What Changed in AI Coding by 2026
The first wave of AI coding tools sold autocomplete. That is now a free feature in the editors most developers already open every morning, which removed the entire business model of a dozen startups. Kite was the clearest casualty, and its story is the template for the rest of this page.
The tools that survived moved up the stack. They review pull requests, run tests, explain unfamiliar code, and take on small chores end to end. That work is harder to copy and easier to charge for, which is why the survivors all look more like teammates than like autocomplete.
The second change is who pays. Inference costs real money on every keystroke, so a coding tool needs either a platform behind it or a high price. Anything in between gets squeezed. You can see the same pattern across the whole graveyard, and the status leaderboard tracks it week by week.
The Tools Still Worth Building On
These four are the ones we would put a team on today. Each still ships features, answers support tickets, and has a clear owner paying the compute bill.
GitHub Copilot — the safe default
Copilot is the boring right answer for most teams. It lives inside the editors and review flows you already use, and it bills through an account your finance team already approved. Suggestion quality is no longer its main advantage; distribution is. Read our ranked Copilot alternatives guide if pricing or policy rules it out, and see vendor terms at github.com/features/copilot.
Cursor AI — the AI-first editor
Cursor asks for more: it wants to be your editor, not a plugin inside it. Teams that accept the trade get the tightest multi-file editing loop available, because the whole interface is designed around asking for changes instead of typing them. Developers who love their existing setup will resist, so trial it with volunteers first. Vendor details at cursor.com.
Codeium — the budget pick that stayed alive
Codeium built its base on a genuinely usable free tier, then converted teams to paid plans with self-hosting options. Check the live Codeium status check before standardising, and browse tools like Codeium for side-by-side context. It is the tool we recommend when the honest constraint is budget.
Tabnine — for code that cannot leave the building
Tabnine sells to banks, health systems, and defence suppliers, which means private deployment and training guarantees are the product. Raw capability trails the leaders slightly. Compliance teams do not care, and that is exactly why the company still has revenue.
At Risk: Useful Today, Uncertain Tomorrow
Devin — the agent that promised too much, too early
Devin made autonomous coding a mainstream idea, and then had to live up to a launch demo. It is genuinely good at scoped chores with strong tests around them, and genuinely unreliable in messy legacy code. Our Devin case study covers the gap between claim and delivery, the status check covers where it stands now, and Devin alternatives covers what to try instead.
Windsurf — acquired, so the roadmap moved
Windsurf is still usable, but an acquisition changed who it serves. When a product's priorities move to a parent company, features that mattered to small teams tend to slip down the list. Read the Windsurf status check and keep Windsurf alternatives on file before you make it a requirement in onboarding docs.
Dead: What Kite Teaches Everyone Else
Kite did local ML completion years before it was fashionable, with real engineering behind it. Then large models made better suggestions available inside editors for free, and a paid completion product had nothing left to defend. The team open-sourced its work and closed down.
The lesson is uncomfortable and useful: being early and technically strong does not protect you when a platform decides your feature is now a checkbox. For the full story see what happened to Kite, why Kite failed, and current replacements in Kite alternatives and best tools like Kite.
Side-by-Side: What Each Tool Is Actually For
| Tool | Status | Best for | Main catch |
|---|---|---|---|
| GitHub Copilot | Active | Teams already on GitHub | Least exciting, hardest to fault |
| Cursor AI | Active | Multi-file refactors | Requires changing editor |
| Codeium | Active | Tight budgets, self-hosting | Smaller ecosystem |
| Tabnine | Active | Regulated codebases | Capability trails leaders |
| Devin | Active, at risk | Scoped, well-tested chores | Expensive supervision |
| Windsurf | Acquired | Existing users, short term | Roadmap serves new owner |
| Kite | Shut down | Nothing — use a replacement | Product is gone for good |
Five Signals a Coding Tool Is About to Die
These checks take about ten minutes and none of them is proof alone. Two together is your cue to stop depending on the tool.
- An empty changelog. Ninety quiet days in this market means the engineers left or moved to something else internally.
- A stale extension listing. Marketplace pages are cheap to update. When they drift out of date, nobody is minding the shop.
- Pricing drift. Cheap tiers disappearing and annual-only plans appearing both say the same thing: the vendor needs bigger cheques from fewer customers.
- Support silence. Ask one real question from a paid account. Two days is healthy. A week of nothing is a measurement, not a mood.
- Model downgrades. Quietly swapping to a cheaper model, shrinking context windows, or adding queues usually means the compute bill got scary.
Every tool page here carries its own check history, so you can see what we saw and when. If you spot a change before we do, report it and we will verify it.
Three Stacks That Work Right Now
The small startup
Copilot for everyone, plus Cursor for the one or two developers doing heavy refactors. Keep prompts and rules in the repository rather than in a vendor account, so switching later is a configuration change and not a rewrite of habits.
The regulated enterprise
Tabnine or a self-hosted Codeium deployment as the standard, with a written data policy your security team signed. Add one narrow agent use case, such as dependency upgrades, and measure it against a human baseline before widening scope.
The agency or consultancy
Whatever the client's stack allows, with your own toolchain as the fallback. Bill supervision time honestly: agents shift effort from writing code to reviewing it, and pretending otherwise is how projects overrun.
Five Mistakes That Cost Teams Real Money
- Buying annual seats in month one. The discount is small and the lock-in is not. Run one real sprint first.
- Skipping the data question. Free tiers often reserve broader rights than paid ones. Read it before installing on a private repo.
- Measuring accepted suggestions. Acceptance rate flatters the vendor. Review time and defect escape rate flatter reality.
- Letting agents touch untested code. Without tests, an autonomous change is a guess you have to verify by hand anyway.
- Writing one tool into onboarding docs. Name the capability, not the vendor, so a shutdown is a swap rather than a rewrite.
What These Tools Really Cost Per Developer
List prices hide the real number. The honest figure is the seat price plus the supervision time each tool adds or removes, and supervision is where budgets quietly disappear. A cheap agent that needs an hour of review per task is more expensive than a paid assistant that saves twenty minutes a day.
Work it out per developer, per month, using your own blended hourly rate. Assume a mid-level engineer costs somewhere between forty and eighty dollars an hour fully loaded. At that rate, a twenty dollar seat pays for itself if it saves half an hour a month, which every tool on the active list clears easily. The interesting question is never whether an assistant pays back. It is whether an agent does.
Three costs teams forget. First, onboarding: a new editor costs a week of reduced output even when developers like it. Second, review load: more generated code means more careful reading, and senior time is the scarcest thing you have. Third, exit cost: the hours spent rebuilding rules, prompts, and habits when a vendor closes. Kite users paid that third cost in full, and the ones who had kept their configuration in the repository paid the least.
Where Agents Help and Where They Hurt
Autonomous coding agents are the most oversold and most misunderstood tools in this category. They are excellent at work that is boring, repetitive, and well covered by tests: version bumps, framework migrations, adding missing test cases, and mechanical refactors across many files. In those lanes an agent genuinely replaces hours of tedious typing.
They hurt in exactly the opposite conditions. Ambiguous requirements, thin test coverage, and old code with undocumented assumptions turn a confident agent into a fast source of wrong answers. One bad assumption becomes twenty files that all look plausible, and a reviewer who trusts the diff ships the mistake. That is why we treat agent output the way a good team treats a new hire: useful from day one, unsupervised much later.
A practical rule: give an agent work you could describe fully in a ticket, and never work you would need a meeting to explain. Then measure the outcome in review time and escaped defects, not in lines produced. If review time goes up and defects stay flat, the agent is costing you money in a way no invoice will show.
Make Yourself Easy to Rescue
Every tool on this page will eventually change price, owner, or existence. The teams that shrug at that news share four habits, and none of them is expensive to adopt this week.
- Keep configuration in the repository. Rules, prompts, and coding standards belong in version control, not in one vendor's settings page.
- Name capabilities in your docs. Write "AI assistant" in onboarding, not a brand name, so a swap is a one-line change.
- Keep one fallback installed. A second assistant on a free tier costs nothing and turns a shutdown into a ten-minute switch.
- Re-check status quarterly. Put a recurring reminder on the calendar and use our tool pages as the source.
What We Expect Next
Expect more bundling and more quiet acquisitions. Platforms will keep absorbing single-feature tools, and the strongest independents will keep moving toward review, testing, and maintenance work where they can prove savings. Agents will get better in narrow lanes long before they are trusted with whole features.
For the wider market context, the Stack Overflow developer survey tracks adoption year over year, and our monthly status reports record each shutdown as it lands. We update this page whenever a status above changes.
Where to Go Next
Comparing two tools head to head? Start at the versus index. Watching the whole market? The status leaderboard ranks every tracked tool by confidence. Building a wider stack? Pair this page with AI tools for writing blog posts and the marketing stack.
Bottom line: pick the tool that fits your editor and your data rules, keep configuration in your own repository, and re-check status every quarter. Do that and a shutdown costs you an afternoon instead of a quarter.