Claude Code users can now slash their AI token costs by 10 to 50 times by routing requests to Chinese language models through Anthropic-compatible endpoints, though this strategy requires careful task selection to avoid quality degradation. The trick works because Claude Code reads its backend address from two environment variables, allowing developers to redirect the entire workflow to providers like DeepSeek, Zhipu GLM, and Moonshot Kimi without changing a single line of code.

How Does Redirecting Claude Code to Cheaper Models Actually Work?

Claude Code isn't permanently locked into Anthropic's servers. Instead, it reads two configuration variables: ANTHROPIC_BASE_URL (the server address) and ANTHROPIC_AUTH_TOKEN (authentication credentials). If a provider exposes an Anthropic-compatible endpoint, usually marked with an "/anthropic" suffix, developers can point Claude Code to that alternative backend with a single variable change. The entire workflow, including agents, parallel tool-use, and multi-step reasoning, runs without modification.

This flexibility emerged because major Chinese AI providers have already built Anthropic-compatible endpoints. DeepSeek, GLM, Kimi, MiniMax, Qwen, and Xiaomi MiMo all offer official "/anthropic" endpoints that plug in within thirty seconds. However, there's a critical catch: these endpoints are "trimmed" versions that strip out several features native to Claude Code.

What Features Get Left Behind When You Switch to Cheaper Models?

The cost savings come with a tradeoff. Chinese provider endpoints do not support Model Context Protocol (MCP) servers, Claude Code's native web search, image or PDF processing, or features behind Anthropic's beta API. This isn't a limitation of any single provider but a shared architectural property across all of them. Understanding this boundary is essential because crossing it blindly turns savings into rework.

The practical implication is straightforward: developers must route tasks strategically. Volume-heavy work like bulk research, iterative landing page drafts, and large-scale content generation can move to cheap models. Final-stage work, anything requiring web search, image handling, or sensitive data processing must stay on Claude. This two-tier approach, sometimes called the "NeuroDrift pattern," lets cheap models handle the housekeeping while Claude handles the irreducible minimum.

Which Chinese Models Cost the Least, and What Are They Good For?

Pricing varies significantly across providers as of June 2026. DeepSeek v4-flash costs $0.14 per million input tokens and $0.28 per million output tokens, making it the cheapest usable default. GLM-4.7-FlashX undercuts even that at $0.07 input and $0.40 output, though it excels at multilingual work. For agentic coding tasks that require tool-use loops, GLM's Coding Plan offers a fixed monthly subscription ranging from $12 to $72 that removes meter anxiety during iterative development. MiniMax M2.7 sits in the middle at $0.30 input and $1.20 output, balancing cost and capability.

The key insight is that no single model works for everything. Instead, developers should route by task type. Large research projects with hundreds of sources benefit from DeepSeek's 1-million-token context window and cache-hit pricing of $0.0028, which makes bulk reading nearly free. Landing page iteration works well on GLM's fixed subscription, avoiding the anxiety of meter-based pricing. Bulk English content favors DeepSeek, while multilingual or Ukrainian output should use GLM for better quality, with Claude handling the final polish.

Steps to Set Up Budget-Conscious Claude Code Workflows

• Budget Profile for Research: Set ANTHROPIC_BASE_URL to "https://api.deepseek.com/anthropic" and ANTHROPIC_AUTH_TOKEN to your DeepSeek API key. Configure ANTHROPIC_DEFAULT_OPUS_MODEL and ANTHROPIC_DEFAULT_SONNET_MODEL to "deepseek-v4-pro" and "deepseek-v4-flash" respectively. This profile handles large research runs, bulk text generation, and scaffolding work at minimal cost.
• Coding and Landing Pages Profile: Switch to GLM's Coding Plan by setting ANTHROPIC_BASE_URL to "https://api.z.ai/api/coding/paas/v4" and using your GLM API key. Set models to "glm-5.2" for Opus and Sonnet tasks. The fixed subscription removes per-token anxiety during iterative development cycles.
• Automatic Routing Within a Single Session: Install the claude-code-router package (npm i -g @musistudio/claude-code-router) to automatically send housekeeping calls and reading tasks to the cheapest provider while routing main reasoning to a pricier, higher-quality model. This hybrid approach maximizes savings without sacrificing output quality.

Where You Absolutely Cannot Cut Corners with Cheap Models

Certain tasks must stay on Claude, regardless of cost pressure. Any workflow requiring MCP servers (Notion, GitHub, Postgres, Figma, Gmail) cannot route through Chinese endpoints because those features don't exist on the trimmed endpoints. Native web search, image processing, screenshot analysis, and PDF handling all require Claude. Final-stage quality assurance for Ukrainian language output should use Claude as the reference, since DeepSeek's Ukrainian quality is noticeably weaker due to its English and Mandarin training corpus. Most critically, any work involving customer private data should never touch Chinese servers due to GDPR and compliance risks.

The honest assessment is that this cost-reduction strategy works best for teams with predictable, high-volume workloads. Research teams running hundreds of source-gathering iterations, content agencies producing bulk material, and development shops iterating on landing pages see the biggest savings. Teams doing sensitive work, heavy MCP integration, or requiring web search throughout the workflow see minimal benefit and should stick with native Claude.

What Does This Mean for the Broader AI Development Landscape?

The emergence of Anthropic-compatible endpoints from Chinese providers signals a shift in how developers approach AI tooling costs. Rather than being locked into a single vendor's pricing, developers can now treat Claude Code as a workflow orchestrator that delegates to the most cost-effective backend for each task type. This flexibility doesn't diminish Claude's value; it repositions it as the premium option for tasks where quality, safety, and feature completeness matter most. For cost-conscious teams, the strategy transforms Claude Code from an all-or-nothing expense into a hybrid infrastructure decision, much like how organizations choose between cloud providers based on workload requirements.