Anthropic prepares for Wall Street

Good day!

There was plenty to follow in AI this week, from a potentially landmark stock market listing to new models, enormous financing plans and some interesting developments in scientific research.

Anthropic filed its S-1 registration statement with the US Securities and Exchange Commission on 26 October, formally beginning the process of taking the company behind Claude public.

The company intends to list on Nasdaq under the ticker ANT, although the number of shares and proposed offering price have yet to be determined.

It's a significant moment for the AI industry. An IPO would give public investors a much clearer view of the economics behind developing frontier AI, including revenue, operating costs, infrastructure commitments and the risks associated with increasingly powerful models.

It will also provide a valuable test of how investors assess a leading AI company whose business depends heavily on expensive computing infrastructure and a rapidly evolving competitive market.

Source: Anthropic's official IPO announcement

OpenAI introduced GPT-6.1 Terra on Monday, targeting professional tasks that require visual understanding, computer use and reliable execution across software applications.

The model is designed to handle activities ranging from interpreting drawings and documents to completing forms and working across applications. Standard API pricing is $1 per million input tokens and $5 per million output tokens, 50% below GPT-5.6 Terra.

OpenAI also reports substantial improvements in computer-use benchmarks, including stronger task completion at lower cost. Those are the company's evaluations, rather than guarantees of performance in every real-world workflow.

Terra is available through the API, ChatGPT Work and Codex, but not ordinary ChatGPT chat. Its commercial appeal is straightforward: capable AI becomes more practical when the cost of running it repeatedly starts to fall.

Source: Introducing GPT-6.1 Terra

Google unveiled Gemini 4 Carbon on 28 October, describing it as its strongest model yet, with particularly advanced capabilities in cybersecurity and complex, long-running tasks.

According to Google's published evaluations, Carbon scored 88.2% on SEC-bench Pro, a benchmark focused on finding security vulnerabilities. The company also reports improvements in coding, reasoning and vulnerability repair.

But Google isn't making the model broadly available immediately. Initial access is being restricted to selected cybersecurity defenders and trusted partners while the company evaluates safeguards.

That is an interesting commercial and security decision. The capabilities that make advanced models useful for finding software weaknesses may also make them valuable to attackers. Google is prioritising controlled access before a wider release.

Source: Introducing Gemini 4 Carbon

Microsoft reported $92.3 billion in quarterly revenue on Wednesday, an increase of 19% compared with the previous year.

Azure revenue grew 48%, while Microsoft Cloud revenue reached $63.9 billion. The company also disclosed $677 billion in remaining commercial performance obligations, reflecting a substantial volume of contracted future business.

Not all of that backlog is AI-related, but the figures demonstrate the scale of Microsoft's enterprise cloud business as customers expand their use of computing infrastructure and AI services.

For investors and businesses following AI economics, the important distinction is between demand, recognised revenue and the cost of delivering the infrastructure. Microsoft's latest results provide another substantial data point in that discussion.

Source: Microsoft's quarterly results

Meta announced new equity financing arrangements this week to support its continued investment in AI infrastructure and product development.

The wider plan, reported by the Financial Times, could raise as much as $95 billion through a combination of near-term share offerings and a further stock-sales programme. The full amount is a potential financing capacity, not money already raised.

The scale is remarkable, particularly alongside Meta's existing spending commitments for data centres, chips and energy.

It also creates a difficult question for shareholders: how much additional capital should a company commit to AI infrastructure today, and how quickly can those investments generate sufficient returns?

Meta is making a substantial financial commitment to its AI ambitions. Investors will be watching the financing terms and the evidence that its AI products can justify the expenditure.

Source: Financial Times — Meta's financing plans

Canada's government announced plans on Thursday to make an equity investment of up to C$1.6 billion in Toronto-based AI company Cohere.

The investment will be made in stages alongside private investors, including Nvidia, Bell and Schwarz Group.

In return, Cohere has committed to maintaining its Canadian headquarters, investing at least C$1.7 billion in additional domestic computing infrastructure by 2030 and creating at least 1,000 technical jobs.

The decision reflects growing interest in sovereign AI: technology that governments and organisations can operate with greater control over their own data, infrastructure and security requirements.

Canada is taking an unusually direct approach by proposing to become an equity investor in a domestic AI developer. That could make Cohere an important example of state-backed competition in the global AI market.

Source: Government of Canada's announcement

French AI company Mistral released the weights of Mistral Large 4 on Thursday, following the model's earlier public preview.

Nicknamed Le Chonk, the model contains approximately one trillion parameters and is designed for multimodal reasoning, coding and cybersecurity tasks.

Unlike models available exclusively through hosted services, Mistral Large 4 can be downloaded and deployed on an organisation's own infrastructure. The licence permits a wide range of commercial uses, although model-as-a-service businesses require a separate commercial agreement.

Mistral reports significant coding improvements since the preview, including better results with fewer generated tokens.

Its release gives businesses and governments another advanced model they can customise and operate under their own infrastructure and data-control requirements, without relying entirely on a provider-hosted API.

Source: Mistral Large 4 open-weight release

Anthropic released a detailed research update on Tuesday describing how Claude is being used in partnerships with six organisations developing medicines.

The examples go well beyond asking AI to suggest new molecules.

Working with Divergence Labs, Claude helped generate roughly 35,000 antibody designs for a potential food-allergy treatment. Separately, researchers working with Generate Biomedicines used Claude Code to help redesign part of a CAR T-cell therapy, producing a shorter genetic design with promising laboratory results.

Other projects explored antibody optimisation, manufacturing documentation and clinical trial planning.

These remain early-stage findings, largely before human clinical testing. Nevertheless, the research offers a useful picture of how AI could support several different stages of drug development, rather than concentrating solely on discovering potential treatments.

Source: Anthropic's drug development research

Google DeepMind introduced Genie 4 on Tuesday, a new world model capable of generating interactive simulated environments with accompanying audio.

Users can influence what happens inside a simulation through natural-language instructions, changing elements such as the weather, camera movement or characters while the world continues to evolve.

The potential applications are interesting. Real-time simulations could eventually support game development, filmmaking, robotics research and training environments where agents learn how to interact with changing surroundings.

Google has begun rolling Genie 4 out to a limited group of trusted testers, with wider access planned.

For now, it remains an early research product, but it demonstrates a different application of generative AI: building environments that respond to actions, rather than simply producing finished images or videos.

Source: Google DeepMind introduces Genie 4

Cadence Design Systems introduced Maestro on Monday, a portfolio of AI models developed specifically for semiconductor and electronic system design.

The models are designed to work alongside existing engineering tools, helping analyse design choices and optimise complex chip-development processes.

Cadence reports initial results showing up to 30% better design quality and five times greater token efficiency compared with general-purpose foundation models. These figures come from its early evaluations, with wider commercial performance still to be established.

An important feature is that customers can run the systems in their own data centres, allowing highly sensitive chip designs to remain within controlled environments.

It's a useful reminder that some commercially valuable AI systems may be highly specialised rather than general-purpose chatbots.

Source: Cadence announces Maestro

O T H E R_A I N E W S

Alphabet reported a 103% increase in Google Cloud revenue, taking the business to $30.7 billion in the third quarter. Growth was supported by enterprise AI infrastructure, AI services and demand for Google's TPU systems.

Amazon reported $46 billion in quarterly AWS revenue, up 39% year on year. The company also reported a $42.4 billion trailing-12-month free cash flow outflow, reflecting heavy investment in AI infrastructure.

OpenAI detailed an expanded system for monitoring unauthorised AI-agent behaviour. In a retrospective simulation of an earlier security incident, the system generated an alert more than 30 hours before the point at which the actual breach occurred.

Bedrock Robotics raised $600 million at a $4.1 billion valuation, supporting the expansion of autonomous excavators and other construction machinery already operating on commercial worksites.

GitHub made Copilot Fusion generally available, allowing its coding assistant to select and coordinate different AI models automatically according to task requirements, performance and cost.

Google introduced document creation and editing directly inside Gemini conversations. Users can now generate, preview and revise documents, spreadsheets, presentations and PDFs without leaving the chat interface, with some format restrictions on mobile.

Anthropic expanded its US government agreement to offer a 40% Claude discount, extending discounted access across eligible federal, state, local, tribal and territorial government organisations.

Perplexity announced FedRAMP Moderate certification, enabling more US government teams to use its live web research and cited-answer capabilities within a certified environment.

OpenAI updated its public Model Spec around five core values, clarifying how models should interpret human intent, minimise harm, communicate honestly, respect human agency and exercise care.

That’s all for this week!

News Digest AI team