The last six months changed the AI coding tool landscape faster than any comparable period. Cursor announced $2 billion in annual recurring revenue and topped Gartner’s developer tool category. Claude Code went from a terminal plugin to an autonomous agent that companies run on long migrations. GitHub Copilot added agentic features that finally make the enterprise license defensible.
This is a grounded look at where things actually stand.
Cursor’s $2B ARR number matters because it’s not a valuation, not a funding round — it’s users paying money every month. That’s the clearest signal that AI-augmented coding has crossed from “interesting experiment” to “workflow dependency.”
What makes Cursor work for teams is the codebase context. It doesn’t just autocomplete functions; it understands the project structure, the naming conventions, the patterns in existing code. When you ask it to add a feature, it generates code that looks like it was already in the repo.
The weakness in 2026 is still the same as 2025: it struggles with tasks that span multiple repositories or require understanding organizational context outside the codebase. Fix that, and the $2B looks conservative.
Claude Code’s positioning has shifted. It’s not competing with Cursor for the day-to-day autocomplete workflow. It’s competing for the 10% of tasks that take 90% of the time — large refactors, complete feature implementations, complex migrations.
The Claude Fable 5 release made this explicit. Autonomous multi-day coding sessions are now the flagship use case. Anthropic is marketing to the engineering manager who wants to hand off a migration to an agent overnight and review the PR in the morning.
Whether that workflow is real or still aspirational depends on the organization. For greenfield projects with clean codebases, it’s often real. For legacy systems with years of undocumented assumptions, it’s still rough.
Copilot’s new agent features aren’t trying to out-innovate Cursor or Claude Code. They’re trying to make the IT department’s job easier by keeping everything inside the Microsoft ecosystem.
For a 5,000-person enterprise with an existing GitHub Enterprise contract, adding Copilot agent features is a single procurement decision. Deploying Cursor or integrating Claude Code is a different conversation — security reviews, legal review, data handling policies.
Copilot is winning the deals it’s winning because of distribution, not because of capability.
Three questions cut through the marketing:
Does it understand your codebase? Not just syntax — the patterns, the conventions, the architectural decisions that aren’t documented anywhere but are obviously there in the code. Tools that index your repository answer this better than tools that don’t.
Can it handle your failure cases? Every AI coding tool fails. The ones worth paying for fail in ways you can recover from: they stop and ask rather than confidently generating wrong code. Evaluate tools on the quality of their failures, not just their successes.
Does it integrate with your review process? The best AI coding tool is the one your team actually uses. If it generates code that bypasses your normal review workflow, you’ll eventually have a security or quality problem. The tools with GitHub/GitLab integration win here.
Token efficiency. The next competition isn’t “which model is smartest” — it’s “which tool completes this task using the fewest tokens.” For teams running agents on complex codebases, token costs are real budget line items.
LLM providers are aware of this. Anthropic explicitly mentioned that Fable 5 uses fewer reasoning tokens than comparable models for equivalent output quality. Expect this to become the primary performance metric by year end.
All ARR and benchmark figures cited are from public announcements and developer surveys as of June 2026.