AI Coding Tools Late 2025: What We Kept vs Cancelled
Eight months after our first AI tools experiment, here's what actually stuck. Real costs, productivity data, and the tools we quietly cancelled.
This post was originally written in October 2025. I've updated it in March 2026 to reflect where the stack actually landed — because a fair bit has changed.
Eight months ago, I wrote about testing AI coding tools in our Brighton studio. The verdict then: cautiously optimistic. The tools showed promise, saved time, but had clear limitations.
It's late 2025 now. The landscape has shifted. Some tools evolved into daily essentials. Others quietly disappeared from our workflow. And a few new ones forced their way in.
Here's the update nobody's writing: what actually survived when the hype wore off, and what we're paying for today versus what we cancelled.
What Changed Since March
The AI coding tool market in 2025 feels different than six months ago. Less breathless hype, more quiet consolidation. Three big shifts:
1. The "AI IDE" wars heated up VS Code added native AI features (finally). Cursor improved dramatically. Every editor now has some form of AI built in or available as a first-party extension.
2. Agentic coding arrived properly The shift from "AI that autocompletes" to "AI that takes actions" happened faster than I expected. Tools like Claude Code can now open files, run commands, write tests, and iterate on a task for hours without supervision. That's a qualitatively different thing.
3. Developer fatigue set in Every week there's a new AI coding tool promising 10x productivity. Most developers I talk to have stopped experimenting and picked their core stack. Tool fatigue is real.
For us, that meant a hard look at what we're actually using versus what we thought we'd use.
The Current Stack: What Survived
After months of real production use, here's what's still in our toolkit. It's shorter than I expected.
Cursor — Primary IDE
Status: Kept Usage: Primary development environment for all client and side project work
Cursor replaced VS Code. It's just better for the way I work now — the multi-file context, the ability to chat with the codebase, the Composer mode for large refactors. I'm not going back.
The thing that makes it worth the subscription is understanding unfamiliar codebases. When I pick up a project I haven't touched in six months, or take on a legacy codebase from another developer, Cursor dramatically reduces the time spent working out what's going on before I can start doing something useful.
Claude Code (Max) — Agentic Work
Status: Kept Usage: Background tasks, longer autonomous work, anything that needs to run while I'm doing something else
Claude Code is Anthropic's agentic coding tool — it runs in the terminal, can access your codebase, run tests, make changes, and iterate. I run it with a Max subscription, which bundles Claude Code access with the full Claude model.
The difference between this and IDE-based AI: I can hand it a task, close my laptop, and come back to a pull request. It's not magic — the output still needs review — but it removes me from the loop on work that's well-defined enough to delegate. Bug fixes, test suites, API integrations, dependency updates. Things that need doing but don't need me doing them.
I've also got it running on my Mac Mini for side project work, which I wrote about separately.
skills.sh — The Glue Between Tools
Status: The thing I didn't expect to care about this much
This is the one I'd have left off a list six months ago, because it solves a problem I hadn't fully named yet.
The problem: Cursor and Claude Code are both AI tools, but they don't share anything. If you've taught Cursor your coding conventions — how you structure components, which patterns you prefer, what your testing approach looks like — none of that transfers to Claude Code. Every tool starts from scratch.
skills.sh fixes this. It's an open ecosystem of reusable AI skills — packages of procedural knowledge you install once and share across tools. A single command (npx skills add owner/repo) installs the skill into your project, and then any AI tool that supports it — Cursor, Claude Code, others — picks it up automatically.
The result is that when I ask Claude Code to write a test, it follows the same conventions as when Cursor writes a test. The AI is consistent. It knows the project's patterns, not just the project's files.
This is the thing that moved AI tooling from "clever autocomplete" to "actually knowing how I work." I use it heavily.
Warp — Terminal
Status: Kept Usage: Daily
Warp is a terminal with AI built in. Not glamorous, not surprising at this point, but it earns its place. The main thing it does is make the terminal less of an interruption to thinking — I can describe what I want to do in plain language and it generates the command. For anything outside my regular repertoire (obscure git operations, specific ffmpeg incantations, sysadmin work on client servers), this is genuinely useful.
It also has a command history and search that's better than zsh's, which shouldn't matter but does.
What We Cancelled
Here's the harder truth.
GitHub Copilot — Gone
Copilot was the first AI tool I used seriously. For a long time it was the foundation of the stack. But once Cursor became the primary IDE, Copilot became redundant. Cursor's AI is better in every way that matters to me — more context, more capable, actually understands the codebase. Copilot's autocomplete is fine, but I don't need fine when I've got Cursor.
Cancelled.
Claude Pro — Replaced by Max
I used to have a separate Claude Pro subscription for architectural discussions, code review, and debugging complex problems. The Max subscription bundles this with Claude Code access, so the separate subscription went away. Same model, better deal, one less thing to track.
Devin — Never Worked
Tried it for two months. It never delivered on the autonomy promise. The output looked plausible and was usually subtly wrong. Claude Code does what Devin promised and actually delivers on it. Cancelled after month two.
ChatGPT Plus — Replaced by Claude
Claude is better at code. Easily. This one was straightforward.
Tabnine — Redundant
Tried as a Copilot alternative. Never better enough to justify running both. Gone before Copilot was.
What the Stack Actually Costs
Current monthly spend:
- Cursor: ~£16/month
- Claude Max: ~£80/month
- Warp: free tier (the paid features haven't been worth it yet)
- skills.sh: free
- Total: ~£96/month
That's more than I was spending on just Copilot and Claude Pro. But the capability difference is significant — Claude Code doing autonomous work is not in the same category as a chat interface. The cost is justified by what it actually does.
The Honest Summary
The stack simplified considerably. I was running six tools. Now I'm running three, with skills.sh connecting them.
The direction of travel is clear: less autocomplete, more autonomy. Less tool-per-task, more shared context. The AI tools that survived are the ones that actually understand the codebase and can take meaningful action in it. The ones that just sat in the editor offering suggestions are gone.
Six months ago I'd have told you the ceiling for AI coding tools was "saves a few hours a week on boilerplate." I'd revise that now. The ceiling is considerably higher — but reaching it requires tools that can actually be trusted to work unsupervised, and that requires shared skills and conventions, not just raw model capability.
That's where skills.sh fits. It's unglamorous infrastructure. It's the thing that makes the rest of it actually work as a coherent system rather than a collection of clever toys.
