The Footnote Layer: Opus 5.5 Leads, With Asterisks Attached
A quiet news day exposes how much of the frontier picture rests on vendor-reported benchmarks with their own safeguards in the loop. Meanwhile the chip export rules are older and broader than yesterday's brief said.
Capability & Integration
- Anthropic’s Claude Opus 5.5, released September 22, is reported by Office Chai to beat OpenAI’s GPT-6 Astra on most of Anthropic’s own benchmarks: Terminal-Bench 4.0 at 66.4% versus 57.9%, and Humanity’s Last Exam with tools at 67.7% versus 57.2%. Astra leads on two evaluations, including the Zapier-run AutomationBench. These are company-reported numbers, and Terminal-Bench 4.0 carries a standard error of roughly ±2.6 points.
Review
Terminal-Bench 4.0 66.4% vs 57.9% — the Opus 5.5 score was run at xhigh effort while Astra’s was at high effort, and both come from Anthropic’s comparison table, so the “lead” is not like-for-like (one third-party guide lists an independent 59.6% for Opus 5.5) — verify and add the effort-setting caveat
- Let’s Data Science reports that Anthropic’s own safeguards intervened in several evaluations, routing some cybersecurity and biology-adjacent tasks to older Claude models. Scores partly reflect the safety layer, not only the model.
- Update: yesterday’s brief said no frontier release surfaced. That was incomplete: Opus 5.5 is three weeks old, and the Gemini 4 timing flag stands, with no Google announcement found.
Unverified
The price cut (sources say between 20% below Opus 5 and $4/$20 per million tokens) and the Artificial Analysis index lead come from secondary summaries; I did not reach Anthropic’s launch page.
Robotics
No significant developments today. The newest dated deployment items I could verify are from August or earlier, and nothing new on Humanoid Ltd., Schaeffler or Siemens advanced the picture from yesterday.
Hardware & Supply Chain
- Update: yesterday’s brief named January 15 as the last confirmed export-control change. A later one exists: on May 31 the Commerce Department’s BIS issued guidance enforcing license requirements on advanced chips going to entities headquartered in China but located abroad, per Reuters via Yahoo Finance. BIS framed it as clarifying requirements in place since 2023, and Nvidia said it changes nothing for the company. An industry source put the scale of prior leakage in the hundreds of thousands of chips.
- The trigger moved from where a chip is delivered to where the buyer’s parent is headquartered, per the same report. I have not seen the guidance text, and the 2025 Malaysia and Thailand draft rule’s final status is unconfirmed.
Environmental & Cultural Impact
- A UN University report puts 2025 data center electricity use at about 448 TWh, rising to 945 TWh by 2030, with roughly a fifth of the 2025 total attributed to AI, per Earth.org. Its water projection, covered by TBS News, is a footprint comparable to the domestic needs of 1.3 billion people by 2030. The report dates to June, so this is context, not a new release.
- Amazon reports 2.5 billion gallons of data center water use in 2025 and 0.12 liters per kWh, which it calls seven times better than the industry average, per Feedbagel. Critics say the comparison sets all of Amazon’s facilities against more power-intensive AI-specific ones.
Unverified
The Amazon figures come from an aggregator summary, not Amazon’s disclosure.
AI in the Wild
“Brainrot” characters such as Tralalero Tralala, an absurdist AI-generated figure, are described by Is4.ai as having accumulated billions of cumulative views. I could not find a dated, verified viral item for this week, so this is a standing genre, not a single clip. Its persistence suggests AI content now circulates as ordinary, low-stakes filler.
Unverified
The view counts come from one trend blog with no traceable data source.
Takeaway
Takeaway
The leaderboard is being written by the vendors who run it and, in some cases, by the safeguards they bolt on, so the most honest reading of “who is ahead” is a margin about the size of the error bars.