56% of Tokens Cost Only 14% of the Money: Save Claude for the Hardest 10%

August, Vercel's AI Gateway—an AI call aggregation layer—handled tokens in which open-source models accounted for 56%, yet consumed only 14 cents of every $1 bill. The remaining tokens accounted for the other 86%.

The source of the gap is simple: routing strategy. Everyday coding, refactoring, documentation, and other ordinary tasks go to cheap models; only the truly hard 10% uses Claude.

X user @mikenevermiss turned this approach into a 5-minute setup.

Give Claude Code a cheaper brain in 5 minutes

First install Claude Code:

npm install -g @anthropic-ai/claude-code

Put a deepseek.json in your home directory and fill in DeepSeek's API address, key, and the current V4 Pro model ID. The model ID is similar to deepseek-v4-pro; refer to DeepSeek's official documentation for the exact value.

Start with this config and DeepSeek takes over:

claude --settings ~/deepseek.json

Testing, refactoring, and writing docs go through DeepSeek. In another terminal, run claude as usual; complex tasks go to Claude. Same Claude Code interface, different brain behind it.

A more direct way is to switch with environment variables:

export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="$DEEPSEEK_API_KEY"
claude

DeepSeek automatically maps the Opus model name to deepseek-v4-pro, and Sonnet and Haiku to deepseek-flash. To switch back to Claude, close the terminal or unset the variables.

Other models are also compatible with the Anthropic protocol. Kimi uses https://api.moonshot.ai/anthropic, and GLM uses npx @z_ai/coding-helper.

Claude is still there; it is just responsible only for the hardest 10% now.

Why nobody is making a big fuss about this

The valuations of big US labs are built on the belief that 'only the most expensive models are good enough.' Anthropic's investors reportedly expect an IPO valuation of at least $2 trillion—a number that requires customers to believe top models are irreplaceable.

The labs are patching the walls too. On October 12, Anthropic announced that new defenses reduced the data 'distillers' extract from Claude by about 90% to 99% (Anthropic's distillation defense report). Distillation is using a strong model's outputs to train a cheap model.

Researchers still disagree about how much of the progress of Chinese models comes from distillation. OpenAI policy lead Dean Ball said distillation may play a bigger role than he thought. But regardless of the reason, cheap models really exist, are publicly downloadable, and anyone can use them.

How much cheaper? Look at the numbers

Artificial Analysis's Coding Agent Index tests AI coding tools on the same set of real tasks. In version 2.0, Claude Code running Opus 5.5 costs $11.42 per task. Codex running DeepSeek V4.1 Flash costs $0.39 per task.

96.6% cheaper. For the price of one Claude task, you can run about 29 DeepSeek tasks.

Six open-source models worth knowing

Prices in the table are per million tokens. Tokens are roughly word fragments, and models bill by token.

Most of these numbers come from third-party evaluations, not vendor boasts. The WebDev leaderboard is produced by real human blind reviews; DeepSeek V4.1 Flash is the top open-source model with a score of 1,670, ahead of Claude Opus 5 (High). Kimi K3.1 Pro ranks #4 on the Intelligence Index; ahead of it are Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full. Ming-3.0 supports Cantonese and other Chinese dialects; MiniCPM-V 4.7 keeps images from leaving the device.

Five integration paths

Ollama, easiest: after installation, run MiniMax with one command:

curl -fsSL https://ollama.com/install.sh | sh
ollama run minimax-m3.1-flash:cloud

The :cloud tag runs the model on Ollama's servers—the model is too large for a laptop.

DeepSeek API, for scripts and apps: register, top up, and create a key. The API is compatible with OpenAI's format; change base_url to https://api.deepseek.com and you can use your existing code.

Claude Code backend swap: the two environment-variable lines above.

OpenCode, free proxy: an open-source coding agent that offers DeepSeek V4.1 Flash free for a limited time. In OpenCode, use /models to select DeepSeek V4.1 Flash Free. The free promotion may end at any time; if requests start failing, choose another model from the list.

Download weights yourself: GLM-5.3-Flash, Ming-3.0, and MiniCPM-V 4.7 can all be downloaded from their official pages. Data does not leave the machine, and there is no usage-based bill. GLM needs heavy hardware such as multiple GPUs; Ming and MiniCPM have a lower barrier.

Small models beating large models: Coinbase

Coinbase fine-tuned Alibaba's Qwen3.5-9B into a fraud-detection model. Fine-tuning cost less than $100. On its own fraud benchmark, this 9B small model beat Claude Opus 4.5, with 55% lower latency.

The conclusion is narrow but useful: Qwen does not comprehensively surpass Claude, but a small model heavily trained for repetitive tasks can beat a general large model in that specific role. If your business has recurring similar tasks, it is worth running a round of tests.

Safety boundaries for using cheap models

• Closed-source models from big labs still top the overall intelligence index. Large refactors, multi-step agent tasks, and high-stakes decisions should still go to Claude.
• Only download weights from official pages. Security researchers found that 'abliterated' (safety training removed) open-source models may hide backdoors. Secondhand uploads by strangers are a high-risk area.
• Know where data goes. With any hosted API, prompts are processed on the other party's servers. Read the privacy policy before sending customer data or personal information. For sensitive content, run open-source weights locally.
• Check the license before using a model in a product. DeepSeek V4.1 Flash and Ming-3.0 are MIT, so commercial use is unrestricted; MiniMax's weights have conditions for commercial use.
• Do not put API keys in code. Use environment variables, and do not paste them into public repos or screenshots.

How to start

Route the everyday 90% to cheap models, and leave only the hardest 10% for Claude. Start with one path and try it on a week of normal work. Record which tasks the cheap models cannot handle; after a week, that list is your 10%.