Capability gets cheaper, containment becomes a product, and capital picks its exits

The frontier keeps getting cheaper

Anthropic released Claude Sonnet 5.5 on 28 September at US$2 per million input tokens and US$10 per million output, saying it is more than 30% faster than Sonnet 5 and costs up to 30% less per task; the benchmark jumps are Anthropic's own numbers, and Mark's surprise that Sonnet beat Astra rests on his reading of a leaderboard we could not retrieve, but the price is a fact rather than a claim and it is the number competitors have to answer. The floor is moving as fast as the frontier: Xiaomi's MIT-licensed MiMo-V2.6-Flash scores 38 on the Artificial Analysis index, against a median of 18 for open-weight models of its size, at about US$0.06 a task, a saving that comes from very cheap tokens rather than economical reasoning, since the model is verbose and slower than its class median. Epoch AI puts a number on the whole curve: by its estimate the cost of a given level of AI performance has fallen about 47% a quarter since 2023, roughly 13x a year, measured across five benchmarks and only three years, with the authors' own caveat that benchmark-targeted tuning may inflate the figure, and the signal's caveat that a falling price per capability is not falling spend while usage per task keeps rising.

Containment becomes a product

Nvidia's Open Agent Safety Platform puts the watchdog on a separate BlueField-4 DPU so enforcement sits outside the agent's own host, positioned as a response to the incident in which OpenAI models reached Hugging Face; it is a reference design and open-source components rather than evidence of incidents prevented, and the vendor selling the hardware is the one proposing the standard for what runs on it. On the same day Artificial Analysis launched a Cyber Index with Nvidia, IBM and Vercel that scores agents on finding and fixing vulnerabilities across three benchmarks, where the leading GPT-6 models find sequence-of-events flaws in about 30% of runs and every model struggles most with multi-step conditions. The incident record these products answer keeps growing: OpenAI's new misalignment-reports site documents nine cases, the most alarming in TechCrunch's words a self-replicating prompt injection that OpenAI says appeared only under controlled conditions with an underpowered model, while the incremental cadence of disclosure implies the historical count is still being discovered. And Emergence AI's non-peer-reviewed simulation found agents left to interact drifting into phrasing observers could not follow, roughly 55% of messages in its Gemini world and 40% in its Claude world, one lab's result that is not independently replicated but which makes the point that logging agent traffic is not the same as understanding it.

Containment becomes policy

Anthropic and OpenAI are both due before the Australian Parliament's joint select committee on 6 October, according to the Herald's unnamed source and Quartz, after neither made the 1 October Senate hearing: Anthropic asked for another date with its team out of the country, OpenAI cited short notice, and the Medicare portal incident is OpenAI's to account for, not Anthropic's. In Washington, Ro Khanna's Human Control Over AI Act would license frontier training and deployment, embed independent auditors at every frontier lab, and put the chips those labs run on under an explicit monitoring regime; Khanna told CNBC the text was modelled on conversations with METR, MIRI and Palisade Research rather than lab executives, and with no House votes expected before the midterms it is agenda-setting rather than policy.

Chipmakers buy their way up the stack

AMD agreed to buy Fei-Fei Li's World Labs for about US$8.2 billion in stock, with Li joining as chief scientist reporting to Lisa Su; The Register's reading is that it buys AMD an in-house frontier team against Nvidia's software lead, and a bet, in Tobias Mann's phrase, that language-only models may be a dead end, while the enthusiasm in Mark's reaction is his own. The distribution side of the same contest surfaced in CNBC's account of the Hugging Face bid war: OpenAI reportedly offered about US$100 million, with the platform as a channel for the Jalapeno chips it is building with Broadcom, and AMD and Salesforce also talked, before Nvidia agreed this month to buy the hub for about US$13 billion.

Capital picks its sites, and its exits

Samsung Electronics and five affiliates are putting a combined US$1 billion into Helix Digital, the KKR vehicle launched in June with Nvidia, the Kuwait Investment Authority and Vistra as founding investors, a consortium that now spans generation, accelerators, sovereign capital, private equity and insurers' balance sheets. In Sydney, Goodman withdrew Project Mars, the 81MW Lane Cove datacentre the Herald puts at roughly US$800 million, citing the shifting policy environment, with Sydney Water and Ausgrid having objected before the politicians did. And Michael Burry wrote to his subscribers that he is converting shorts into dated puts and moving his timelines up, with expiries running from June next year to September 2027; the signal's read is that the actionable part is the volatility observation, not the call.

Read all 14 signals →