I tried out the newly supported OpenAI GPT-5.6 Sol / Terra / Luna on Kiro
This page has been translated by machine translation. View original
Introduction
On July 14, 2026, Kiro added the ability to select OpenAI's GPT-5.6 Sol / Terra / Luna.
GPT-5.6 also became GA on Amazon Bedrock today, and Kiro supports it on the same day. Details on the Bedrock side are summarized in the following article.
Here is a comparison between the models previously available in Kiro and the models added this time.
| Category | Model Family |
|---|---|
| Previously available (Claude + open weights) | Claude Opus / Sonnet / Haiku, DeepSeek / MiniMax / GLM / Qwen |
| Added this time (+ OpenAI) | GPT-5.6 Sol / Terra / Luna in addition to the above |
GPT-5.6 offers three tiers — Sol, Terra, and Luna — to choose from based on difficulty and frequency of use.
| Tier | Use Case | Credit Multiplier | Context | Benchmark (CAI) |
|---|---|---|---|---|
| Sol | Most challenging multi-step tasks | 2.4x | 272K | 80 |
| Terra | Everyday multi-step tasks | 1.2x | 272K | 77.4 |
| Luna | High-frequency, throughput-focused | 0.6x | 272K | 74.6 |
CAI (Coding Agent Index) is a benchmark metric published in Kiro's model documentation.
Kiro's credit multiplier and Bedrock's USD pay-as-you-go pricing are separate systems.
| Status | Region | Target Plans |
|---|---|---|
| Experimental | us-east-1, eu-central-1 (cross-region inference) | Pro / Pro+ / Pro Max / Power |
Kiro's official model documentation describes the following characteristics of GPT-5.6's behavior style.
- Hidden chain-of-thought (internal reasoning is not displayed; only the final output is returned)
- Tool coordination via lightweight program generation (generates intermediate programs as needed to process results and reduce round-trips for tool integration)
- Transparent usage (Bedrock requires the Responses API, but Kiro only requires specifying the model name)
Verification
This time, we specified GPT-5.6 Sol in a headless execution of the Kiro CLI and confirmed that it works correctly. A comparative verification of all tiers including Terra and Luna is out of scope.
Model Selection in IDE and CLI
The three tiers of GPT-5.6 are also displayed on the IDE's model selection screen.
We also confirmed that the three tiers of GPT-5.6 can be selected from the /model command in the CLI.
Running GPT-5.6 Sol in Headless Mode
We specified gpt-5.6-sol in --model and asked for an overview of Amazon Bedrock.
kiro-cli chat --no-interactive --model gpt-5.6-sol "Amazon Bedrock とは何ですか？3文で簡潔に説明してください"
The following response was returned.
Amazon Bedrockは、AWSが提供する生成AIアプリケーション開発向けのフルマネージドサービスです。複数の主要な基盤モデルを統一
APIで利用でき、用途に応じたモデルの評価・カスタマイズ・検索拡張生成（RAG）が可能です。AWSのセキュリティや各種サービスと連携
し、生成AIシステムを効率的に構築・運用できます。
The footer displayed credit consumption and execution time.
▸ Credits: 0.10 • Time: 3s
Summary
Kiro now allows you to choose between both Claude and OpenAI, creating an environment where you can use different models depending on the nature of the task. Personally, I used to go through Codex CLI when I wanted to get a second opinion from GPT-5.x, so it's great that this can now be done entirely through Kiro's model switching. The status is Experimental (a stage where changes may occur in the future), but I'd like to start using it on actual tasks to understand the differences in output tendencies compared to Claude and the characteristics of each GPT-5.6 tier.
