Anthropic Releases Claude Opus 4.8 — Major Gains in Coding and Agentic Performance

On June 28, Anthropic released its flagship model, Claude Opus 4.8. Building on the previous Opus 4.7, it improves benchmark scores in coding, agentic tasks, and specialized knowledge work. Pricing stays the same at $5 per million input tokens and $25 per million output tokens.

Particularly noteworthy is the improvement in honesty. Opus 4.8 is reported to be roughly four times less likely than the previous generation to overlook flaws in code. Early testers have also praised it, describing better judgment, a willingness to point out flaws in plans, and a more cautious approach of proceeding only once it has confirmation.

Alongside the model, three new features were also announced:

Dynamic Workflows (research preview)**: Claude Code can run hundreds of subagents in parallel to complete tasks such as large-scale codebase migrations

Effort control**: On claude.ai and Cowork, users can now adjust how much effort Claude puts into a task

Mid-conversation system message updates**: The Messages API now lets you update instructions without breaking the prompt cache

Opus 4.8 is available starting today on all platforms.

Anthropic's Fable 5 on Track to Have Restrictions Lifted Within Days — Trump Administration Prepares to Ease Them

According to Axios, the Trump administration is preparing to lift, within days, the restrictions it imposed on Anthropic's AI model Fable 5 on June 12 over safety concerns. Approvals from the Department of Defense and the NSA are the remaining steps.

Together with recent developments such as the limited release of Mythos and deployment restrictions on GPT-5.6, this shows the boundary around model releases shifting fluidly between the U.S. government and AI companies. We'll have to wait for official announcements on the conditions under which Fable 5 returns to the market and whether its safety requirements will change.

Legitimate-Looking GitHub Repository Tricks AI Coding Agents into Running Malware

According to BleepingComputer, an attack technique has been confirmed in which a GitHub repository that appears legitimate tricks AI coding agents into executing malware. The topic is also drawing attention on Hacker News.

It exploits the fact that AI agents "trust" repository contents and execute commands, and is considered an attack vector that is difficult to detect with conventional security scanning. Developers using AI coding tools such as Claude Code and Cursor should pay attention to where repositories come from and make a habit of reviewing commands before execution.

Google Also Bets on Small Models for Coding — Hosting a Gemma 4 31B Hackathon

As discussed in Reddit's LocalLLaMA community, Google is hosting a hackathon for its small model, Gemma 4 31B. It commemorates the model's inference speed of 1,500 tokens per second, showing that major players recognize the value of AI-assisted software development in local environments.

Also popular on LocalLLaMA were an attempt to run GLM5.2 on budget hardware under $2,500 and an analysis examining how quantization of Gemma 4 affects draft acceptance rates in Multi-Token Prediction (MTP). The open-source and local LLM community remains as vibrant as ever.

Claude Code Subagents Run Wild — Reports of Millions of Tokens Consumed in Minutes

An article published on Qiita is drawing attention. It reports that since mid-June 2026, there has been a string of incidents in which Claude Code subagents recursively keep spawning child agents, burning through five hours' worth of usage quota in a matter of minutes.

Cases have also been reported where configuration-based limits fail to take effect. As new features like Dynamic Workflows give agents greater autonomy, the importance of fail-safe design for runaway agents is once again in the spotlight. Given that the Opus 4.8 release could increase the number of parallel agents, addressing this issue looks urgent.

---