Anthropic's S-1 became public on October 26, 2026, for a proposed IPO. The registration statement has not yet become effective and the offering's price range has not been set. Anthropic's announcement is here: https://www.anthropic.com/news/s1-sec
An S-1 is a registration statement. Investor.gov describes a registration statement as a filing with the SEC making required disclosures, one that often includes a prospectus describing the offering and the company to investors, and Form S-1 is the form often used to register offerings (https://www.investor.gov/introduction-investing/investing-basics/glossary/registration-statement). In my view, it is not a confession of how the company plans to retain customers. This post is about switching costs for enterprise buyers of AI models. I look at cloud purchasing, developer tooling, integration and vendor approval. Everything about switching-cost implications below is my analysis, not a disclosed plan.
What a filing can and cannot tell a buyer
Where a filing notes that customers can terminate with limited or no notice, that describes contract flexibility. It says nothing by itself about how hard or expensive migration is. Those are separate questions. A customer can leave a contract quickly and still need months to replace the model behind a production workflow.
So the useful question for a buyer is what sits around the model API: cloud purchasing, developer tooling, prompts and tool definitions, and vendor approvals. The sections below take those in turn. The examples below are hypothetical, not measured migration results.
Where switching costs can accumulate
Cloud purchasing
Claude is offered through major clouds, including Amazon (https://platform.claude.com/docs/en/build-with-claude/claude-platform-on-aws), Microsoft Foundry (https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/claude-models-billing) and Google Cloud (https://platform.claude.com/docs/en/build-with-claude/claude-models-on-google-cloud/anthropic-managed).
Existing cloud purchasing may simplify some steps. It does not remove model terms, access setup, security review or purchase-order requirements.
One possibility worth considering: if usage counts toward a commitment, finance may come to favor it.
The distributors also sell competing models
The clouds that distribute Claude also offer other models, such as Nova on Amazon (https://aws.amazon.com/about-aws/whats-new/2025/12/nova-2-foundation-models-amazon-bedrock) and Gemini on Google (https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/google-models). Microsoft's relationship with OpenAI is described in its current announcement: https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/. For a buyer, replacing a model with another one on the same cloud may preserve some integrations, such as authentication and networking. It still requires validation of the new model's behavior.
Claude Code configuration
Anthropic introduced Claude Code as a limited research preview on February 24, 2025 (https://www.anthropic.com/news/claude-3-7-sonnet). Teams using it may have built up established workflows around CLAUDE.md files, hooks, and skills or custom commands. Features and configuration can also differ by provider; see https://code.claude.com/docs/en/feature-availability.
The cost of retraining and revalidating a team is something to estimate for your own organization, for example by listing the configuration in use and the workflows that depend on it. It is not impossible to model.
Prompts, tool descriptions and gateways
A gateway can simplify routing and supported API formats. LiteLLM documents that parameters unsupported by a provider need handling (https://docs.litellm.ai/docs/completion/drop_params), and Portkey documents its SDK support (https://portkey.ai/docs/product/ai-gateway/universal-api). Replacing a model is still not necessarily a config-only production migration.
Hypothetically, a prompt tuned to one model could produce subtly different output formatting or tone on another, so quality drifts without any alert firing. Or a feature you rely on could be unsupported, and the request fails outright. Both kinds of problem are worth looking for before a switch.
MCP is an open protocol under the Agentic AI Foundation, part of the Linux Foundation (https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/). It is implemented through applications, clients and servers (https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture).
Security and compliance review
Adding a model vendor can mean substantial review work. How much depends on the organization and the deployment. Some evidence, such as existing assessments of a cloud provider, may be reusable, while new data flows and new vendors still need assessment. A vendor's safety positioning is not a substitute for evaluating its controls, terms and deployment.
Renewal pricing and model retirement
Renewal pricing is a risk buyers should assess, not something I am predicting. Existing contracts have their own terms. Buyers can read the filing's risk factors for themselves.
On retirement, Anthropic has committed to preserving the weights of publicly released models (https://www.anthropic.com/news/deprecation-commitments). Preservation is not permanent endpoint access for customers, and it is not a public release of weights. For timing, use the current deprecation policy, including its notice scope and any partner-specific calendars: https://platform.claude.com/docs/en/about-claude/model-deprecations. Where available, negotiate suitable notice and successor pricing. That is something to ask for, not an entitlement.
What I would do before a renewal
- Define an adapter or gateway boundary where it makes sense. Native SDK imports can stay where they earn their keep, but the boundary should be explicit.
- Build representative evals with edge cases, output and tool correctness, latency, cost and clear acceptance criteria. Keep evaluating a candidate model separate from the implementation and deployment work of moving to it.
- Inventory spend and configuration: which cloud offer and agreement each workload uses, and where CLAUDE.md files, hooks and custom commands live.
- Keep an alternative provider's approval current, so a switch does not start from zero.
- Review contracts for notice periods, successor pricing and data-handling terms.
Where this leaves me
Customer configuration and integration can contribute to switching costs alongside cloud purchasing and vendor approvals. The practical step is to inventory what you have built around a model and estimate what revalidation would take.
If you have migrated a production workload from one model to another, in either direction, I would like to hear what broke first and how you measured it.