February 2026: AI Coding on a Budget

AI is moving so fast it's nearly impossible to keep up with it or have enough time in the day! I am having a blast though.

🚀 Current Workflow Lately

Claude Code and Github Copilot most of the time (Copilot is not as good as Claude Code but they do try to catch up fast and they have the most bang for your buck when it comes to access to the best models: Claude Opus 4.5, GPT 5.2, etc).

Claude Code is just the best!!

Some people seem to think that it is only the model intelligence that matters, and the agent harness / coding tool around the models aren't a big deal. This is sooooo wrong.

Why Claude Code + Cheap Models > Expensive Models + Bad Tools

I can get such a better, smoother experience using Claude Code with Deepseek v3.2 and/or Minimax M2.1 and/or GLM 4.7, compared to using Claude Opus 4.5 in some tools like Opencode.

I am only paying 28-40 cents per million tokens if I use the official Deepseek Anthropic-compatible API endpoint with Claude Code.

I use Claude Code in Windows often in WSL (Linux sort of "inside" Windows, a lot of things just work better in WSL) and have my ~/.bashrc file set up with custom commands to run Claude Code with either Deepseek or Minimax. GLM 4.7 is the cheapest option here... $3/month still I believe!

## Claude Code Custom Commands
CLAUDE_BIN="$HOME/.local/bin/claude"
CLAUDE_SETTINGS="$HOME/.claude/settings.json"

### Helper function to swap settings temporarily
### There are better/other ways to do it, but I know this works
### I haven't tested other versions but you can

_claude_with_config() {
    local json_content="$1"
    shift

    # Ensure .claude directory exists
    mkdir -p "$HOME/.claude"

    # Backup original settings
    [ -f "$CLAUDE_SETTINGS" ] && cp "$CLAUDE_SETTINGS" "$CLAUDE_SETTINGS.backup"

    # Write JSON directly - no variables to expand
    echo "$json_content" > "$CLAUDE_SETTINGS"

    # Run claude
    "$CLAUDE_BIN" "$@"
    local exit_code=$?

    # Restore original settings
    [ -f "$CLAUDE_SETTINGS.backup" ] && mv "$CLAUDE_SETTINGS.backup" "$CLAUDE_SETTINGS"

    return $exit_code
}

## DeepSeek Configuration
claude-ds() {
    local config='{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "sk-examplekey",
    "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "DeepSeek-V3.2",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "DeepSeek-V3.2",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "DeepSeek-V3.2",
    "ANTHROPIC_MODEL": "DeepSeek-V3.2",
    "ANTHROPIC_REASONING_MODEL": "DeepSeek-V3.2-Speciale",
    "API_TIMEOUT_MS": 600000
  },
  "includeCoAuthoredBy": false
}'
    _claude_with_config "$config" "$@"
}

## This version of the Deepseek command adds the skip permissions flag
claude-ds2() {
    local config='{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "sk-examplekey",
    "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "DeepSeek-V3.2",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "DeepSeek-V3.2",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "DeepSeek-V3.2",
    "ANTHROPIC_MODEL": "DeepSeek-V3.2",
    "ANTHROPIC_REASONING_MODEL": "DeepSeek-V3.2-Speciale",
    "API_TIMEOUT_MS": 600000
  },
  "includeCoAuthoredBy": false
}'
    _claude_with_config "$config" --dangerously-skip-permissions "$@"
}

## MiniMax Configuration
claude-mm() {
    local config='{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "eyJexamplekeyFJP5CLrCE1f-sxKKg",
    "ANTHROPIC_BASE_URL": "https://api.minimax.io/anthropic",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "MiniMax-M2.1",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "MiniMax-M2.1",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "MiniMax-M2.1",
    "ANTHROPIC_MODEL": "MiniMax-M2.1",
    "API_TIMEOUT_MS": 3000000,
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  },
  "includeCoAuthoredBy": false
}'
    _claude_with_config "$config" "$@"
}

claude-mm2() {
    local config='{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "eyJhbGcfakeexamplekeyNpfWyandBFJP5CLr9999CE1f-sxKKg",
    "ANTHROPIC_BASE_URL": "https://api.minimax.io/anthropic",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "MiniMax-M2.1",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "MiniMax-M2.1",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "MiniMax-M2.1",
    "ANTHROPIC_MODEL": "MiniMax-M2.1",
    "API_TIMEOUT_MS": 3000000,
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  },
  "includeCoAuthoredBy": false
}'
    _claude_with_config "$config" --dangerously-skip-permissions "$@"
}

## Claude Menu Command
claude-menu() {
    echo "Available Claude Custom Commands:"
    echo "  claude-ds    - Run Claude with DeepSeek"
    echo "  claude-ds2   - Run Claude with DeepSeek (skip permissions)"
    echo "  claude-mm    - Run Claude with MiniMax"
    echo "  claude-mm2   - Run Claude with MiniMax (skip permissions)"
}

🎯 Why Claude Code Stays Ahead

All of those coding tools are lagging behind Claude Code, and it seems like the main developer on that is just way smarter (and faster) so it's unlikely to change anytime soon. He's the one coming up with all the novel methods to save context or split context, he knows how to delegate tasks with the right context to subagents, etc.

For example, the Skills feature is genius. Most people thought it was a gimmick - not me! It's immediately obvious how smart it is with the ability to not shove the entire amount of text at the model. It can choose to load more progressively.

I can see a lot of other things that could possibly use a similar feature/addition like context pruning... with a dedicated model that does that in the background. Might be a good thing.

💰 Use Smartest Models for Planning, Cheap Models for Doing

You don't have to use the pricey Claude 4.5 Opus and Sonnet the entire time - I see that the internet is still full of angry people when they can't do much with a $20 Claude subscription. Do they upgrade to MAX for $200? People tend to use Opus for everything or think they need at least Sonnet.

The Strategy:

1. Planning Stage: Use Claude Opus 4.5, GPT 5.2, Deepseek v3.2, or Gemini 3 Pro
2. Execution Stage: Use GLM 4.7, Minimax M2.1, or Deepseek v3.2
3. Tool Usage: Cheap models work just fine for file operations, searches, etc.

📝 Spend Longest Time in the Planning Stages

Figure out details so it won't guess

If you pay for Claude, use it for planning stages, spend much more time here than feels comfortable. Or just use Deepseek v3.2 here (it is so good, I am not sure why I ignored it for a while!).

Look up SPEC.md methods or different PLAN approaches. I like this plugin called Superpowers. Github has some spec command you can download that is probably good although it seemed too time consuming. I am still trying to find the perfect plan tool or workflow to plan really good to where it can one-shot a complex project.

There are easy ways around this, even better than what I had discovered before... read on!

⚡ GLM 4.7

It's a good "do-er" model, not as smart as Claude or Deepseek v3.2, but it is fast and cheap. Only $3/month!

🏆 Claude Code Is Always Several Steps Ahead of the Rest

I am currently using Claude Code with Deepseek v3.2 which only costs:

I also use it a TON with Minimax M2.

My Current Workflow:

1. Plan with aicodeprep-gui using multiple models (and the new Flow Studio)
2. Get everything fine-tuned
3. Get a good plan and possibly a speckit going (if it's a big enough software project)
4. Load Claude Code up
5. Give it the markdown plan and just let it run
6. Give it some MCP servers to help automate testing

Works so well.

🤖 Gemini CLI

It is getting better! I have not used it that much but I am going to see what it can do vs Claude Code. It is still free!

I use Gemini 3 Flash and Pro in Copilot for debugging... its great, better than Claude usually for that.

You can get the 3-month $300 Google Cloud credit, which allows you to use Nano Banana Pro (REALLY amazing) in the CLI/API.

🌌 Google Anti Gravity IDE

It's good! It's free to use (rate limited but usable) and you can use Opus 4.5 + Gemini 3 Pro free.

It is a VS Code Clone.

Link: https://antigravity.google

🌍 Updates to aicodeprep-gui: Now Multilingual!

I made it multilingual as some people asked about it - originally I made it for myself which is why I didn't bother with any of that before. Now you can switch language.

I noticed a lot of visitors to this site from China, Korea, India... so this should help!

💡 Quick Tips & Best Practices

🎓 Lessons Learned So Far in 2026

1. The tool matters MORE than the model - Claude Code with Deepseek > Bad tool with Claude Opus
2. Planning is worth the investment - Spend 3x longer planning vs coding
3. Multi-model workflows work - Different models excel at different things
4. Context curation is king - Don't dump everything, be strategic
5. Cheap models are getting REALLY good - GLM 4.7 and Deepseek v3.2 rival expensive models for many tasks

📊 Cost Comparison: February 2026

Here's what I'm actually spending monthly for unlimited AI coding:

Compare this to:

• Claude Pro MAX: $200/month (and you still hit limits!)
• Cursor Pro: $40/month
• Other premium IDEs: $30-100/month

🚀 What's Next?

I'm experimenting with:

• Multi-model review workflows
• Automated spec generation
• Context pruning with dedicated models
• MCP server automation strategies

Check back next month for updates!

Last updated: February 4, 2026