My AI Stack (May 2025)
I've decided to start a monthly series tracking my AI tool ecosystem. This serves two purposes: documenting my own workflow changes and providing a practical reference for those curious about how I'm using these tools day-to-day. Each update will cover subscription changes, new discoveries, and assessments of what's working (and what isn't).
Core LLM Subscriptions ($240/month total)
OpenAI: ChatGPT Plus ($20/month)
- Status: Downgraded from Pro ($200/month) in April
- Why the change: While o3 is really really good, I wasn't using it frequently enough to justify the Pro subscription. ChatGPT Plus has generous usage limits and meets most of my needs. Deep Research was great when it first came out, but Gemini's Deep Research is miles better, and I'm not really using the Operator or Sora (or GPT-4.5). I do miss o1-pro, I initially signed up for the Pro plan to get access to it, but o3 covers the needs more than enough.
- Future plans: Will reconsider Pro once o3-pro becomes part of the plan.
Anthropic: Claude Max ($200/month)
- Status: Upgraded from Claude Pro ($20/month) in May
- Why the upgrade: The recent addition of credits for Claude Code made Max significantly more valuable. Running multiple Code instances against my codebase, integrating with my Obsidian vault, and using it as a planning assistant has been just fantastic.
- Verdict so far: Extremely satisfied with the value. May eventually downgrade to Pro if OpenAI improves their models' instruction following and enables MCPs in Codex (or if there's a solid Aider MCP integration).
Google: Gemini Advanced ($20/month)
- Status: Maintained subscription
- Strengths: Gemini-pro-2.5 might be the highlight of last month. Exceptional for writing tasks, handles large codebases well, and offers substantial response lengths ideal for summarization. The Deep Research feature (updated in April 2025 to use pro-2.5) outperforms competitors.
- Bonus value: Also making extensive use of the free API credits (25 daily per Google account).
AI-Powered Products & Applications ($10/month active)
Still In My Toolkit
GitHub Copilot ($10/month)
- Status: Long-term subscriber, maintaining
- Key integration: Recently enhanced with Gemini 2.5.
- Why I keep it: Despite having Claude Max, Copilot's real-time coding assistance remains valuable. I particularly enjoy using Roo; it's really fun to give a bunch of tasks to it and just "let it rip".
Warp Terminal (Free plan)
- Status: Regular user, staying on free plan
- What works: Clean interface with thoughtful LLM integration.
- Decision point: Would have considered the paid plan if not for Claude Max subscription.
- Verdict: Excellent terminal experience that balances aesthetics with functionality.
On the Chopping Block
Perplexity Pro ($20/month)
- Status: Have yearly subscription but likely canceling
- Why it's on the chopping block: Usage has declined significantly as other tools improved. While it still outperforms ChatGPT 4o, Gemini, and Claude for web search (though not o3), I've found myself using dedicated Deep Research features instead.
- Feature loss: Previously used it to access Deepseek models, but they removed the "No Web Search, just call LLM" option.
Recently Canceled
Cursor IDE ($20/month)
- Status: Canceled in May after being a user since August 2024
- What changed: Quality noticeably declined after they reduced context limits (around December/January). The MAX requests are great but I don't see the point of paying additional money for them, when I can just call the API and know exactly what I'm feeding into the model.
- What I miss: Their non-LLM autocomplete functionality is excellent.
- Future plans: May reconsider if they offer a bundle of ~50-100 MAX requests monthly.
Superwhisper ($10/month)
- Status: Canceled after 3 months
- Use case: Voice-to-text for "vibe coding".
- Why I left: Gradually used it less as my work habits changed (more cafe coding, less talking). For simple commands, Apple/Siri transcription proved sufficient.
- Comparison note: Found it smoother and more responsive than Wispr Flow.
- Verdict: Good product that I might revisit if my workflow changes again.
Highlights of April 2025
- Gemini-pro-2.5: Best release of April, even considering o3. I can't overstate this. This model is multiple tiers above any of the previous Gemini's. It's the current best overall model; I prefer o3 for coding (mildly), and Sonnet for its speed and ability to rapidly iterate through multiple distinct tasks, almost like a hyperactive assistant tackling a checklist.
- o3: Finally happy to have access to this; I have been waiting for this since December last year. I really enjoy that it treats all requests as mini Deep Researchs (goes back and updates its todo list, randomly runs python code to check its results), and I like the code that it writes (very smooth and clean). It's probably the smartest model out there (but I'm learning that smarts are not just what I need from a model).