AI / Field Guide
Which AI coding assistant should you actually run?
Ten tools, two tiers, and one uncomfortable finding: most working engineers no longer pick one. This page sizes up the AI code assistants developers really use in 2026 - the editors that sit around your cursor and the agents that live in your terminal - and gives a verdict on each rather than a feature grid.
About 9 min read. Prices, versions, and survey figures verified August 2026.
The field splits into two tiers, not ten products
Every tool below belongs to one of two shapes. The IDE tier puts the model next to your cursor: inline completion, a chat panel that can see the open file, and an agent that edits within the editor's model of the project. You stay in the loop keystroke by keystroke. The terminal-agent tier hands the model a shell, a file tree, and a goal, then gets out of the way for minutes at a time. You review the diff, not the typing.
That distinction predicts almost everything else - pricing model, how much context the tool burns, how bad a mistake gets before you catch it, and whether it can run unattended in CI. It also explains why so few developers commit to one product.
70 percent run two to four tools
The Pragmatic Engineer surveyed 906 engineers between January 27 and February 17, 2026. Seventy percent used two to four AI tools; only 15 percent used exactly one, and another 15 percent used five or more. Single-vendor loyalty is the minority position.
Adoption is near-universal, trust is not
JetBrains' January 2026 wave of more than 10,000 professional developers found 90 percent regularly using at least one AI tool at work and 74 percent on a developer-specific tool. Stack Overflow's 2026 survey put overall use-or-plan-to-use at 84 percent while trust in the output kept falling.
Money follows the agents
Industry estimates put the AI coding tools market near 12.8 billion dollars in 2026. GitHub Copilot still leads on paid seats, reported around 4.7 million, but the growth is in agentic tools, and JetBrains measured Copilot workplace adoption at 29 percent against Cursor and Claude Code tied at 18 percent each.
The IDE tier: assistants that live around your cursor
These are editors, or extensions to one. The model sees what you have open, completion fires as you type, and an agent mode handles the multi-file work without leaving the window. Best when you are actively authoring code and want to steer every few seconds.
Cursor
The default AI editor
A VS Code fork with the AI parts rebuilt rather than bolted on. Tab completion predicts multi-line edits and next-cursor position, Cmd+K does scoped inline edits, and Agent handles multi-file changes with a Plan mode in front of it. Anysphere reported passing 2 billion dollars in annualized revenue in early 2026, which buys a shipping cadence nobody else in the editor category matches. It also ships its own in-house model, Composer, tuned for low-latency agent loops and available only inside Cursor.
Verdict: the best place to do fast, interactive editing with AI. Hobby is free, Pro is 20 dollars a month, and the Pro+ and Ultra tiers exist because heavy agent use will exhaust a 20 dollar credit pool. cursor.com/docs
GitHub Copilot
The incumbent, and the safe procurement answer
Still the widest install base by a distance, and the only one of these your security team has probably already approved. It runs inside VS Code, Visual Studio, JetBrains IDEs, Neovim, and github.com itself, with agent mode, a cloud coding agent that opens pull requests, and automated code review on PRs. The free tier gives 2,000 completions and 50 chat requests a month - genuinely useful for a student or a hobby repo.
Verdict: pick it when the org already pays for GitHub and you want one vendor, one invoice, and one audit trail. Pro is 10 dollars a month with 15 dollars of AI credits; Pro+ is 39 and Max is 100, each with a larger credit pool. github.com/features/copilot
Devin Desktop
Formerly Windsurf, rebranded June 2, 2026
Cognition acquired Windsurf and on June 2, 2026 relaunched it as Devin Desktop, folding the editor into the same brand as its autonomous Devin agent. The rebrand shipped real features, not just a logo: an Agent Command Center that manages local and cloud agents in one Kanban view, Spaces for sharing context between agents, and Devin Local, a Rust rewrite of Cascade with subagent support. It also opened the Agent Client Protocol so Codex, Claude's agent, and OpenCode can run inside it.
Verdict: the most interesting bet in the category - an editor that assumes you will run somebody else's agent inside it. Do not call it Windsurf; that name is retired. devin.ai
Zed
Fast editor first, AI second
Written in Rust with GPU-accelerated rendering, and the only editor here where the AI is optional rather than the point. Agent panel, inline assist, and bring-your-own API key or your existing subscription; there is a free tier and no requirement to use the hosted models at all. Multiplayer editing and the built-in terminal make it a credible daily driver even with every AI feature switched off.
Verdict: the choice for developers who resent how heavy VS Code forks have become and want AI on their own terms. Thinner extension ecosystem is the price. zed.dev
JetBrains AI Assistant and Junie
In-IDE assistant plus a coding agent
JetBrains ships two things: AI Assistant, the in-editor completion and chat layer, and Junie, an agent that plans and executes multi-step tasks inside IntelliJ, PyCharm, GoLand, and the rest. Both are bundled into the AI Pro subscription that comes with an All Products Pack. Their own January 2026 survey put AI Assistant at 9 percent regular use and Junie at 5 percent - modest numbers, but the tools have something nobody else does: the IDE's real static analysis and refactoring engine underneath, so a rename is an actual rename rather than a regex the model hopes is right.
Verdict: obvious if your team already lives in JetBrains IDEs, especially for Java, Kotlin, and large statically typed codebases. Not worth switching IDEs for. jetbrains.com/ai
The terminal-agent tier: assistants you hand a task to
No editor, no completion, no cursor. You describe an outcome, the agent reads files, runs commands, edits, tests, and reports back. The unit of review is a diff or a pull request. Best for work that spans many files, for refactors, and for anything you want running unattended.
Claude Code
The reference terminal agent
Anthropic's CLI reached 18 percent workplace adoption in JetBrains' January 2026 wave, with the highest loyalty scores in the survey at 91 percent satisfaction and an NPS of 54. In the Pragmatic Engineer survey it was named most-loved by 46 percent of respondents against Cursor's 19 and Copilot's 9. The reason is architectural: a small set of composable primitives - CLAUDE.md memory, skills, subagents, hooks, MCP servers, and plugins that bundle all of them - rather than a fixed feature list. It also runs headless, so the same tool that answers your prompt can gate a CI job.
Verdict: the best agent for complex, multi-file, long-running work, and the most configurable tool in this list. Included with Claude Pro at 20 dollars a month; Max at 100 or 200 buys 5x or 20x the usage. Full reference on the Claude Code cheatsheet.
OpenAI Codex CLI
The fast follower with the aggressive sandbox
Codex runs in the terminal, in a VS Code extension, on the web, and as a GitHub reviewer, all off one ChatGPT login. Its distinguishing design choice is separating what the agent is allowed to do from when it asks: sandbox_mode controls filesystem and network access, approval_policy controls the prompts. Subagents reached general availability on March 14, 2026, with explorer, worker, and default roles running in parallel, and codex exec gives you a read-only non-interactive mode for pipelines.
Verdict: strongest choice if your team already pays for ChatGPT, and the best sandbox story in the category. Compared head to head in Claude Code vs Codex.
Aider
Open source, git-native, still excellent
Aider predates the current wave and has aged better than most of it. It pairs with any model through your own API key, builds a repository map so it can work in large codebases without loading everything, and commits each change with a sensible message so the git history is the audit log. No subscription, no telemetry story to argue about, and total control over which model sees your code.
Verdict: the pick when you must bring your own model, run against a local model, or keep costs strictly metered per token. Less polished than the funded tools and it shows. aider.chat
Cline
Apache-2.0, provider-agnostic, now with a CLI
Cline started as a VS Code sidebar agent and spent 2026 growing into a platform: a matured standalone CLI, an SDK, parallel agents, and support for 30-plus model providers including local ones. It asks approval at each step by default, which is either the feature or the friction depending on your temperament. Open source under Apache 2.0 with millions of installs.
Verdict: the best open-source middle ground - agentic power with an explicit approval gate, and no lock-in to a single model vendor. cline.bot
Amp
Sourcegraph's agent, with an unusually transparent bill
Amp runs in the terminal with editor integrations for VS Code, JetBrains, Neovim, and Zed. Sourcegraph's pitch is pricing honesty: a free tier that gets you a working agent, and pay-as-you-go access to frontier models at cost with no markup on top. The free tier's funding model has changed at least once during 2026 - it launched ad-supported and Sourcegraph has since revised that - so check the current terms before you standardize a team on it.
Verdict: worth a look if you want a capable agent without a subscription and you would rather see the model bill than a credit abstraction. ampcode.com
The field at a glance
Entry pricing as published in August 2026. Every paid tier here meters usage in some way, so the headline number is a floor, not a ceiling.
| Tool | Shape | Entry price | Best at |
|---|---|---|---|
| Cursor | AI editor (VS Code fork) | Free, Pro $20/mo | Interactive editing, inline completion, tight feedback loops |
| GitHub Copilot | Extension for many IDEs | Free tier, Pro $10/mo | Org-wide rollout, PR review, breadth of IDE support |
| Devin Desktop | AI editor plus cloud agents | Free tier, Pro $20/mo | Managing many agents at once, running third-party agents inside one shell |
| Zed | Native editor with optional AI | Free tier, bring your own key | Editor speed, collaborative editing, AI you can turn off |
| JetBrains AI | Assistant plus Junie agent | Bundled with AI Pro subscription | Refactors backed by real static analysis in JVM and Python codebases |
| Claude Code | Terminal agent | Claude Pro $20/mo | Long multi-file tasks, deep customization, headless CI runs |
| Codex CLI | Terminal agent plus web and IDE | Included with ChatGPT Plus $20/mo | Sandboxed autonomy, parallel subagents, one login across surfaces |
| Aider | Open-source terminal agent | Free, you pay the model API | Git-native commits, local models, strict per-token cost control |
| Cline | Editor extension plus CLI | Free, you pay the model API | Step-by-step approval, 30-plus providers, open licence |
| Amp | Terminal agent plus editor plugins | Free tier, then at-cost model usage | Transparent billing, no subscription commitment |
Gotcha: credit pools are not comparable across vendors. Cursor grants a dollar-denominated pool, Copilot grants AI Credits, Anthropic grants a usage multiple over a rolling window, and Codex draws on your ChatGPT quota. The only honest comparison is a week of your own work on each.
Our pick per workflow
Cursor for editing, Claude Code for the hard parts
The dominant combination in 2026 is not a compromise, it is a division of labour. Keep Cursor open for the work where you are the author - writing a component, fixing a test, following a stack trace, accepting a completion every few seconds. Hand Claude Code the work where you are the reviewer - a refactor across forty files, a migration, an upgrade that touches every import, a bug you can describe but not locate. They cost 40 dollars a month together and they do not overlap much.
If you can only pay for one
Take the terminal agent. Completion is the feature you will miss least; unattended multi-file work is the one you cannot replicate by typing faster.
If procurement owns the decision
GitHub Copilot. It is already on the approved list, it covers every IDE your team uses, and the free tier lets sceptics try it before a seat is bought.
If your code cannot leave the building
Aider or Cline against a self-hosted model. Both are open source, both let you point at any endpoint, and neither requires a vendor subscription.
Four things the marketing pages leave out
Benchmarks move faster than your habits. Terminal-Bench and SWE-bench leaderboards reshuffle every few weeks, and the gap between the top two agents is usually smaller than the gap between a good prompt and a lazy one. Switching tools because a scoreboard moved two points is a bad trade against the week you spend re-learning a workflow.
The model matters more than the wrapper. Most of these tools can run most frontier models. What differs is the harness around the model: how it searches your repo, how it decides what to load into context, how it recovers from a failed command. That harness is worth real money, but do not confuse it with the intelligence.
Rules files are now portable, mostly. AGENTS.md has become the near-universal convention, and Cursor reads it directly. Claude Code reads CLAUDE.md and not AGENTS.md, but a one-line @AGENTS.md import bridges the two, so a repo can carry one set of conventions for every agent that visits it.
Trust is falling while adoption rises. Stack Overflow's 2026 numbers show a growing share of developers who use these tools daily and explicitly do not trust their output. That is not a contradiction, it is a review discipline problem - and it is exactly what the AI-assisted development workflow guide is about.
Settle the close calls
Claude Code vs Cursor
Terminal agent against AI editor, and why the answer is often both.
Read the verdictCursor vs Copilot
The dedicated AI editor against the one your company already bought.
Read the verdictClaude Code vs Codex
Two terminal agents with genuinely different safety models.
Read the verdictClaude vs GPT for coding
The model layer underneath every tool on this page.
Read the verdictOnce you have picked, the reference pages get you productive: the Claude Code cheatsheet, the Cursor cheatsheet, and the VS Code shortcuts cheatsheet that still applies to every fork on this page. To give any of these agents access to your own tools and data, start with the Model Context Protocol explainer.
Building AI features into a product rather than using AI to write it? That is a different layer: see AI APIs and SDKs and the AI app stack guide. If you want a working prototype before you write any code at all, AI app builders covers that category.