Infrastructure, Interoperability, and the IPO Clock: Western AI's Defining Week

Three events from the past 48 hours, taken together, reveal something about where the Western AI industry is heading: OpenAI shipped a purpose-built inference endpoint for structured decisions; Anthropic handed its Agent Skills packaging standard to a neutral foundation; and Meta reported quarterly capital expenditures that, for the first time, exceeded its operating cash flow. None of these is a headline model release. All three are, in their own way, more consequential than one.

OpenAI's Decisions API: A Dedicated Endpoint for Bounded Choices

On October 22, 2026, OpenAI moved the Decisions API to general availability↗, ending a public beta that had run since early October. The API is powered exclusively by GPT-6 Luna, a model that does not appear in OpenAI's standard chat or completion endpoints and is not available for general text generation. Its sole function is to evaluate a shared input context — text, images, or native JSON — against a set of developer-defined questions and return structured, typed answers.

The three answer types are precise and deliberately narrow:

• Predicate: Returns a probability between 0 and 1 for a yes/no condition — for example, "does this support ticket describe a billing issue?"
• Choice: Selects from a predefined list of options and returns the chosen label alongside per-option probabilities and a confidence score.
• Score: Evaluates input against an ordered rubric of levels and returns a probability-weighted average — useful for severity classification or quality grading.

According to OpenAI's technical documentation↗, a single request can contain up to 400 independent questions, all evaluated against the same shared context. GPT-6 Luna's context window sits at 1,050,000 tokens, which means the API can process very long documents — legal contracts, medical records, lengthy conversation histories — without chunking. Latency is vendor-stated at approximately 150 milliseconds, roughly 10× faster than the Responses API for equivalent bounded tasks, though that figure will vary with input length and network conditions.

Pricing and the Competitive Angle

The economics are structured to encourage high-volume, automated use. OpenAI charges $0.10 per million input tokens↗, with no output-token fee — a sensible model given that the outputs are typed scalars rather than prose. Prompt caching drops the cost to $0.01 per million cached tokens, which matters for workloads that repeatedly evaluate the same system prompt or rubric against changing inputs.

The competitive framing is explicit: OpenAI's own materials position the Decisions API against TypeSafe's Jev model, a specialist decision endpoint that has been available at lower list prices. At roughly 2.4× Jev's price, OpenAI is betting that enterprise compliance requirements — HIPAA eligibility, Zero Data Retention, regional processing in the US and EEA — plus the ability to handle images and 400-question batches will justify the premium. That is a reasonable bet for regulated industries; less obvious for cost-sensitive, high-throughput consumer applications.

"The Decisions API is not a general-purpose model. It is infrastructure for the agentic layer — the routing and classification substrate that sits beneath every agent that needs to make fast, bounded choices without generating explanatory text."

What the Decisions API signals, more broadly, is that OpenAI is disaggregating its model stack. GPT-6 Luna is not GPT-6; it is a purpose-built derivative optimised for a specific inference pattern. Expect more of this: as the agentic use case matures, the monolithic "one model for everything" approach becomes economically and architecturally inefficient.

Anthropic Hands Agent Skills to the Linux Foundation

On the same day, Anthropic announced the donation of its Agent Skills standard to the Agentic AI Foundation (AAIF)↗, a body operating under the Linux Foundation. The transfer includes the project name, logo, GitHub organisation, and the `agentskills.io` domain.

Agent Skills originated inside Claude Code in December 2025 as a way to package specialised knowledge and workflows for AI agents. A skill is a folder: at minimum, a `SKILL.md` file containing metadata, a description, and instructions; optionally, scripts, templates, and reference materials. The architecture uses progressive disclosure to manage context window costs — at startup, an agent reads only the skill's name and description (roughly 100 tokens); the full `SKILL.md` body loads only when a task triggers the skill; scripts and reference materials are accessed only when explicitly required.

Scale and Adoption

The numbers are striking. According to the Linux Foundation's announcement↗, the standard is now integrated into more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code. Approximately 1 million skills have been published, with nearly 280 million installations recorded. For a standard that is less than a year old, that is a meaningful installed base.

The AAIF's portfolio now covers ten projects — up from three at launch — spanning agent capabilities, model interoperability via the Model Context Protocol (MCP), reinforcement learning tooling, and secure execution environments. Lead maintainers going forward are David Soria Parra and Mahesh Murag, with engineers from Anthropic, Google, and OpenAI contributing in their individual capacities rather than as corporate representatives.

"By placing Agent Skills under neutral stewardship, Anthropic is making a calculated bet that a larger, interoperable ecosystem is worth more than proprietary lock-in — the same logic that drove the MCP donation earlier this year."

The security dimension deserves attention. Because skill packages can include executable scripts, they are not inherently safe. Industry reports from Snyk and JFrog have already identified malicious skills in the wild. The AAIF's roadmap includes versioned releases, a formal specification-change process, and standardised security guidance — necessary infrastructure for a standard that is now embedded in production systems at scale.

Anthropic's Parallel Cloud Move

Separately, on October 22, Anthropic launched "Claude Models on Google Cloud (Anthropic managed)"↗ — a new deployment option that gives Google Cloud customers day-zero access to Claude features under Anthropic's commercial terms, with Anthropic acting as data processor rather than Google. The service supports Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5, with Opus 5.5 Fast Mode available at $8 per million input tokens and $40 per million output tokens, delivering up to 2.5× faster output generation for interactive workloads.

The distinction from the existing Google-managed Claude offering is deliberate:

• Anthropic managed: Day-zero feature access, Anthropic as data processor, billing via Google Cloud Marketplace.
• Google managed: Delayed feature access, Google as data processor, native integration with Model Armor, Agent Engine, and VPC Service Controls.

For enterprises that need the latest Claude capabilities immediately — particularly those building on the agentic tooling Anthropic ships frequently — the Anthropic-managed path removes the integration lag that has historically made Vertex AI a second-class citizen for Claude deployments.

Meta's Earnings: When AI Capex Exceeds Operating Cash Flow

The third data point from this week is financial rather than technical, but it may be the most structurally significant. On October 21, Meta reported Q3 2026 results↗ that beat revenue expectations — $64.6 billion, up 26% year-over-year — while simultaneously alarming investors with a capital expenditure figure that has no precedent in the company's history.

Meta spent $34.9 billion on capex in Q3 alone, up from $31.1 billion in Q2 and $19.4 billion in Q3 2025. For the first time, quarterly capex exceeded operating cash flow of $32.8 billion↗, producing negative free cash flow of approximately $2.1 billion — compared to positive free cash flow of $10.6 billion in the same quarter of 2025. Full-year 2026 capex guidance was narrowed to $135–$145 billion.

The stock fell 6.98% to $648.85 on October 22, with investors citing three concerns:

• The sheer scale of infrastructure spending, with no clear near-term revenue offset.
• The announcement of a new class of non-voting Class C shares, raising dilution concerns.
• Operating margins compressed to 20% (from 40%), partly due to a $9.89 billion legal settlement related to youth safety.

What the Spending Is For

Management's answer is the Muse AI agent suite, which recorded more than 10 million setups in the US and Canada within six weeks of its launch. The Muse ecosystem — spanning the Muse personal agent, Muse Wonder model, Muse Code for engineering workflows, and the forthcoming Muse Charm hardware companion — is Meta's attempt to build a vertically integrated AI platform that competes with OpenAI's ChatGPT ecosystem and Anthropic's Claude products.

The infrastructure required to run that platform at Meta's scale — billions of daily active users across WhatsApp, Instagram, Facebook, and the Meta app — is genuinely enormous. The JEM ad-ranking model alone, which drives the advertising revenue that funds all of this, requires continuous retraining at a scale that most AI labs cannot approach.

Whether the market will reward this level of spending depends on whether Muse achieves the kind of retention and monetisation that justifies the infrastructure. The early adoption numbers are encouraging; the free cash flow trajectory is not.

The Week's Underlying Logic

Taken together, these three developments point toward a maturing industry that is increasingly focused on infrastructure, interoperability, and financial sustainability rather than raw capability benchmarks.

• OpenAI is disaggregating its model stack into purpose-built endpoints, trading generality for efficiency at the agentic layer.
• Anthropic is building open standards infrastructure — Agent Skills, MCP — that makes the broader ecosystem more capable while reducing the risk that any single vendor's proprietary format becomes a lock-in mechanism.
• Meta is making the largest infrastructure bet in the industry's history, accepting negative free cash flow in the near term on the thesis that AI-native products will eventually justify the spend.

None of these moves is about who has the best benchmark score this week. All of them are about who controls the infrastructure, the standards, and the distribution channels that will determine which models and products reach users at scale over the next several years.

The Anthropic IPO clock is also ticking in the background. The company confidentially filed its S-1 with the SEC in June 2026↗ and is reportedly targeting a public debut after the November US midterm elections, with analysts suggesting a valuation target exceeding $2 trillion against an annualised revenue run rate of approximately $65 billion. The Agent Skills donation and the Google Cloud launch are, among other things, signals to prospective public-market investors: Anthropic is building ecosystem infrastructure, not just models.

The week's quieter story — that no major new frontier model shipped from any of the leading Western labs — is itself informative. The race to release is giving way, at least temporarily, to the harder work of making what already exists reliable, interoperable, and economically sustainable.

---

*Lukas Hoffmann is Neuron's Europe & Frontier Correspondent, based in Berlin.*