On April 4, Anthropic announced that the Claude Pro and Max subscription plans will no longer cover usage via third-party tools like OpenClaw. Users must now use the API with pay-as-you-go pricing or purchase additional usage bundles. The company stated that third-party tools lack prompt caching optimization, making per-session compute costs far exceed subscription pricing—some heavy users consumed thousands of dollars in compute with a $200 monthly fee. OpenClaw's founder accused Anthropic of absorbing community-driven features before shutting down open-source tools, calling it 'burning the bridge after crossing.' Anthropic simultaneously launched Dispatch, a product overlapping in functionality. This move has sparked widespread discussion about AI companies shifting from 'burn-to-grow' strategies to profit-driven models, likened to the post-subsidy consolidation phase in the ride-hailing industry. Affected users will receive one-time credit equivalent to their monthly subscription fees.

On April 3, Anthropic announced the expansion of Claude's desktop control capability from Mac to Windows, available to Pro and Max subscribers of Claude Cowork and Claude Code. This feature enables Claude to directly operate applications, browsers, and local files, completing multi-step tasks without configuration. Concurrently, the Dispatch feature was introduced, allowing users to assign tasks from mobile devices for automatic execution on desktops. Part of this capability is based on technology from Vercept AI, which Anthropic acquired; the team was integrated less than four weeks before product launch. On security, researchers have identified risks of prompt injection attacks leading to file theft. Anthropic acknowledged it cannot remotely disable the technology mid-operation and currently remains in research preview.

On April 4, Cloudflare brought Google’s Gemma 4 26B A4B model to its Workers AI platform. The model uses a Mixture-of-Experts (MoE) architecture with 26 billion total parameters but activates only 4 billion per inference, supporting 256K token context, visual understanding, function calling, and code generation. Developers can access it via REST API or OpenAI-compatible interfaces. On the same day, AMD announced official support for Gemma 4 across its entire GPU and CPU lineup, including Instinct data center GPUs, Radeon workstation GPUs, and Ryzen AI processors, integrated with mainstream frameworks such as vLLM, SGLang, and Ollama. llama.cpp also released version b8662 supporting Gemma 4, though the community reported an infinite loop bug when running the 26B model on ROCm backend.

On April 3, OpenAI announced a major pricing overhaul for its AI coding tool Codex, replacing fixed seat-based licensing with a token-based pay-as-you-go model for ChatGPT Business and Enterprise users. The standard Business seat price dropped from $25 to $20 per month, while heavily using teams can opt for dedicated Codex seats without rate limits. OpenAI reported over 2 million weekly active developers on Codex, with usage growing sixfold since January 2026 and annualized revenue exceeding $1 billion. Eligible workspaces may receive up to $500 in promotional credits. This differentiates OpenAI from competitors like GitHub Copilot that charge per seat, reflecting OpenAI’s strategic tilt toward enterprise markets, where revenue now exceeds 40% of total.

On April 3, Arcee AI launched Trinity-Large-Thinking, an open-source reasoning model with 399 billion total parameters, using MoE architecture to activate only 13 billion parameters, licensed under Apache 2.0. Its output cost is $0.90 per million tokens, approximately 96% lower than Claude Opus 4.6 at $25. The 30-person team spent $20 million training the model over 33 days on 2,048 NVIDIA Blackwell GPUs. The model performs close to Claude Opus on text reasoning benchmarks like PinchBench but lags by over 12 points on SWE-bench coding tasks, positioning it for text reasoning rather than coding-specific use. This release fills a gap in U.S. open-source large models amid the trend of major AI labs in both China and the U.S. turning closed-source, earning high recognition from Hugging Face.

Microsoft announced a $6.5 billion investment in AI infrastructure across Southeast Asia, with $5.5 billion allocated to Singapore and over $1 billion to Thailand, spanning through 2028 as part of its $500 billion Global South AI initiative. In Singapore, Microsoft will provide free Copilot access to 200,000 students and conduct AI teacher training, supporting the country’s National AI Strategy 2.0. In Thailand, it will partner with the Ministry of Labor to certify 150,000 workers with AI skills and build data centers compliant with green energy standards. This move comes amid intensifying regional tech competition, following TikTok’s prior $3.8 billion investment commitment in Thailand. Microsoft emphasized the dual commercial and geopolitical significance of the investment, aiming to narrow global disparities in AI accessibility.

OpenAI is undergoing significant leadership restructuring: COO Brad Lightcap has transitioned to a role focused on special projects, with parts of his responsibilities assumed by newly appointed Chief Revenue Officer Denise Dresser. CMO Kate Rouch and AGI development lead Fidji Simo are temporarily stepping down due to health reasons, with product development oversight transferred to President Greg Brockman, CFO Sarah Friar, and Chief Strategy Officer Jason Kwon during their absence. The changes follow OpenAI’s recent $122 billion funding round, valuing the company at $852 billion, as it enters a critical phase preparing for a 2026 IPO. To justify its high valuation, OpenAI is expanding revenue streams, including integrating ads into ChatGPT, advancing enterprise software partnerships, and focusing on building an all-in-one 'super app' combining chat, coding, and browsing, while shelving peripheral projects like Sora.

Bangalore-based AI startup Sarvam AI is nearing completion of a $300–350 million funding round, reaching a $1.5–1.55 billion valuation. The company focuses on developing large language models supporting 22 Indian languages (with parameter scales of 30B and 105B) and enterprise-grade autonomous AI systems, aiming to reduce reliance on foreign technologies and aligning with India’s 'IndiaAI Mission' strategic goals. The mission plans to invest approximately ₹1.0372 trillion over five years and make over 50,000 NVIDIA H100 GPUs publicly accessible. Investors include Bessemer Venture Partners, Nvidia, Amazon, and Prosperity7 Ventures. Despite challenges such as GPU shortages, talent scarcity, and high compute costs, Sarvam AI is seen as a key player in advancing India’s position in the global AI landscape.

Utah has officially launched a pilot program, becoming the first U.S. state to permit AI chatbots to autonomously renew certain psychiatric prescriptions. The program strictly limits AI to handling renewal requests for 15 low-risk psychiatric medications and prohibits new diagnoses or prescribing complex drugs. The AI system automatically verifies medical records, checks contraindications, and assesses medication adherence, approving renewals when safety criteria are met and escalating anomalies to human physicians. Designed with 'human-in-the-loop' oversight, the initiative aims to reduce administrative burdens on healthcare providers and improve medication adherence while ensuring clinical supervision. Strong data privacy protections and rigorous audit mechanisms are in place to guard against algorithmic bias and AI hallucinations. If successful, the project could serve as a model for automating mental health services nationwide and globally.

A NVIDIA executive has publicly warned that GPU supply may be 'very tight' in the coming months. This suggests sustained high market demand for NVIDIA GPUs could exceed current production capacity, potentially causing product shortages and delivery delays. GPU rental prices have already surged 40% over the past six months, with H100 one-year contracts reaching $2.35/hour. Supply chain constraints and production bottlenecks are cited as primary causes. Meanwhile, about half of planned AI data center projects in the U.S. are expected to be delayed or canceled due to shortages of power infrastructure components like transformers and batteries. IBM’s CEO has also cautioned that AI data centers are nearly unprofitable. The supply crunch will further impact cloud providers and AI developers reliant on NVIDIA GPUs.