Four Layers, Four Different Risks
Developers rarely buy one AI tool. They accumulate four, each with a different failure mode and a different blast radius when it goes away.
Name the layer before you compare products. The editor assistant is a preference. The model API is a dependency your users feel.
Verified Status of the Developer Stack
Every row below is checked on a schedule, and anything that looks like a shutdown is reviewed by a person before the label changes.
| Tool | Layer | Status | What it means for you |
|---|---|---|---|
| GitHub Copilot | Editor | Active | Safest default; distribution and billing already exist |
| Cursor AI | Editor | Active | Strongest AI-first editor if your team will move |
| Codeium | Editor | Active | Best value per seat; wide editor coverage |
| Tabnine | Editor | Active | Self-hosted option when code cannot leave the network |
| OpenAI API | Model | Active | Broadest tooling; pin versions and watch deprecations |
| Anthropic API | Model | Active | Strong long-context reasoning; good second provider |
| Cohere | Model | Active | Enterprise and retrieval focus rather than consumer chat |
| Hugging Face | Model hub | Active | Open weights and hosting; read the status history |
| Devin (Cognition) | Agent | Active | Useful on scoped chores; see alternatives |
| Windsurf | Editor | Acquired | Roadmap now serves its owner; usable, not standardisable |
| MosaicML | Infrastructure | Acquired | Absorbed into a larger platform; training plans changed |
| Kite | Editor | Shut down | Dead; the replacement list is here |
Shutdown Risk Scoreboard
Risk here scores the company, not the quality of its completions. Funding, distribution, shipping pace, and pricing behaviour decide who is still selling seats next year.
Deprecation Is the Real Outage
Most developer pain in this category is not a company dying. It is a model version being retired while your prompts quietly stop behaving. The failure is silent: valid responses, worse content, no error in your logs.
Treat the model like a dependency. Pin the version, keep a small evaluation suite of real inputs with expected shapes, and run it before you move. Read the provider's deprecation policy in the OpenAI deprecations list and the Anthropic model deprecations page, and put the dates in the same calendar as your certificate renewals.
How Developer AI Tools Actually Die
Across the tracked graveyard, four causes explain nearly every developer-tool failure. None of them is "the completions got worse".
Seven-Day Developer Stack Audit
Run this once a year and no vendor decision becomes an emergency.
- Inventory by layer. List every AI tool in the editor, the build, and production, with owner, plan tier, and monthly cost.
- Find the hard dependency. Mark anything whose disappearance would break a shipped feature. That short list is your only real risk.
- Pin versions. Replace every implicit "latest" model reference with an explicit version in configuration.
- Build the golden set. Fifty real inputs, expected output shapes, runnable in one command. This is your insurance policy.
- Check the terms. Confirm training exclusion and retention on each paid seat, and save a dated copy of the policy.
- Test a second provider. Point the adapter at an alternative model for one afternoon and record what broke.
- Write one page. Approved tools, review rules for agent diffs, deprecation calendar, next review date.
Five Mistakes Teams Keep Making
- Spreading vendor SDK calls across the codebase. One adapter turns a migration from a project into a config change.
- Calling "latest" in production. The upgrade lands without a test run and quality drifts unnoticed.
- Letting agents merge unattended. Scoped chores are fine; ambiguous work in legacy code is where a wrong assumption multiplies.
- Installing free tiers on private repositories. The generous tier is usually the one that reserves rights to your code.
- Standardising on an acquired tool. Use it if it helps, but do not make it required while its roadmap belongs to someone else.
What This Actually Costs
Editor assistants land around ten to forty dollars per developer per month, and enterprise tiers with policy controls sit above that. The model API is the line that surprises finance, because it scales with usage rather than headcount: a feature that costs cents in staging can cost real money at launch traffic. Agents are priced per task or per seat and are the easiest spend to justify or cancel, because their output is measurable in merged pull requests.
The number nobody budgets is switching cost. A stack with one adapter, pinned versions, and a golden evaluation set moves providers in a day. A stack with vendor calls scattered through the codebase takes weeks, and those weeks arrive exactly when a deprecation notice does.
Keep Exploring the Graveyard
Developer tooling overlaps with the rest of the stack, and the portability question repeats everywhere. If you are auditing more broadly, see AI tools for coding and code review, AI tools for research, AI tools for writers, and the full collection index.
Want the underlying data? Browse every tracked tool from the homepage, compare survivors on the leaderboard, read the latest postmortems, or report a tool you believe is winding down. For background on how the wider ecosystem is consolidating, the Stack Overflow developer survey tracks adoption year over year.
Bottom line: buy per layer, keep one adapter between your code and any model, pin versions, hold a golden evaluation set, and never let an acquired tool become required. Do that and the next shutdown costs you a subscription, not a release.