Kiro now supports the full GPT-5.6 family—Sol, Terra, and Luna—across its IDE, CLI, and web experience. That gives developers three levels of capability and credit consumption inside the same spec-driven workflow.

The important distinction is that this is Kiro model access, not a new OpenAI API price. Kiro meters usage with its own credit multipliers and exposes a 272K-token context window for these models. Direct OpenAI API pricing and availability through other providers are separate decisions.

Quick comparison

Model in Kiro	Current credit multiplier	Best fit	Main trade-off
GPT-5.6 Sol	2.4x	Hard architecture, ambiguous refactors, highest-stakes implementation	Highest Kiro credit use
GPT-5.6 Terra	1.0x	Daily feature work, specs, debugging, multi-file changes	Less headroom than Sol on the hardest tasks
GPT-5.6 Luna	0.1x	Routine edits, tests, documentation, repetitive implementation	Use a stronger model when judgment dominates volume

These are Kiro’s current published multipliers as of August 25, 2026. They can change independently of OpenAI’s token prices. Kiro labels the GPT-5.6 family experimental and currently serves it from the United States regardless of the region selected in a Kiro profile, which matters for teams with data-location requirements.

Why GPT-5.6 fits Kiro’s workflow

Kiro is built around more than a chat box. Its spec workflow turns an idea into requirements, design, and implementation tasks; steering files carry team standards; checkpoints make changes reviewable; and property-based tests can turn requirements into executable quality checks.

That structure gives the model better context than a one-line coding prompt. It also makes model choice more deliberate: the most capable model does not need to handle every step.

A practical flow is:

1. Use Terra to turn a feature request into requirements and a design.
2. Escalate only the genuinely difficult architecture or migration decision to Sol.
3. Use Luna for repetitive task execution, test scaffolding, documentation, and low-risk cleanup.
4. Run tests and inspect the checkpoint before accepting the implementation.

This is model routing at the development-workflow level. It can control Kiro credit consumption without abandoning the shared spec.

When to choose Sol

Choose Sol when a wrong decision will be expensive to unwind:

• redesigning a service boundary;
• planning a large cross-repository migration;
• debugging a failure with several plausible root causes;
• reviewing security-sensitive authentication or authorization changes;
• resolving conflicting requirements across a long specification.

Sol’s 2.4x multiplier makes it a poor default for mechanical tasks. Use it for the reasoning bottleneck, then hand the approved plan back to Terra or Luna.

When Terra is the default

Terra is the balanced choice and the cleanest default for most professional work in Kiro. At a 1.0x multiplier, it fits:

• generating and refining feature specs;
• implementing multi-file changes;
• fixing bugs that require repository context;
• writing tests alongside production code;
• reviewing a task list before execution.

OpenAI also reports that Terra completed successful Terminal-Bench 2.1 tasks in Kiro at roughly an 82% cost reduction. Treat that as a vendor-reported Kiro result, not a universal promise about your repository or a direct API price comparison. Task mix, prompts, retries, and Kiro’s own metering all affect real cost.

When Luna makes sense

Luna’s 0.1x multiplier makes it attractive for high-volume work where the desired change is already clear:

• adding tests from an approved design;
• applying a known pattern across several files;
• updating documentation;
• small, well-scoped bug fixes;
• summarizing code or producing a first-pass inventory.

The cheap multiplier should not become an excuse to skip review. If Luna starts retrying, misunderstanding architecture, or producing broad diffs, switching to Terra can be cheaper than continuing with a weaker fit.

The 272K Kiro context limit

Kiro documents a 272K context window for all three GPT-5.6 variants. That differs from the larger context figures developers may see for the model family on other surfaces.

Plan for the context Kiro actually exposes. Keep specifications focused, split unrelated work, and use steering files for stable rules. A million-token model specification elsewhere does not mean Kiro will place an entire monorepo into one request.

Cost and availability boundaries

Three surfaces are easy to conflate:

• Kiro: uses Kiro subscriptions, credits, multipliers, and a 272K context limit.
• Direct OpenAI API: uses OpenAI model pricing and API eligibility.
• Amazon Bedrock in AWS GovCloud: GPT-5.6 Terra and Luna are available in GovCloud US-East and US-West through Bedrock. AWS did not include Sol in that specific announcement.

Bedrock GovCloud availability does not mean Kiro runs a project through your Bedrock account, and it does not turn a Bedrock model into direct OpenAI API access. Treat identity, billing, regional controls, and data terms separately for each surface.

Recommended starting configuration

For most teams, start with Terra as the default, Luna for clearly mechanical tasks, and Sol behind an explicit escalation rule. Put that policy in your steering guidance so the team does not choose models by instinct alone.

Then measure:

• credits consumed per completed task;
• retries and abandoned attempts;
• review findings;
• test failures after generation;
• time from approved spec to merged change.

The cheapest request is not the goal. The useful metric is the cost of a correct, reviewed change.

Bottom line

GPT-5.6 gives Kiro a credible model ladder. Terra is the practical default, Luna is the volume tool, and Sol is the escalation path for decisions where deeper reasoning justifies higher credit use.

The strongest reason to use the family in Kiro is not a benchmark. It is the combination of model routing with requirements, design, task checkpoints, steering, and automated validation. Use those controls; otherwise you are paying for a spec-driven environment while working as if it were an ordinary chat interface.

For the product experience behind that workflow, read the Kiro one-week review. For broader choices, compare the best AI coding agents and review the AI Testing & Evaluation hub.

Primary sources

• OpenAI: GPT-5.6 in Kiro
• Kiro model reference
• Kiro GPT-5.6 changelog
• AWS: GPT-5.6 Terra and Luna in AWS GovCloud