Anthropic Introduces "Agent SDK Credits," Resuming Third-Party Tool Access

On May 13, Anthropic announced that it will introduce "Claude Agent SDK credits" for Claude's paid plans. The move addresses excessive resource consumption by third-party apps, adopting a fixed-budget system separate from the subscription allowance. As a result, Claude usage via third-party tools is expected to resume on June 15.

Until now, third-party usage had been restricted because resource consumption was difficult to control. While the new credit system gives developers clearer budget management, it effectively caps heavy users, and some developers have voiced concerns.

SoftBank Group Declares Its AI Investment Strategy Is "OpenAI-Centric"

At the earnings briefing for the fiscal year ending March 2026, SoftBank Group CFO Yoshimitsu Goto explicitly stated that the company is "currently thinking OpenAI-first" when it comes to AI-related investment targets. The remark comes amid reports of a stake in Anthropic as well, signaling a strategic approach to resource allocation.

DAgger Makes a Comeback in the Era of LLM Agents — a 4B Model Beats 8B Systems

The paper "Revisiting DAgger in the Era of LLM-Agents," featured on Hugging Face Daily Papers, applies the classic iterative learning method DAgger (Dataset Aggregation) to post-training of LLM agents, organizing SFT, RL, OPD, and DAgger within a unified framework. Notably, a 4B-parameter SWE agent trained with this recipe outperforms published 8B systems, and the 8B agent approaches 32B-scale performance. It is a significant result demonstrating that effective agents can be built even with smaller models.

SWE-ZERO 12M: 112B Tokens of Execution-Free Agent Coding Data

"SWE-ZERO 12M Trajectories," which has appeared on Hugging Face's trending datasets, is an execution-free agent coding dataset spanning 112 billion tokens. It is generated using only standard POSIX commands, without requiring test execution in Docker containers, and covers more than 120,000 PRs, over 3,200 repositories, and 16 languages. At 5.7 times the size of AgentTrove (19.4B tokens), the previous largest dataset, it is intended as intermediate training data for agents.

AgentLens: The "Lucky Pass" Problem in SWE Agent Evaluation

AgentLens, released by a Microsoft research team, tackles the problem of SWE agent evaluation being dominated by a single binary signal: whether the patch passes tests. Analyzing 2,614 OpenHands trajectories, the team identified that 10.7% of passing trajectories were "lucky passes" — chaotic processes involving regression cycles, blind retries, and missing verification. Lucky-pass rates vary from 0.5% to 23.2% depending on the model, and ranking by quality scores can shift standings by up to five places.

The Cognitive Load of AI Use: Research Suggests "Thinking Degrades After Just 10 Minutes"

A discussion about how AI use affects human cognitive function has been drawing attention on Reddit's MachineLearning community. Research reported by Wired suggests that as little as 10 minutes of AI use can adversely affect problem-solving ability. A BCG study also found that using AI effectively, when combined with other tasks, multiplies workload two- to threefold. The reality of this trade-off — less burnout but more cognitive overload — is resonating with practitioners.

MTP + TurboQuant: 40% Faster Local Inference with Qwen 3.6

A benchmark reported in the LocalLLaMA community shows that combining Multi-Token Prediction (MTP) with TurboQuant improved local inference of Qwen 3.6 27B from 21 tok/s to 34 tok/s. The results were achieved on a MacBook Pro M5 Max, with a 90% acceptance rate. The llama.cpp patch and quantized models have already been released.

Cloud Infrastructure Market Grows 35% Year Over Year — Fueled by the AI Boom

According to Synergy Research Group, the global cloud infrastructure market recorded rapid 35% year-over-year growth in Q1 2026. Rising AI workloads are seen as the primary driver. The shares of the top three providers reportedly show little change.

OpenAI Launches a Codex Switching Campaign — Two Months Free for New Enterprise Customers

OpenAI has launched a campaign granting two months of free access to its coding agent Codex for enterprises signing new contracts. The move explicitly targets customers switching from competitors' existing products, underscoring how competition in the enterprise AI coding market is intensifying.

---