The IDE Landscape Shifted in 2026

The agentic coding market looked very different eighteen months ago. Cursor held the IDE throne, GitHub Copilot anchored Microsoft’s developer ecosystem, and “agentic” mostly meant “the tab bar autocompletes a bit further.” Then Claude Code arrived and reset what developers expected from a coding agent. Two of the largest cloud providers responded by replacing their existing tools entirely. Amazon quietly retired Amazon Q Developer and relaunched as Kiro in mid-2025—a spec-driven IDE with enterprise audit trails. Google followed at I/O 2026 on May 19, unveiling Antigravity 2.0, built on Gemini 3.5 Flash and designed for parallelized agent throughput. By September 2026, developers choosing a primary coding environment face three technically distinct options, each optimized for a different workflow. This comparison examines how they actually differ—on architecture, benchmarks, pricing, and the real-world tradeoffs nobody mentions in launch posts.

AWS Kiro: Planning as a First-Class Feature

Kiro’s core premise is that the biggest cost in software development isn’t writing code—it’s writing the wrong code. The tool enforces a three-phase workflow before a single line is generated: a requirements document in EARS notation, a technical design document, and a dependency-ordered task breakdown. Only after those three artifacts exist does the agent begin implementation.

How the Spec Workflow Works in Practice

EARS (Easy Approach to Requirements Syntax) is an ISO-adjacent notation method that structures user stories with a mandatory condition-action-response format. Kiro auto-generates this from a feature description, then waits for developer sign-off before proceeding. The resulting task list decomposes work into steps estimated at 5–15 minutes each, with explicit dependency ordering so parallel execution is safe. For teams running regulated software—think healthcare tooling or financial services infrastructure—this produces an audit trail from intent to implementation that Kiro’s competitors cannot currently match. Kiro has HIPAA certification and is pursuing FedRAMP High authorization.

The trade-off is ceremony. For a small bug fix or a quick prototype, three mandatory planning documents feel like architecture review for a parking lot refactoring. Developers in unstructured environments report that Kiro’s upfront discipline slows them relative to tools that start generating immediately. The approach also assumes specification documents stay synchronized with code as the project evolves—an assumption that holds at first commit and degrades steadily over weeks unless teams actively maintain the specs.

Pricing and Enterprise Controls

Kiro uses a credit-based pricing model: 50 credits free, 1,000 credits at $20/month (Pro), 2,000 credits at $40/month (Pro+), and 10,000 credits at $200/month (Power). Overages run $0.04 per credit. The predictability is intentional—enterprise procurement teams can budget against credit consumption rather than guessing at token-level API costs. Kiro also ships native SAML/SCIM support with confirmed integrations for Okta and Entra ID, and centralized billing across teams. These are table-stakes for enterprise security reviews, and no other tool in this comparison offers them out of the box today.

On security transparency: Kiro has accumulated three formal CVEs plus one inherited vulnerability, all patched and all visible in standard CVE databases through AWS security bulletins. This means automated scanning tools catch Kiro exposure automatically. That’s actually a signal of mature disclosure process, not negligence.

Google Antigravity 2.0: Parallel Agents, Browser Testing, Speed

Antigravity launched at Google I/O 2026 as a replacement for Gemini Code Assist—and the rebranding understates how much changed. Where Gemini Code Assist was an IDE plugin, Antigravity is a full agent platform: a VS Code–based desktop application, a CLI, and a public SDK. Its default model is Gemini 3.5 Flash, which hits roughly 289 output tokens per second on independent benchmarks—approximately four times the throughput of Claude Opus 4.7 or GPT-5.5 at comparable task sizes. For tasks where latency matters more than raw accuracy, that difference is tangible.

The Browser Subagent Advantage

Antigravity’s most distinctive architectural feature is the Browser Subagent: an embedded Chromium instance that agents can control directly. For full-stack development, this closes a loop that other tools leave open. Rather than generating a UI component, stopping, and waiting for a developer to manually verify rendering, Antigravity can spin up a browser subagent to load the result, check for console errors, and validate layout—all within the same agent session. This “verify while building” approach fundamentally changes how UI-heavy tasks flow compared to Kiro’s sequential spec-then-implement model or Claude Code’s terminal-native approach where browser verification requires external tooling.

Antigravity also supports up to five parallel Cascade agents running simultaneously, enabling true multi-file concurrent development. The platform scores 76.2% on SWE-bench—meaningfully above human baseline but below Claude Code’s frontier numbers. For many production tasks, the throughput advantage at 289 tokens/second compensates for lower per-task accuracy: faster iteration beats slower perfection in prototyping cycles.

What “Free Preview” Costs You in Stability

As of September 2026, Antigravity is still in public preview with a free tier. The planned pricing is $19.99/month (AI Pro) and $99.99/month (AI Ultra with 5x limits)—but “planned” is doing real work in that sentence. Neither an enterprise SLA nor formal compliance certifications are documented. Rate limiting can extend up to five days in current preview usage, which is workable for individual evaluation but unacceptable for any team that depends on coding throughput. Antigravity also has no SSO integration and no audit logging—neither are planned for the initial GA release.

Security disclosure is the sharpest contrast with Kiro. Antigravity has zero formal CVEs despite four documented vulnerabilities discovered by security researchers. None appear in standard CVE feeds. The most severe involved a backdoored workspace that persisted across all subsequent sessions, requiring a full reinstall to remediate. That an issue of this severity bypasses automated CVE scanning is a gap security teams should flag explicitly before enterprise adoption.

Claude Code: Benchmark Leader, Higher Setup Bar

Claude Code is the performance benchmark against which the others are measured. Running on Claude Opus 4.8, it scores 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro—the highest published scores for any coding agent as of September 2026. Claude Fable 5 posts 80.3% on SWE-bench Pro. These numbers matter: SWE-bench Verified correlates meaningfully with real-world bug-fix accuracy on open-source codebases, and the gap between 88.6% (Claude Code) and 76.2% (Antigravity) is large enough to show up in practice on complex multi-file tasks.

Claude Code generated roughly 135,000 GitHub commits per day in mid-2026—approximately 4% of all public GitHub commits. It runs as a terminal-native CLI with optional IDE plugins, making it composable with CI/CD pipelines, headless scheduled runs, and MCP-based tool integrations in a way that GUI-first IDEs cannot easily replicate. Parallel sub-agents and dynamic workflows (which split large tasks across multiple agents and resume progress on codebases up to ~750,000 lines) are first-class features.

Token Consumption Is the Hidden Cost

Claude Code’s benchmark advantage comes with a real cost: it uses approximately four times more tokens than comparable tools for equivalent tasks—6.2 million tokens versus 1.5 million in one documented head-to-head comparison. At the $20/month tier, developers report hitting limits within 10–20 productive weekly sessions, far fewer than most assume when signing up. Moving to the $100/month tier helps but doesn’t resolve the ceiling; the $200/month tier is what heavy users actually require. Pricing at those levels is comparable to Antigravity’s Ultra tier, eliminating the apparent cost advantage.

Claude Code also demands meaningful setup investment before it performs well. CLAUDE.md configuration files, skills definition, and MCP integration are not optional niceties—without them, the agent lacks the project context to make correct architectural decisions. One developer benchmarking comparison estimated that most engineering time in the first week went to Claude Code configuration rather than actual coding. For teams willing to invest that setup time, the performance ceiling is the highest available. For teams wanting immediate productivity, the friction is real.

Head-to-Head: The Comparison Matrix

Who Should Use What

The right answer depends more on your workflow shape than your benchmark preferences.

Choose Kiro if your team operates in a regulated industry, needs HIPAA or FedRAMP compliance, already runs on AWS, or has a culture of upfront specification discipline. Kiro’s audit trail and enterprise controls are the most mature in this group. The SWE-bench gap relative to Claude Code matters less when the risk of shipping the wrong feature is higher than the risk of slightly imperfect code. Brownfield codebases with established architectural constraints also benefit from Kiro’s approach—the spec documents enforce consistency rather than leaving architectural decisions to each agent session.

Choose Antigravity if you’re a greenfield team that values rapid iteration over audit trails, already runs on Google Cloud, or simply wants to evaluate without committing budget. The free preview tier makes Antigravity the lowest-risk evaluation option available right now. The Browser Subagent is genuinely differentiated for full-stack teams—if your workflow includes frequent UI verification loops, no other tool handles that end-to-end. The enterprise-readiness gap is the real constraint: if you need SSO, compliance, or guaranteed uptime, Antigravity is not production-ready for enterprise deployment today.

Choose Claude Code if you want the highest benchmark performance, need deep MCP ecosystem integration, or are building autonomous pipelines—scheduled tasks, CI/CD integration, headless agent runs. It rewards teams willing to invest in setup and teams whose bottleneck is code quality rather than speed. The token consumption and limit economics mean honest budget modeling matters: plan for the $100–$200/month tier if you’re using agents heavily, and treat the $20 tier as trial access rather than sustainable production capacity. You can also see how it performs on legacy codebases specifically if that’s your primary concern.

None of these tools is the final answer. All three reset coordination context at session boundaries, meaning accumulated architectural intent degrades over long-lived projects without explicit spec maintenance (Kiro) or CLAUDE.md discipline (Claude Code). The common failure mode in every tool reviewed here is what happens six months into a project when the original intent is gone and the agent only has the code to read. That problem remains unsolved across the board—and it’s the benchmark the next wave of tools will need to crack.

For teams evaluating now, the most rational sequence is to run all three in parallel on the same real codebase task, measure actual time-to-PR, and compare against your compliance requirements. Antigravity’s free preview window makes that evaluation essentially free.

Further Reading

• Claude Code vs Cursor vs Augment Code: August 2026 — The previous head-to-head that benchmarked Claude Code against Cursor in August, useful for understanding how Claude Code has evolved.
• Kiro vs Antigravity (Augment Code) — The most detailed security comparison available, specifically covering CVE disclosure differences and enterprise readiness gaps.
• SWE-bench Pro Leaderboard (September 2026) — The live benchmark tracker for coding agent accuracy, updated monthly with new scores and pricing data per point.