What changed on this site

Two new pages. Memory covers the feature that reached Kiro CLI 2.29.0 (local V3 sessions) and Kiro IDE 1.4 this month: Kiro learns a repository’s conventions and corrections in one session and recalls them in the next. It is the fourteenth idea on Core Concepts, and the page explains why rules that must always apply still belong in steering. Turn on the command sandbox is a recipe for the operating-system sandbox that IDE 1.3 and CLI 2.31.0 put around agent-run shell commands.

The change most likely to affect you is Kiro CLI 3.0, which Kiro says makes V3 the default by the end of October and retires inline shell autocomplete, the CLI desktop app, and Classic sessions. The CLI Guide now has a checklist for the switch: upgrading agents, checking permissions, replacing use_aws, and moving ACP clients. It also records CLI 2.28.0 through 2.32.0.

Models: Claude Sonnet 4.0 reached end of life on 14 October and left the tables; Kimi K3 (2.4x, US only) joined them; and Kiro’s model documentation now lists GPT-5.6 Luna at 0.6x (1.2x above 272K tokens), down from 1.1x, with no changelog entry. The model table and the credit calculator use the new numbers.

Kiro Crew gained sections for 0.8.0 (27 September, which the 4 October pass missed) and 0.9.0 (16 October). 0.8.0 matters for safety: owner-written MCP autoApprove lists now take effect and crew logs are on by default. What’s New in Kiro, History, the IDE, CLI, and Web concept pages, Hooks (tool categories), Steering, and the glossary were updated to match. Sources are in research/kiro-facts-2026-10-21.md.

The headline is Workflows (30 September): reusable, multi-step plans in which each step runs as its own agent session in the background while you keep working in the main chat. They shipped in IDE 1.2, CLI 2.26.0, and Kiro Web on one shared runtime, opt-in everywhere. They have a new Workflows concept page and a recipe, Write your first workflow recipe, that builds an implement, review, and loop recipe step by step. The glossary keeps them apart from Kiro Crew’s Python Workflows.

Two more new pages. The credit calculator turns your daily usage and model choice into monthly credits and the cheapest plan; it and the model table in Choosing a Model now read from one data file. What’s New in Kiro lists the last month of releases in plain English, with what to do about each one.

Kiro Crew 0.7.0 (24 September) is recorded on the Kiro Crew page: settings that apply live, work that survives Gateway restarts, OpenCode and goose backends, Fargate remote crews, and pull request watches beyond GitHub. 0.7.2 fixed WhatsApp group access.

The change most likely to break something is in CLI 2.24.0 (23 September): project .env files no longer load automatically into sessions, MCP servers, or tools. Export the variables in your shell first. The CLI Guide covers it alongside 2.25.0 (/powers install, the SessionEnd hook) and 2.27.0 (saved prompts as slash commands, folder and line-range references in steering).

On models, Claude Opus 5.5 (28 September, 2.0x) and Claude Sonnet 5.5 (2 October, 1.3x) joined the tables in Choosing a Model and Credits & Pricing. Opus 5.5 breaks the old rule that every Opus costs 2.2x. Kiro’s effort documentation now lists xhigh for Opus 4.7, so that claim was corrected, and Kiro Web gained a reasoning effort selector on 22 September.

Earlier, the 25 September pass recorded CLI 2.22.0 (/fullscreen) and 2.23.0 (effort controls inside /model), the Enterprise-only Claude Fable 5.1 preview, and the Kiro for Students programme. Sources are in research/kiro-facts-2026-09-24.md and research/kiro-facts-2026-10-04.md.

The consequential change is pricing. On 14 September the GPT-5.6 tiers moved from a 272K context window to 1M, and the family was re-priced onto the two-tier scheme OpenAI and Amazon Bedrock use: Sol went from 2.4x to 4.4x, Terra from 1.0x to 2.2x, and Luna from 0.1x to 1.1x, with every one of those rates doubling on requests above 272K tokens. Luna had been the cheapest non-open-weight model in Kiro since July; it is now dearer than Auto. Choosing a Model and Credits & Pricing were rewritten around the new numbers, including the worked cost example and the cheapest-to-dearest spread, which is now 88x rather than 48x.

Three release updates landed alongside it. IDE 1.1 (14 September) introduced Agent Artifacts — durable documents, diagrams, code, and images you review in Agent Focus and return to later, rather than losing them in a chat transcript — plus native ARM64 builds for Windows and Linux and a Code OSS 1.131 foundation. CLI 2.21.4 (11 September) added a configurable /sessions search scope and a --v2 flag for running one session on the older agent harness without changing your saved default. Crew 0.6.0 (5 September) added selectable agent harnesses, remote crews, worker dispatch, an Apps Launchpad, and much longer timeouts, and raised the minimum Python version to 3.12.

Sources for this pass are recorded in research/kiro-facts-2026-09-17.md.

The biggest restructure so far: each search intent now has a page of its own instead of sharing a broad chapter with three other topics.

Under the hood, every page carries structured data (TechArticle, BreadcrumbList, and FAQPage where a chapter has an FAQ), a generated social preview image, and a lastmod date in the sitemap. This changelog and its RSS feed are new.

Every chapter that had not been touched earlier in August was re-verified against the official documentation, and the at-a-glance diagram was regenerated. Kiro CLI 2.20.0 and IDE 1.0.395 were recorded at the end of the month.

Four chapters joined the guide, the diagrams were regenerated, and the site now remembers which chapters you have read.

Every chapter was re-read against the July 2026 state of Kiro: IDE 1.0, Sonnet 5, CLI V3, and agents defined in Markdown.

The Workspace Game replaced the old “try it yourself” sections. Canonical URLs were normalised, the 404 page was marked noindex, and the sitemap and analytics were wired up.

The site launched with the complete learning path: what Kiro is, its history, core concepts, a beginner tutorial, an advanced tutorial, use cases, a cheat sheet, and the practice pages.

Before launch every open fact was verified against kiro.dev and the inline TODO markers were removed. Learning diagrams were generated for the concept pages.