Eighteen months, six tools, no scoreboard math
We have written separately about how ChatGPT reshaped our research phase, how Codex changed our build loop, why we build on OpenCode and what the arrival of Freebuff means. This piece is the one we would hand a client's developer over coffee: every tool we have used day to day since early 2025, with the opinions we would actually say out loud.
No benchmark tables. The benchmarks move every month and we do not trust ours anyway. What does not move quickly is how a tool feels when you use it daily, and how it behaves when the invoice arrives.
GitHub Copilot: great, watch the meter
Our verdict on GitHub Copilot is genuinely positive and comes with one loud asterisk: it is easy to run up a big bill. The tool is excellent — the inline suggestions still feel like magic on the third day of a project, and letting it loose on test files is one of the simplest productivity wins in the industry.
The bill part is structural, not a gotcha in disguise. Per-seat pricing plus consumption-based agent spending means the cost scales with how much you use it, and the branch between “this saved us an afternoon” and “why is this line item the size of a salary” is closer than you expect. We keep it, but we watch the meter the way you watch a taxi at night.
Codex: the same shape, more rope
OpenAI's Codex is the same story at a higher gear. It is a genuinely capable coding agent — our earlier piece covers how it earned its place in the build loop — and the pattern that emerged is identical to Copilot: brilliant on well-scoped tasks, and the marketing budget is the meter again.
Agents that bill by use are wonderful when your tasks are disciplined and expensive when they are not, because the failure mode is not the tool getting worse — it is the tool diligently working longer than it needed to. Scoped briefs, explicit acceptance criteria, and a human stopping rule: that is the whole trick, and it costs nothing.
ChatGPT: from convincing to collaborating
Our history with ChatGPT as a coding tool started ridiculous and got good — and it started out of pure necessity. The first time we leaned on it properly was the day we ran out of Codex usage, with work open and a meter sitting at zero. ChatGPT was harder to convince: you had to argue your way into it agreeing to write code at all, and when it finally did, the solution arrived as a zip file to download and paste in. Clunky delivery by today's standards, considerably harder than the tool we had just exhausted — but it did work, and we got it to do what we needed.
That era is over, and the tool we use now is closer to the research partner we described in the first of these stories. Planning, comparison shopping between architectures, drafting the ugly first version of a spec — that is where it earns its seat, and frankly where it has always earned it best. The zip jokes write themselves in retrospect, but there is something worth remembering in them: even in its clunkiest form, the value was real.
Cursor: a deserved shout
Cursor deserves its own shout, because it is the editor-native flagship that the whole category gets measured against, and for people whose comfort zone is a code editor rather than a terminal it is the most natural way into agentic work. Plenty of developers we respect run their whole day inside it.
We are a terminal-first studio, so our experimentation there is lighter than the others on this list and our opinion is correspondingly modest. What we can say without hedging: Cursor proved there is a large audience for agents that live inside the editor, and the CLI tools we prefer are all chasing the standards it set. That is influence worth acknowledging.
OpenCode: the quiet favourite
OpenCode is the one that seems to be really good. There is no koan there; that is the review. It is open-source, it runs in the terminal next to our build pipeline, the configuration lives as files in our repository, and after the honeymoon period with every other tool faded, this is the one whose weekly usage has kept growing quietly.
In our earlier piece we made the ownership argument: inspectable, portable, model-agnostic. Eighteen months of the model landscape reshuffling roughly every quarter has made that argument by itself. The harness stayed put while everything around it changed, and that turns out to be the property worth paying for — or in this case, not.
Freebuff: free, and not yet metered
Freebuff rounds out the set, and the practical note from months of use is disarmingly simple: we have never once run out of daily usage on the free plan. For a category where the paid tools have trained everyone to watch credits like a fuel gauge, a coding agent whose ceiling you simply have not hit is a genuinely odd experience.
Everything we said in the dedicated piece stands: ad-funded rather than subscription, so read the privacy terms properly and keep it aligned with its risk tier. But as a permanent second opinion sat next to the paid tools, and as proof that agent pricing had further to fall, it has taken a firm seat in the kit.
The scoreboard, for real
So after eighteen months, if you held us to one line each: Copilot — great, watch the meter. Codex — more rope, same meter. ChatGPT — the zip era is gone, and it is now our research desk. Cursor — the editor-native benchmark, well earned. OpenCode — quietly excellent and ours. Freebuff — free, and the lid has not come off the box yet.
The honest conclusion is the one we keep returning to: the tools rotate, the meters tick, and none of it changes the quality of your output. What changes output is specification, review and ownership — the process piece we will keep banging on about. Pick the tool that fits your stack and your budget this quarter; just make sure the process is yours.
