Anthropic ships Sonnet 5.5 as cyber safeguards move down a tier
Anthropic's second Claude 5.5 model lands with frontier-class benchmark deltas at a mid-tier price — and the safety machinery that used to sit only on the top model. Meanwhile a chipmaker buys into the financing layer of the AI buildout, and Congress gets its most detailed AI-safety bill yet, with no path to a vote.
Anthropic released Claude Sonnet 5.5 on September 28, and the interesting part is not the score — it is that Sonnet 5.5 beats Anthropic's own top model on the hardest agentic evaluation at a fraction of the token cost. On Terminal-Bench 4.0, the new model scores 70.6% against 10.3% for Sonnet 5 and 66.4% for Opus 5.5, per Anthropic's published numbers, while GDPval-AA v2.1 puts it at 1,844 against Opus 5.5's 1,846 — a statistical dead heat. Pricing is $2 per million input tokens and $10 per million output tokens with 1M context and thinking always on, and Anthropic says it runs more than 30% faster and costs up to 30% less per task because it burns fewer tokens.
The tier inversion is the story. For most of 2026 the assumption has been that capability lives at the top of the price list and you pay for it; Sonnet 5.5 collapses the gap between the frontier and the workhorse, which matters more to buyers than any single benchmark record. Anthropic also put cyber safeguards and anti-distillation classifiers on a Sonnet model for the first time, with higher-risk cyber requests falling back to the older Sonnet 5, and it credits safety-first framing in its own launch copy. A caveat worth carrying: The Next Web notes Anthropic calls $2/$10 "the same as Sonnet 5," while its own June reporting had those as introductory rates due to expire August 31 — unresolved, so treat "no price increase" as a claim rather than a fact. Haiku 5.5 is promised "in the coming weeks," which would complete the family ahead of Anthropic's expected IPO push.
Samsung Electronics and five affiliates will put $1 billion into Helix Digital Infrastructure, the AI infrastructure company KKR launched in June with Nvidia, the Kuwait Investment Authority and Vistra as founding investors. Samsung Electronics carries $500 million of that, with Samsung C&T, SDS, SDI, Life Insurance and Fire & Marine Insurance funding the rest, according to Samsung's announcement. Helix is led by Adam Selipsky, the former AWS chief executive, and KKR says the vehicle already holds more than $10 billion in long-duration commitments; KKR has now deployed over $75 billion across digital infrastructure and power globally, roughly $9 billion of it in Korea.
Read it as vertical integration into the financing layer. Samsung is not just selling memory and cooling into the buildout — it is buying equity in the vehicle that decides which data centers and power plants get built, which is a different kind of leverage. Two honest limits: this is a commitment, not a closed and priced round, and no valuation, instrument or specific project was disclosed.
Representative Ro Khanna will introduce the Human Control Over AI Act, which would ban models that recursively self-improve or autonomously modify their own core objectives, containment or shutdown controls until federal guardrails exist and a regulator approves the activity. CNBC, which reviewed a summary exclusively, reports the bill would create an FDA-style federal agency with a licensing system for frontier model training and deployment, independent auditors embedded inside labs and reporting directly to the agency, security standards covering sandbox testing and kill switches, controls on the advanced chips frontier labs use, mandatory liability insurance, and criminal penalties for employees who disable safeguards. Khanna told CNBC the bill draws on conversations with METR, MIRI and Palisade Research rather than lab executives, and cited "a civilizational extinction risk" alongside misuse risk as the two threats to take seriously.
The wording matters as much as the content: as of September 29 there is no bill number, no text and no reported co-sponsors, so say "will introduce," not "introduced." Congress is also expected to leave Washington for the midterms without voting on any AI bill. Compare the ban-with-fine-print approach to the broad call we covered this month — Ezra Klein calls for a ban on self-improving AI — the difference being that Khanna's version names the licensing mechanism a ban would need to be enforceable.
What to watch: whether Haiku 5.5 arrives on the same price curve, and whether Khanna's bill ever gets a number.
Sonnet 5.5 beating its own flagship at a fifth of the price — do models need a top tier at all anymore? Tell us in the comments.
