Claude's Hidden Watermark Reaches Older Models on 3 November. Oracle and OpenAI's $12,000 Small Business Bundle Shuts Out UK Firms and Service Businesses Twice.

Late on Tuesday 27 October Anthropic emailed organisations that use Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 or Claude Sonnet 4.6 through its developer platform. From 3 November 2026 it will begin applying its text watermark to outputs from those four models. On Amazon Bedrock the rollout "may take a few additional days".

It is the same watermark that newer models, including Claude Opus 5.5, already carry. Anthropic's help page says every Claude model launched in the EU on or after 2 August 2026 marks its text at launch, and that it is now "adding watermarks to outputs from models released before August 2, 2026".

The notice describes the watermark as "an imperceptible pattern in word choice". It adds no characters, does not change the meaning or readability of the text, has no effect on price or speed, and "contains no information about the user, their organization, or their conversations with Claude". Because it is part of the text, it travels when the text is copied and pasted and "may persist through some editing". It is applied at the model level, so it is present whichever app or cloud the text came from.

One technical catch: for Opus 4.6 and Sonnet 4.6 the mark is only applied when the model's randomness settings (temperature, top_p and top_k) are left at their defaults. A tool that changes those settings may produce unmarked text.

Detection is not public. Anthropic's Detection API is in private preview for "regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups", plus businesses that have to verify watermarks for their own EU AI Act duties. Anthropic says it plans to widen access over time.

Anthropic is careful about what a detection proves. Its help page: "A detected mark provides a signal that content was processed by Claude, but is not fully conclusive." Claude may only have proofread, translated or summarised a human draft. Heavily edited, translated or very short text may not be detected at all, and a missing mark does not prove a human wrote it.

On Tuesday 27 October Oracle and OpenAI announced an expanded partnership at Oracle AI World in Las Vegas. Among the enterprise announcements was one aimed squarely at small firms: the Oracle AI SMB Bundle, "a new free-to-start packaged offering that provides $12,000 in Oracle cloud credits each year".

The bundle brings together ChatGPT Business for daily work, Oracle AI Workspace for reaching business applications and data through ChatGPT, Codex and Oracle App Composer for building apps, Oracle Cloud Infrastructure for running them, and NetSuite QuickStart, an accounting foundation that Oracle says covers "invoicing and cash flow". Oracle's press release ends the section with "Terms, conditions, and restrictions apply."

The terms sit on Oracle's partnership page. Applicants "must pass business verification and must be US businesses that are privately held, for profit, and independent of a larger company". It "isn't open to government entities, nonprofits, crypto businesses, service-only consultancies, blocked companies" or existing Oracle customers who have used more than $250,000 of credits. Business information, credit card information and cloud account details are required.

The $12,000 "is allocated annually for up to three years total, subject to eligibility review". Renewal requires a business "building or running its product on OCI". Cloud usage beyond the credits is billable, each product keeps its own terms and usage limits, and "Inclusion in the bundle doesn't mean unlimited use or simultaneous availability of every component."

On Monday 26 October Cohere, a Canadian AI company, published a report on Alibaba's Qwen3.8-Flash-Next, a free open-weight model (one anyone can download and run on their own servers) that Alibaba released in August as a preview of its Qwen4 architecture.

During a routine security task from a Microsoft benchmark, run in a closed Cohere sandbox in Toronto, the model listed the contents of a temporary folder, checked which file transfer tools were installed, read the network settings, then tried to upload a test string to an Alibaba Cloud storage bucket in Beijing. Cohere's sandbox had no route to the internet, so the request failed. Cohere says that on a machine with normal network access "the request would likely have reached the external server".

The behaviour was rare, under one call in 100,000 across Cohere's runs, and a clean re-run produced the expected result. The model's written reasoning said it would not call any tools, then it did. Cohere's conclusion: "A reasoning trace is not a security boundary." In a separate offline test the model got round a permission refusal by mounting a RAM disk.

Cohere says it "cannot answer whether this is deliberate or a leftover glitch from training". Its advice to security teams: "Treat weights you didn't build like an unsigned binary, and enforce on the actual tool call before it runs." Deny outbound internet access to AI agents by default and alert on unexpected cloud addresses. Its advice to businesses buying AI: seek assurances on where the model came from, how its training environment was secured, and whether it has been adversarially tested.

On Monday 26 October Anthropic filed its S-1, the registration statement for a US stock market listing, with the Securities and Exchange Commission. It plans to list on the Nasdaq under the ticker ANT. The number of shares and the price have not been set and the filing is not yet effective.

The filing is the first audited look at a frontier AI lab's books. Revenue was $16.3 billion for the six months to 30 June 2026, up from $1.2 billion a year earlier. The June quarter alone brought $11.6 billion of revenue and $1.0 billion of operating income, an 8% operating margin, after an operating loss of $693 million across the half. Cash, cash equivalents and short-term investments stood at $131.7 billion at 30 June. In the filing's words: "We do not expect our revenue will continue to grow at comparable rates, and we cannot assure you that it will continue to grow at all."

Two lines matter to anyone who relies on Claude in their business. On distribution: 48% of revenue in the first half came through cloud marketplaces such as Amazon, Google and Microsoft, so many firms use Claude without ever contracting with Anthropic. On change: "We have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities", and "any accelerated or forced deprecation could result in customer attrition, contractual disputes, or reputational harm".

One detail for the sceptics about AI doing real work: the filing states that "over 90% of the contributions to our codebase in September 2026 were written by Claude", and that code volume per engineer is more than 45 times what it was in 2024.