What you need to know
- Seven serious contenders. Claude Code, Google Antigravity, OpenAI Codex, Cursor, Kiro, GitHub Copilot and Windsurf are all credible in June 2026. There is no longer one obvious default.
- Four philosophies, not seven products. The real choice is terminal-native agent vs IDE-embedded vs spec-driven vs agent-first platform. Pick the philosophy first, then the product.
- Billing has changed under your feet. Flat-rate predictability is fading; usage-based plans now dominate. Heavy agentic runs cost real money, so model your busiest month before you commit a team.
- Lock-in is the quiet decision. Whether your prompts, specs and config live as plain files in the repo — or inside a vendor's platform — determines how cheaply you can switch later.
For most of 2024 and 2025, choosing an AI coding tool meant choosing between a small handful of options and accepting whatever workflow they imposed. That is no longer true. The last ninety days alone have reshaped the field, and a builder in Bengaluru or Bristol now faces a genuine shortlist rather than a default. The good news is that the decision is more rational than it looks once you stop comparing logos and start comparing philosophies. This guide gives you a framework, a comparison table and a clear set of questions — not a verdict.
The seven contenders, and the four philosophies behind them
It is tempting to rank the seven tools on a single line. Resist that. The more useful move is to recognise that they cluster into four distinct ways of working, and the cluster matters more than the brand.
- Terminal-native agents — Claude Code and the Codex CLI. The agent lives in your shell, reads the codebase, plans, executes with real developer tools, evaluates its own output and adjusts. Claude Code in particular operates at the project level rather than the snippet level.
- IDE-embedded tools — Cursor, GitHub Copilot and Windsurf. The intelligence sits inside an editor you already know, completing, refactoring and chatting within the familiar window.
- Spec-driven — Amazon's Kiro, which leans on written specifications and structured tasks rather than free-form prompting.
- Agent-first IDE and platform — Google Antigravity, designed from the ground up to let the agent drive the build rather than assist a human typing.
That last category is the newest. Google launched Antigravity 2.0 at I/O on 19 May with Gemini 3.5 Flash, a new Antigravity CLI, an SDK and a reset of the Google AI subscription — a clear signal that the agent-first idea is being taken seriously. We covered the launch in detail in our piece on Antigravity 2.0 and Google's agent-first developer platform.
What just changed: a turbulent ninety days
If your impression of these tools is more than a quarter old, it is out of date. The recent run of releases is unusually dense:
- Cursor shipped Composer 2.5.
- Anthropic doubled Claude Code limits — off the back of a SpaceX compute deal — and then released Claude Opus 4.8.
- GitHub switched on usage-based "flex" billing on 1 June, alongside a new Copilot Max plan.
- Windsurf bundled Devin and then rebranded to Devin Desktop.
- OpenAI released GPT-5.5 and expanded Codex with Sites, Annotations and business plugins.
- Kiro launched a simplified credit-based plan with parallel Spec task execution.
- Google launched Antigravity 2.0 with the CLI, SDK and subscription reset noted above.
The pattern underneath all of this is consolidation of capability and divergence of business model. The tools are converging on what an agent can do — read a repo, plan, edit, test — while diverging sharply on how they charge you and how tightly they hold your workflow.
Do not benchmark these tools on a toy task. Give two finalists the same real ticket from your backlog — ideally one that touches three or four files and needs a test written — and compare the diffs, the rework and the token bill. A fifteen-minute bake-off on real work tells you more than any leaderboard.
The comparison at a glance
The table below maps each contender to its vendor, interface style, pricing model and the kind of work it suits best. Treat "best for" as a starting hypothesis to test against your own codebase, not a ruling.
| Tool | Vendor | Interface style | Pricing model | Best for |
|---|---|---|---|---|
| Claude Code | Anthropic | Terminal-native agent | Subscription tiers; limits recently doubled | Project-level autonomy in the shell |
| OpenAI Codex | OpenAI | Terminal-native (plus Sites, plugins) | Subscription with usage tiers | Teams already on the OpenAI stack |
| Cursor | Cursor | IDE-embedded | Subscription | Editor-first developers wanting deep completion |
| GitHub Copilot | GitHub / Microsoft | IDE-embedded | Usage-based "flex" billing; Copilot Max plan | Teams living inside GitHub and VS Code |
| Windsurf (Devin Desktop) | Windsurf | IDE-embedded, Devin-bundled | Subscription | Builders wanting an autonomous teammate in-editor |
| Kiro | Amazon | Spec-driven IDE | Credit-based plan | Large, planned features built from written specs |
| Antigravity | Agent-first IDE / platform | Google AI subscription (recently reset) | Greenfield builds where the agent drives |
One data point is worth holding in mind while you read the table. A Pragmatic Engineer survey of 906 software engineers in February 2026 found Claude Code the most-used AI coding tool, with a 46% "most loved" rating. That tells you where the centre of gravity sits — but popularity is a signal about the average engineer, not a verdict about your particular workflow.
"Best for" is not a moat. Cursor shipping Composer 2.5, Codex gaining Sites and plugins, and Antigravity arriving with a CLI all show that categories blur fast. Choose for the next two quarters, and re-evaluate rather than marry one tool for life.
Pricing: the quiet shift to usage-based billing
The single most important change for your budget is not a feature — it is the business model. GitHub switching on usage-based "flex" billing on 1 June, with a new Copilot Max plan, is the clearest example, and we unpacked the mechanics in our analysis of GitHub Copilot's move to usage-based billing. But the direction is industry-wide: Kiro's credit-based plan, Google's subscription reset and the various OpenAI and Anthropic tiers all push the same way.
Usage-based billing rewards light, surgical use and punishes sprawling, exploratory agent runs. An agent that reads your whole repository on every task is wonderful for capability and unforgiving on a per-token plan. For a small team in Pune or Manchester running dozens of agentic tasks a day, the difference between a flat subscription and metered usage can be the difference between a predictable line item and a nasty month-end surprise.
Before rolling a tool out to a team, run a one-week pilot with billing alerts switched on. Capture cost-per-merged-PR, not cost-per-token — that is the number your finance lead will actually care about, and it lets you compare a metered tool against a flat-rate one honestly.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →Vendor lock-in: the decision you make without noticing
The third axis is the one builders most often overlook until it is expensive. Ask a blunt question of any tool: if you switched it off tomorrow, what would you lose? With terminal-native agents and spec-driven tools, much of the value lives as plain files in your repository — prompts, specifications, configuration and task lists that travel with the codebase. With fully managed, agent-first platforms, more of the value lives inside the vendor's environment.
Neither is wrong. An agent-first platform like Antigravity can deliver a smoother, more cohesive experience precisely because it owns the whole loop. But it is worth being clear-eyed about the trade. The deeper a tool's understanding of your project lives inside a proprietary IDE or cloud, the more switching later costs you — in retraining, in re-tooling and in the institutional memory the agent accumulated. Spec-driven Kiro is interesting here exactly because the specs it produces are durable artefacts you keep regardless of which agent reads them next.
Avoid standardising your whole team on a single platform purely because the demo dazzled. If your agent's context, history and configuration cannot be exported as files you control, you are buying convenience now and paying for it at migration time.
A framework for choosing — not a winner
Put the three axes together and the decision becomes a short sequence of questions rather than a popularity contest. Work through them in order.
- Where do you already work? If you live in the terminal, a terminal-native agent such as Claude Code or the Codex CLI will feel native. If your team is wedded to its editor, an IDE-embedded tool such as Cursor, Copilot or Windsurf removes friction. If you are starting something genuinely new, an agent-first platform such as Antigravity is worth a serious look.
- How planned is your work? Large features with clear requirements suit a spec-driven tool like Kiro, which now executes Spec tasks in parallel. Exploratory or reactive work suits a conversational agent.
- What can your budget absorb? If your usage is spiky and exploratory, a flat subscription protects you. If it is light and surgical, a usage-based plan may be cheaper. Decide before you scale, not after the first invoice.
- How much lock-in can you tolerate? If portability matters — and for most teams it should — favour tools whose artefacts are plain files in the repo.
- Run the bake-off. Shortlist two, give them the same real ticket, and let the diffs decide.
For teams whose work spans both autonomous agents and conventional engineering, it is also worth understanding how these coding tools relate to the broader agent landscape. Our comparison of Claude, Codex and Gemini computer-use agents covers the wider capability question, and if you are wiring agents into a larger system, our 2026 map of agent frameworks and SDKs is the natural next read. If spec-driven development is your direction, our deep dive on Amazon's Kiro agentic IDE is worth your time.
The honest conclusion
There is no single right answer in June 2026, and anyone selling you one is selling. Claude Code's lead in adoption is real and well earned, but the most-loved tool on average is not automatically the best tool for your repository, your budget or your appetite for lock-in. The field has matured to the point where the smart move is to choose deliberately along three axes — philosophy, pricing and portability — pilot two finalists on real work, and stay willing to switch when the next ninety days inevitably reshuffle the deck again.
If you ship with any of these tools, that work is exactly the kind of evidence the people hiring across India and the UK are looking for. A Verified Builder profile is a simple, understated way to make it findable.