Claude Code, Anthropic's AI coding assistant, is no longer locked to Claude models. Developers can now route the tool through any model that speaks the Anthropic Messages API, from open-weight alternatives to competing frontier models, using just three environment variables. This shift transforms Claude Code from a branded product into a flexible harness where the underlying model becomes a swappable choice.

Why Would Developers Want to Swap Models Inside Claude Code?

The appeal comes down to economics and specialization. Frontier models like Claude Opus excel at open-ended reasoning and research, but much agentic work is repetitive: reading files, applying edits, running tests. Paying premium rates for mechanical tasks wastes money when a cheaper or open-weight model handles the same work for a fraction of the cost. Some developers plan complex tasks with one model and delegate execution to a leaner alternative, creating a cost-efficient workflow that keeps the same tools and interface intact.

Beyond cost, developers cite several practical reasons for model flexibility:

• Availability: Switching models keeps work flowing when one provider experiences throttling or downtime, ensuring uninterrupted productivity.
• Privacy: Sensitive code that cannot leave internal systems can run on local models or private endpoints instead of cloud APIs.
• Avoiding lock-in: The next strong model is one command away rather than requiring a full tool migration, reducing switching costs.
• Experimentation: A one-line configuration change is the cheapest way to test whether a new model suits your specific coding workflow.

As one developer explained, the deeper principle is that "harness matters more than the model." Your skills, tools, context, and agent setup are durable assets, while the model underneath is increasingly a commodity you can swap without losing the expertise built into your workflow.

How to Run Any Model Inside Claude Code

The technical setup is straightforward. Claude Code communicates with models through Anthropic's Messages API, and developers control where requests go, how authentication works, and which model runs entirely through three environment variables. No forking, no binary changes, and no deep configuration required.

• ANTHROPIC_BASE_URL: Points Claude Code at a different endpoint, whether that is a gateway service like OpenRouter, a local proxy, or a custom shim that translates between protocols.
• ANTHROPIC_AUTH_TOKEN: Sends authentication as an Authorization Bearer header, the standard format most gateways expect for API requests.
• ANTHROPIC_MODEL: Names the specific model to run, using provider/model-slug format such as "z-ai/glm-5.2" or "openai/gpt-5.1."

The fastest entry point is OpenRouter, a gateway service that natively speaks the Anthropic Messages protocol. Because OpenRouter exposes an Anthropic-compatible endpoint at https://openrouter.ai/api/v1/messages, Claude Code appends the /v1/messages part itself. This means developers set the base URL to https://openrouter.ai/api, not the full path. A common gotcha: including /v1 yourself results in requesting /api/v1/v1/messages, which fails in ways that are annoying to debug.

Once those three variables are exported, launching Claude Code runs entirely on the chosen model with all tool calling, file edits, and skills intact. OpenRouter's catalog includes models from dozens of providers: GLM 5.2 from Zhipu, GPT-5.1 from OpenAI, DeepSeek Chat v3.1, Qwen3 Coder, and many others.

For developers who want to avoid manual setup, a simple shell function can automate the process. The function takes a model slug as an argument, sets the three variables only for that invocation, and launches Claude Code. This keeps environment state from leaking into the shell and makes switching models as simple as typing a single command followed by a model name.

What Does This Mean for the Broader AI Coding Landscape?

This flexibility arrives as the AI coding market fragments. Different models have distinct temperaments: some excel at open-ended reasoning, others at rapid execution. Rather than forcing developers to choose a single tool and live with its defaults, Claude Code now lets them match the model to the task. A developer might use Opus for complex architecture decisions but switch to a faster, cheaper model for routine refactoring.

The shift also reflects a broader principle articulated by industry leaders: the real competitive advantage is not picking the best model, but building a learning loop on top of models where human capital and token capital compound. Your skills, tools, and context are the durable asset. The model is increasingly a swappable commodity.

Meanwhile, enterprise data access is becoming another critical layer. CData Software launched Connect AI Developer Edition, a free platform that gives developers governed access to enterprise systems like Salesforce, Snowflake, NetSuite, and Microsoft 365 through SQL, Python, the command line, and MCP (Model Context Protocol). The platform works out of the box with Claude Code, Codex, Cursor, and LangChain, among others, letting developers pull live enterprise data without asking IT for permission at every step.

CData also released the CData Connect AI Python SDK, an open-source tool providing database-compatible access to governed enterprise data, and CData CLI, a command-line interface designed for developers using tools like Claude Code to scaffold connectivity without digging through documentation.

"Developers have been forced to choose between moving fast and meeting the governance their company requires. That tradeoff doesn't hold up anymore," said Raviv Levi, Chief Product and Technology Officer at CData.

Raviv Levi, Chief Product and Technology Officer at CData

The combination of model flexibility and data access represents a shift in how developers build AI applications. Rather than being locked into a single vendor's stack, developers can now compose tools from multiple providers, swap models based on task and cost, and access enterprise data through standardized interfaces. Claude Code's evolution from a branded product to a universal harness reflects this broader trend toward modularity and choice in the AI development ecosystem.