The Battle for AI-Powered Development: Amazon Kiro vs. Anthropic Claude Code

Two opposite bets on AI coding, and what a year of competition did to both

Jul 15, 2025

When I first wrote this comparison in the summer of 2025, the question looked clean. Amazon's Kiro and Anthropic's Claude Code were two newcomers making opposite bets on how AI would reshape software, and you could pick a side based on which bet you believed.

A year later, the cleaner answer is that both bets were partly right, and the two tools moved toward each other. Claude Code grew the structure Kiro was built around. Kiro shipped, went generally available, and repriced in a way that made some of its early fans flinch. The "process versus flexibility" framing that made the original piece easy to write has narrowed. So this is the same head-to-head, brought current: what each tool actually is now, what it costs, and how to choose between them in a field that is no longer two newcomers but a crowded market.

The honest update: the gap I described in 2025 closed from both ends. Claude Code added the scaffolding, Kiro grew up and got pricier, and the choice is now about defaults rather than philosophy.

Meet the contenders, a year on

Amazon Kiro: from preview to general availability

When this post first ran, Kiro had just launched in public preview. As of November 2025 it reached general availability. Built on VS Code OSS and still powered by Anthropic's Claude through Amazon Bedrock, Kiro's pitch hasn't changed: turn fast, messy AI prototyping into documented, production-ready software through an enforced process.

"From vibe coding to viable code" stayed Amazon's tagline through the GA launch.

The core loop is the same. Ask Kiro to add a feature and it refuses to dive straight into code. It generates formal requirements, drafts a technical design with data-flow diagrams and interfaces, breaks the work into sequenced tasks with dependencies, then implements with testing and quality checks built in. The bet is that enforcing process upfront prevents the technical debt that fast AI-assisted coding tends to leave behind.

What's new since 2025: Kiro added an "Auto" agent mode for users who want less ceremony, a quiet admission that the rigid spec-first flow isn't right for every task. And Amazon now runs a deliberate dual track, keeping Q Developer alongside Kiro rather than folding one into the other.

Anthropic Claude Code: the terminal tool that grew a skeleton

Claude Code started as the opposite of Kiro: maximum capability, minimal ceremony. A CLI-first tool that slots into your existing workflow instead of replacing it, strong at autonomous codebase exploration, multi-file edits, and deep reasoning.

That description is still true, and it's no longer the whole story. The biggest change in a year is that Claude Code stopped being structure-free. It now ships Skills (reusable, packaged capabilities), native subagents, hooks that fire on defined events, an explicit plan mode, and dynamic workflows that can fan out parallel work and verify it. None of that existed when I wrote the original.

That matters for this comparison more than any pricing line. The original piece floated a "convergence question," wondering aloud whether Claude Code might someday add more structure. It did. The terminal-native tool grew a skeleton you can build process on, which is exactly the territory Kiro staked out first.

Technical showdown

The model difference

Both tools still run on Anthropic's Claude family, so Anthropic keeps collecting from both sides of the fight. The model lineup has moved on:

The Claude family is now Opus 4.8 / Sonnet 4.6 / Haiku 4.5 / Fable 5. Claude Code runs on Opus 4.8, Anthropic's strongest model. Kiro draws on Claude through Amazon Bedrock. The shape of the difference holds from the original: Claude Code reaches for the top-tier reasoning model, which shows up most on complex, multi-step work. Kiro trades some of that ceiling for a managed, predictable flow.

The 2025 comparison table, updated for where both tools stand now:

| Feature | Amazon Kiro | Anthropic Claude Code |

|---|---|---|

| AI model | Claude via Amazon Bedrock | Claude Opus 4.8 |

| Interface | Full IDE (VS Code OSS) | CLI + IDE plugins |

| Workflow | Structured (spec → design → tasks), plus a new Auto mode | Flexible, now with Skills, subagents, hooks, plan mode |

| Context approach | Guided through specs | Autonomous exploration, now packageable via Skills |

| Target user | Teams wanting enforced process | Power users who build their own structure |

| Pricing | Credit model ($20 to $200/mo, plus overage) | Subscription ($20 Pro, $100 to $200 Max) |

| Status | Generally available (Nov 2025) | Generally available |

The model row is the one constant: Claude Code on Opus 4.8, Kiro on current Sonnet via Bedrock. The interface and workflow rows are where the two tools have crept toward each other.

Two ways of reading a codebase

Both tools sit on Claude's large context window, and they still use it differently.

Kiro reads with a goal. It analyzes code against the requirements and design docs it generated, a focused approach that's strong when the spec is clear and weaker when the work doesn't fit a template.

Claude Code reads to map. It explores project structure and dependencies on its own, building understanding without you feeding it context by hand. With Skills and subagents now in the picture, that exploration can be packaged and repeated, which narrows the gap on the "but Kiro is repeatable" argument.

Development experience: the philosophies still show

Kiro gives you a full IDE on VS Code OSS, with familiar keybindings, extensions, visual diffs, one-click approvals, task tracking, and multimodal input for things like turning a screenshot into UI. It's for developers who want everything in one environment.

Claude Code lives in your terminal with optional IDE plugins, scriptable for automation and CI/CD, permission-gated for safety, and composable with standard Unix tools. The audience is developers who want to enhance their setup, not adopt a new one. That hasn't changed, even as Claude Code added more structure underneath.

Performance in the real world

Benchmarks, with a caveat

The original cited roughly a 70% success rate on SWE-Bench for Claude Code. That was a 2025 figure, and benchmark numbers in this space age fast as new model versions ship, so treat it as a marker of the era rather than a current scoreboard. The directional claim holds better than the digits: Claude Code's access to the heavier model gives it an edge on complex, multi-step bug-fixing, and that edge is what shows up when the task gets gnarly.

Kiro, no longer in preview, has more real-world mileage behind it than the "early promise, limited data" line from a year ago. Its strength reads consistently: well-documented, pattern-following code, delivered fast inside its structured flow, with more friction on unfamiliar paradigms or legacy systems.

The quality personalities

The character sketches from the original still ring true with another year of use behind them.

"Claude Code is like working with a careful surgeon, it asks the right questions before cutting anything open." (community feedback synthesis)

Claude Code produces careful, maintainable code and sometimes over-engineers, needing a human to trim it back. It catches edge cases and architectural concerns and takes its time on hard problems.

"Kiro is the reliable contractor, shows up on time, does competent work, rarely gets it wrong, sticks to established patterns." (early-adopter assessment)

Kiro delivers consistent, documented work along known patterns and moves quickly within its lane.

The economics, where the year was loudest

Kiro repriced, and people noticed

This is the section that changed the most. Kiro launched in 2025 on simple interaction tiers: a free preview, then $19 and $39 a month. The pricing turned contentious fast. When Amazon first revised it in August 2025, The Register reported the new scheme being dubbed wallet-wrecking. By general availability that November, Kiro had settled into a credit model:

Free: 50 credits
Pro: $20/month, 1,000 credits
Pro+: $40/month, 2,000 credits
Power: $200/month, 10,000 credits
Overage: $0.04 per credit

The worry I flagged as a question in the original, that heavy automation could burn through quota faster than expected, is exactly the friction a credit meter creates. The lesson for buyers: with a credit model, your cost tracks how hard you drive the tool, not a flat tier, so model your real usage before committing.

Claude Code's subscription reality

Claude Code pricing rides Anthropic's broader plans. Pro is $20/month; the Max tiers run $100 and $200 a month for heavier use, with effectively generous-to-unlimited usage for most workflows. Team and API pricing offers flexibility but needs cost discipline.

Kiro went from predictable tiers to a credit meter that bills you for how hard you push it. Claude Code stayed on flat subscription tiers. If budget predictability is the deciding factor, that flipped during the year.

Security and enterprise readiness

Both tools carry enterprise-grade backing, and not much here needs softening.

Claude Code maintains security controls built in from the start: read-only by default with explicit permission gates, detailed audit logging, and enterprise deployment through AWS Bedrock and Google Vertex.

Kiro inherits AWS's security infrastructure, with its long list of compliance certifications, trusted-command controls that block dangerous operations, and the enterprise features you'd expect from an Amazon product now that it's past preview.

Choosing your tool, not your philosophy

The original framed this as picking a development philosophy. With the two tools closer together, it's more useful to frame it as picking a default.

Kiro fits when

A team is fighting the maintainability of AI-generated code and wants documentation and tests enforced, not optional.
The work lives in modern web/cloud stacks where its patterns are strong.
A single, complete IDE environment beats a composable toolkit for your team.
You've modeled your usage against the credit pricing and it pencils out.

Claude Code fits when

You want maximum capability and the heavier reasoning model on hard, multi-step work.
You work across diverse languages and want the tool to adapt to you, not the reverse.
You're comfortable in the terminal and want to script and automate around it.
You want to add your own structure now (Skills, subagents, hooks) rather than accept a fixed one.

The hybrid, still real

Plenty of developers run both: Claude Code for daily coding, Kiro for structured feature work. That hasn't gone away, though it's a smaller niche than it was, because Claude Code can now cover more of the structured ground itself.

The wider field has filled in

A year ago I wrote that Microsoft and Google were "preparing" their offensives. That framing is retired. The market is crowded and established now, not two newcomers with giants warming up offstage.

The current field around Kiro and Claude Code includes Cursor ($20 Pro / $200 Ultra), Windsurf ($20 / $200), GitHub Copilot ($10 Pro / $39 Pro+), Devin ($20 / $500), Replit, plus OpenAI's Codex and Google's Antigravity. The competitive question is no longer whether the incumbents will respond. They have, and so has a wave of independents.

The total opportunity still looks large. The original cited a projection of the AI coding market reaching $97.9 billion by 2030, again a 2025 figure that's worth re-checking against newer estimates rather than quoting as today's number. The point survives the digits: the market is big enough to hold structured tools, flexible tools, and a dozen hybrids of the two.

The bottom line

A year on, the cleanest read is that nobody fully won the philosophy argument because the philosophies converged. Claude Code, the flexible terminal tool, grew Skills, subagents, hooks, plan mode, and workflows, the structure Kiro built its whole pitch around. Kiro, the structured IDE, went generally available, added an Auto mode for less ceremony, and moved to a credit model that cost it some early goodwill.

So the choice is no longer "do you want AI to impose structure or stay out of your way." Both can now do both. The real questions are narrower and more practical: which default do you want to start from, which pricing model fits how hard you'll push the tool, and whether you'd rather add your own structure (Claude Code) or accept a strong one out of the box (Kiro).

If you're building your own setup around either, I've put my full Claude Code harness in the open as a cookbook, claudelicious, and the anti-slop guardrails I run on AI output in slopless. For the practical setup side, see the companion piece, Claude Code Best Practices.

The transformation was never about choosing sides. It's that the side you choose matters less than it did, and the field around the choice got a lot more interesting.

Run Data Run

Discussion about this post

Ready for more?