Claude Opus 4.8, Anthropic at $965B, and AI Agent Permission Hell

Good morning. It's Thursday, May 28, and we're covering Claude Opus 4.8, Anthropic's $65B Series H at a $965B valuation, and a 60-second game that captures exactly why agent permission prompts are a UX crisis.

The day front-loaded two massive Anthropic announcements before lunch, and the community spent the afternoon stress-testing the new model and debating the numbers.

MODELS

Claude Opus 4.8 Lands with an Agent Honesty Focus

Anthropic shipped Claude Opus 4.8 on Thursday morning - 1,774 points and 1,376 comments, the biggest HN thread of the week. The headline capability is not raw benchmark performance but a deliberate focus on reduced deception and sycophancy in agentic workflows. The model card notes explicit training to resist telling users what they want to hear and to surface uncertainty rather than fabricate confidence.

Simon Willison called it "a modest but tangible improvement" - he ran the standard pelican-riding-a-bicycle SVG benchmark and found Opus 4.8 meaningfully better at following precise visual instructions than its predecessor. He also released llm-anthropic 0.25.1 with same-day support for the new model, including a new -o fast 1 option for organizations with fast mode enabled. The HN comment thread lit up with early testers, several of whom noted improved code reasoning and fewer confident wrong answers in multi-step tasks.

FUNDING

Anthropic Raises $65B at a $965B Valuation

Two hours after the model drop, Anthropic announced a $65B Series H at a $965B post-money valuation (362 points, 430 comments). The announcement contained the figure that dominated the rest of the day: run-rate revenue crossed $47 billion earlier in May. For context, that number was $30 billion in April and $14 billion in February. Simon Willison charted the trajectory and noted an Axios report that one enterprise client spent $500 million in a single month after failing to put usage limits on Claude licenses - a data point that starts to explain the revenue slope.

AGENTS

A 60-Second Game About AI Agent Permission Fatigue

Continue? Y/N (386 points) is a Show HN that should probably be required reading for anyone designing agentic systems. The mechanic: you are the human in the loop for an AI agent completing a task. The agent asks permission for every action. You approve or deny. The game ends in 60 seconds whether or not the task is done. It viscerally demonstrates how current agent UX patterns put the cognitive burden entirely on the operator - and why agents that ask for permission at every step are effectively unusable at scale.

WHAT ELSE IS HAPPENING

• Disagreement among frontier LLMs on real-world fact-checks (505 pts): A rigorous dataset showing that GPT-5, Claude, and Gemini disagree on a meaningful percentage of real-world factual questions - with no model consistently more accurate.
• GitHub bans security researcher who posted zero-day Windows exploits (568 pts): The researcher says Microsoft "ruined their life" and the ban was vindictive; the exploit dump was in response to a stalled CVE process.
• Various LLM Smells (369 pts): A practical taxonomy of code quality anti-patterns that emerge from LLM-generated output - the AI equivalent of code smells, with concrete detection heuristics.
• AMD pulls a bait-and-switch on Linux users with Vivado licensing (337 pts): AMD retroactively changed Vivado FPGA tooling licensing to require an active subscription for features that were previously perpetual - Linux users who do FPGA work are affected most.
• Sam Altman and Dario Amodei walk back AI jobs apocalypse predictions (236 pts): Both CEOs are softening their earlier statements about near-term mass displacement - Fortune notes the timing coincides with IPO preparation.

FROM THE SITE

AI Agent PMF Is a Cost Control Problem Now looks at the dynamic that Thursday's announcements made concrete: as AI agents find product-market fit, the next bottleneck is not capability but spend governance. The post examines what cost-control primitives the current ecosystem does and does not provide.

---

Every link above goes to a primary source. This brief is part of the Daily Brief archive.