1. Anthropic Surpasses OpenAI in Enterprise Adoption
Data from the May 2026 Ramp AI Index indicates a significant shift in enterprise AI spending, with 34.4% of participating businesses now paying for Anthropic services compared to 32.3% for OpenAI. While OpenAI maintains a massive consumer base, Anthropic's business adoption has quadrupled over the past year, driven largely by the popularity of its agentic coding tools. This shift highlights the growing enterprise preference for specialized agentic workflows over general-purpose consumer models.
- • Anthropic's business adoption reached 34.4% in April, surpassing OpenAI's 32.3%.
- • Anthropic's business adoption has quadrupled over the past year.
- • Claude Code is estimated to author 4% of all public GitHub commits.
This milestone signals a potential turning point in the enterprise AI market, suggesting that Anthropic's focus on agentic coding tools is successfully capturing business budgets that were previously dominated by OpenAI.
2. Anthropic Introduces Agent SDK Credits
Anthropic has reversed its April 2026 ban on using Claude subscriptions for third-party autonomous agents. Subscribers will now receive fixed, non-rollover monthly Agent SDK credits ranging from $20 to $200, depending on their plan. These credits are billed at API rates and are separate from standard chat usage, addressing previous compute inefficiencies caused by agents bypassing prompt caching mechanisms. The policy will be fully implemented by June 15, 2026.
- • Agent SDK credits are fixed, non-rollover monthly allocations.
- • Programmatic usage is billed at API rates, separate from interactive chat limits.
- • The policy change addresses technical inefficiencies related to prompt caching.
This change provides a clear path for developers to integrate third-party agents with Claude while allowing Anthropic to manage compute costs and infrastructure stability.
3. Fastino Labs Releases GLiGuard Moderation Model
GLiGuard is an encoder-based safety moderation model that treats moderation as a text classification task rather than a generation task. By focusing on classification, the model achieves significantly higher throughput and lower latency than larger decoder-only models. It evaluates four safety dimensions—safety classification, jailbreak detection, harm category detection, and refusal detection—in a single forward pass, making it suitable for production-grade AI guardrails.
- • 300M parameter model optimized for text classification.
- • Achieves 16.2x higher throughput than larger decoder-only models.
- • Evaluates four safety dimensions in a single forward pass.
For developers building AI applications, GLiGuard offers a lightweight, efficient alternative to larger models for real-time safety moderation, reducing both latency and operational costs.
4. Ardent Launches Postgres Sandboxes for AI Agents
Ardent allows developers to spin up database sandboxes in under six seconds, even for terabyte-scale datasets. The platform uses a replication stream and copy-on-write technology to create isolated environments that mirror production databases. It includes features for granular access control, credential protection, and automated PII redaction, enabling AI agents to safely test database interactions in a production-like environment.
- • Sandboxes spin up in under 6 seconds.
- • Compatible with any hosted PostgreSQL database.
- • Includes built-in PII redaction and access control.
Testing AI agents against real database schemas is often slow and risky; Ardent's approach enables faster, safer iteration for agentic workflows that require database access.
5. Modal Optimizes Serverless GPU Scaling
Serverless computing is increasingly vital for AI inference, where workloads are highly variable. Modal has optimized its infrastructure to spin up new replicas rapidly, reducing the time required to scale inference capacity from multiple kiloseconds to just tens of seconds. This improvement allows developers to handle unpredictable traffic spikes more effectively without maintaining idle GPU capacity.
- • Scaling time reduced from kiloseconds to tens of seconds.
- • Optimized for variable inference workloads.
- • Reduces the need for over-provisioning GPU capacity.
Rapid scaling is critical for cost-effective AI inference; Modal's infrastructure improvements allow developers to build more responsive applications while minimizing compute waste.
6. Cactus Needle: A 26M Parameter Open Model
Cactus Needle is a Simple Attention Network (SAN) distilled from Gemini 3.1. With only 26 million parameters, it is designed to run locally on Mac or PC hardware, achieving speeds of 6,000 tokens per second for prefill and 1,200 tokens per second for decode. The model weights are fully open, aiming to redefine AI capabilities for edge devices like phones and wearables.
- • 26M parameter model distilled from Gemini 3.1.
- • Capable of local fine-tuning on consumer hardware.
- • Achieves high token throughput for prefill and decode.
This model demonstrates the potential for highly efficient, small-scale models to perform specific tasks locally, bypassing the need for cloud-based API calls for simple interactions.
7. Notion Launches Developer Platform for AI Agents
Notion is expanding its productivity software into the agentic space with a new developer platform. This update allows teams to connect external data sources, custom code, and AI agents directly into their Notion workspaces. The move signals Notion's intent to become a central hub for agentic workflows, enabling users to automate tasks and interact with data without leaving the application.
- • New platform supports integration of external data and custom code.
- • Enables teams to embed AI agents directly into Notion workspaces.
- • Strategic expansion into agentic productivity software.
By turning its workspace into an agent hub, Notion is positioning itself as a primary interface for enterprise AI, allowing developers to build custom agentic tools that live where the work happens.
8. UK AI Safety Institute Reports Rapid Cyber Capability Growth
The UK AI Safety Institute (AISI) reports that AI models are rapidly advancing in their ability to complete complex cyber tasks. Testing on models like Claude Mythos Preview and GPT-5.5 shows a capability doubling time of approximately 4.5 months, an acceleration from previous estimates. These models are now capable of solving complex cyber ranges, with performance appearing limited primarily by token usage rather than inherent ability.
- • Cyber capability doubling time is estimated at 4.5 months.
- • New models are successfully solving complex cyber ranges.
- • Performance appears limited by token usage rather than inherent ability.
The rapid advancement of AI cyber capabilities necessitates new security frameworks and monitoring strategies for developers building agentic systems that interact with sensitive infrastructure.