Anthropic cuts Fable's price and gates the unrestricted version
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September. They are the same underlying model shipped under two different safeguard regimes. Fable 5.1 is generally available on the Claude API and through AWS, Google Cloud and Microsoft Azure. The price of cache reads falls 75%, to $0.25 per million tokens; Anthropic says that cuts the cost of a typical workload by around 25% against Fable 5, and up to around 45% for context-heavy agentic work where cache reads dominate the bill. Input and output pricing is unchanged at $10 and $50 per million tokens. On Anthropic's own evaluations Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1 against 24.7% for Fable 5 and 29.0% for Opus 5, and 55.8% on Terminal-Bench 4.0.
Mythos 5.1 — the same weights with more permissive cybersecurity and life-sciences safeguards — is reachable only through two vetted programmes and inside Anthropic's own Claude Security product, and Anthropic states it is “currently only available to a set of US organizations,” with wider access being coordinated with the US government. So the commercially available model got cheaper and the unrestricted one did not become available in Europe. One thing did move the other way: Anthropic's new Enterprise Frontier Safeguards store customer data in infrastructure the customer controls rather than Anthropic's, with human review defaulting to the customer. It rolls out in phases from this autumn, and eligible customers can run Fable 5.1 with zero data retention until it arrives. Data residency and capability access are now being unbundled, and they are moving in opposite directions.
Google ships its third Flash in six weeks, with a cyber sibling of its own
Google rolled out Gemini 3.8 Flash on 2 September, three weeks after 3.7 Flash and, in its own words, its third Flash release in six weeks. Google attributes the gains to the model simply working harder — more reasoning steps, more iterative tool calls, and more tokens spent at higher effort levels. Google reports 54.9% on HLE-Verified. Introductory pricing is $0.75 per million input tokens and $3.75 per million output, the same rate 3.7 Flash carried, and it holds until 31 December. On 1 January 2027 it goes to $1.50 and $7.50 — a doubling, per Google's own footnote. Model that as a price with an expiry date, not a price.
Alongside it, Gemini 3.8 Flash Cyber went to trusted partners under a new Fairwind Program. Google's Chrome security team reports it produced 2.6 times more correct patches to Chrome vulnerabilities than the best commercial models, which are much larger; Wiz, now part of Google, reports 7.5–9.7% higher recall on its internal penetration-testing benchmark at 2.3–5.2 times lower cost. Both readings are in-house. Two labs, two days, and both put their strongest security model behind a list you have to be admitted to.
Dell's AI order book is now twice a quarter's revenue
Dell reported second-quarter fiscal 2027 results on 1 September: record revenue of $47.0bn, up 58% year on year, of which $16.4bn came from AI-optimised servers. It booked a record $60.9bn of AI server orders in the quarter and closed with a $95.0bn backlog, up from $51.3bn three months earlier. Full-year revenue guidance rose by $25bn to roughly $192bn, with about $74bn of that expected from AI servers.
The backlog is the number to hold onto. It is now roughly double a full quarter's revenue across Dell's entire business, which means most of the AI hardware that has been paid for has not been built, shipped or racked. Anyone sizing inference capacity in Europe over the next eighteen months is queueing behind that, whether they buy the machines or rent the tokens.
Quick Hits
- Anthropic ships an AI Act text watermark and a detection API not everyone can have. — Models released after 2 August 2026 now carry an invisible statistical watermark, under the Code of Practice on Transparency of AI-Generated Content that Anthropic signed in July, alongside about 190 organisations in total according to the European Commission. The detection API is in private preview, available to regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises with their own verification obligations. Provenance you can check is becoming a permission, not a feature. (Anthropic)
- A third of firms have skipped a software purchase because they could just build it. — McKinsey's State of AI 2026 survey, published on 25 August, finds 32% of respondents decided against buying at least one software product or feature because agentic coding tools let them build it internally — 41% in technology and 39% among healthcare payers and providers. Among the roughly 6% McKinsey classes as high performers, it is close to half.
- Cisco puts a personal agent in front of 90,000 employees. — MyAgent is rolling out company-wide, running supervised autonomous workflows across Outlook, Webex, Jira and SharePoint. It sits on Circuit, which Cisco describes as a “secure, governed, multi-model agnostic AI platform”; Cisco says MyAgent “only accesses approved models, approved systems, and enterprise-appropriate data pathways.” Agentic interactions on Circuit grew nearly 350% quarter over quarter, on Cisco's numbers.