claude Code to Copilot CLI, the first named enterprise rollback of an Anthropic product in favor of a competing internal tool.
- Claude Code had become widely adopted inside Microsoft over the past six months, making this a deliberate strategic reversal rather than a low-usage cleanup, and it comes the same day Microsoft shipped the Copilot desktop app.
- The memo frames Claude Code as a learning instrument that served its purpose, which is the language enterprises use when replacing a vendor they have benchmarked and now want to internalize.
Trending on X
- Anthropic overtakes OpenAI in enterprise The Ramp AI Index showing Anthropic at 34.4% vs OpenAI at 32.3% enterprise spending share is circulating widely, with builders treating it as a structural shift rather than a monthly fluctuation and speculating whether OpenAI's response accelerates GPT-5.6 to market.
- Scaling law and thinking tokens ceiling Ethan Mollick's post noting that adding thinking tokens continues to improve model performance on hacking, math, and science tasks with no visible plateau is sparking debate about whether test-time compute has become the dominant axis of model improvement over parameter scaling.
- Microsoft winding down Claude Code internally The leaked Rajesh Jha memo is generating sharp discussion about whether enterprise AI tool consolidation will always favor incumbents with distribution and whether Anthropic's land-and-expand motion is structurally vulnerable to Microsoft's platform control.
- SAP Concur AI transformation moment swyx's story about a developer publicly criticizing SAP Concur at a conference while a SAP employee in the audience responded by inviting them to advise on AI transformation for 6,800 employees is being shared as an illustration of how enterprise software incumbents are now actively recruiting their critics to accelerate AI rebuilds.
- Google I/O Gemini and Omni expectations With Google I/O confirmed for May 19-20, discussion is heating up around an unannounced video generator called Omni and a new Gemini version, with the community debating whether Google's open-source Apache 2.0 move on Gemma 4 today is a pre-I/O signal about a broader openness strategy shift.