OpenAI and AWS announced on August 24, 2026, that OpenAI’s GPT-5.6 model family is available inside Kiro, Amazon’s agentic development platform built around structured, spec-driven coding. The move gives Kiro developers a new model option alongside the Claude models the platform has relied on since launch — and it marks the first time OpenAI models have been built into Kiro’s workflow.
The announcement itself is a formal partnership statement rather than a first-time technical rollout. GPT-5.6 has actually been available inside Kiro on an experimental basis since mid-July 2026, shortly after the model family reached general availability in ChatGPT, Codex, and the OpenAI API on July 9. What changed on August 24 is that OpenAI and AWS put their names jointly behind the integration, with executives from both companies describing it as a step toward giving developers more control over the tradeoff between intelligence, speed, and cost.
Colleen Kapase, OpenAI’s Vice President of Strategic Global Partnerships and Ecosystems, said the integration gives developers “more room to match intelligence, speed, and cost to each stage of the software development lifecycle.” AWS’s Vice President of Agentic AI, Swami Sivasubramanian, echoed that framing, saying the addition helps “accelerate AI-native development using Kiro.”
Table of Contents
What Is Kiro?
Kiro is a standalone development environment built by AWS — not a wrapper around an API, but a full IDE forked from VS Code (Code OSS), along with a command-line interface and a web version. It reached general availability on November 17, 2025, after a four-month public preview that drew more than 250,000 developers.
What sets Kiro apart from many AI coding tools is its emphasis on structure before code. Rather than generating output directly from a prompt, Kiro converts a developer’s intent into a written specification: clear requirements, a technical design document, and a task list, before an AI agent starts implementing anything. The idea is to reduce the kind of unstructured “vibe coding” where an assistant guesses at implementation and developers spend the next hour patching gaps.
Kiro also includes property-based testing, a verification method introduced at general availability. Unlike conventional unit tests that check a handful of specific cases, property-based tests run hundreds of times against randomly generated inputs, looking for edge cases that a human tester might miss. That capability is central to what OpenAI’s announcement describes as developers being able to “check correctness of implementation” using GPT-5.6 in Kiro.
The GPT-5.6 Model Tiers: Sol, Terra, and Luna
GPT-5.6 isn’t a single model — it ships in three tiers designed to sit at different points on a cost-versus-capability curve. Sol is the flagship, built for the hardest multi-step engineering work. Terra is positioned as a balanced option for everyday development tasks. Luna is the fastest and cheapest tier, intended for high-frequency, lower-complexity work.
Inside Kiro, each tier consumes the platform’s credit system at a different rate: Sol carries a 2.4x credit multiplier, Terra a 1.2x multiplier, and Luna a 0.6x multiplier, according to Kiro’s own documentation. All three models are limited to a 272,000-token context window within Kiro — notably smaller than the one-million-token window OpenAI and AWS have made available for the same models through Amazon Bedrock directly. That gap reflects a platform-level constraint Kiro has chosen to apply, not a limitation of the underlying model.
One other detail worth knowing: GPT-5.6 in Kiro uses a hidden chain-of-thought reasoning process. Developers see the model’s final output, not the intermediate reasoning steps — a design choice Kiro says is expected behavior and doesn’t affect output quality, though it does mean less visibility into how the model arrived at a given answer.
Pricing and Availability
Kiro’s pricing runs on a subscription-plus-credits model, separate from OpenAI’s standard API pricing. The platform offers a free tier with 50 monthly credits, then four paid tiers: Pro at $20 per month, Pro+ at $40, Pro Max at $100, and Power at $200, each with a larger monthly credit allowance. Users on paid plans can also purchase add-on credits at $0.04 each once they exceed their monthly allotment.
GPT-5.6 access is currently restricted to the four paid tiers — Free-tier users do not have it — and to two AWS regions: US East (Northern Virginia) and AWS Europe (Frankfurt), with cross-region inference support to help route requests. AWS and OpenAI describe the rollout as gradual and experimental, and neither company has published a timeline for expanding to additional regions or to the free tier.
It’s worth separating this from OpenAI’s direct API pricing for GPT-5.6, which is billed per million tokens: $5 input / $30 output for Sol, $2.50 input / $15 output for Terra, and $1 input / $6 output for Luna. Those figures apply to direct API use, not to how the models are metered inside Kiro’s credit system.
Performance Claims and Benchmarks
OpenAI and AWS point to one specific efficiency claim tied to Kiro: on Terminal-Bench 2.1, a benchmark that tests complex command-line workflows, GPT-5.6 Terra completed successful tasks in Kiro at roughly an 82% cost reduction. Neither company’s announcement specifies what that reduction is measured against — whether it’s compared to an earlier model generation in Kiro, an unstructured workflow, or something else — so the figure should be read as a company-reported estimate rather than a fully documented, independently verified result.
Broader coding benchmarks, published by OpenAI for the GPT-5.6 family generally, offer more context. On the Artificial Analysis Coding Agent Index, Sol scores 80, ahead of GPT-5.5’s 76.4 and Anthropic’s Claude Fable 5 at 77.2. On Terminal-Bench 2.1, Sol scores 88.8%, again ahead of the same comparison models. But GPT-5.6 doesn’t lead every coding measure: on SWE-Bench Pro, which evaluates real-world GitHub issue resolution, Claude Fable 5 scores 80% and Claude Mythos 5 scores 80.3%, both ahead of Sol’s 64.6%.
OpenAI itself notes that its cost and latency comparisons across models are estimates derived from simulating production behavior, not measurements from live deployments, and that “real-world results may vary substantially.” That caveat applies to the benchmark figures above as much as to the Kiro-specific cost-reduction claim.
Business Context: The OpenAI–AWS Partnership
The Kiro integration is the latest step in a fast-moving expansion of OpenAI’s commercial relationship with AWS. The two companies first announced OpenAI models coming to Amazon Bedrock in April 2026. GPT-5.6 reached general availability on Bedrock in mid-July, alongside its Kiro debut. In early August, AWS and OpenAI launched Daybreak Red and Daybreak Blue on Bedrock for vetted customers — Daybreak Red providing access to GPT-5.6 Cyber, a purpose-trained cybersecurity model, and Daybreak Blue providing GPT-5.6 Sol with safeguards calibrated for defensive security work. By mid-to-late August, AWS had expanded GPT-5.6’s availability on Bedrock to more than 25 regions through cross-region inference, and OpenAI had cut Sol’s API pricing by more than 20% for a three-month window.
Neither company has disclosed the financial terms of the broader partnership or any revenue-sharing arrangement specific to Kiro.
How This Compares to Claude in Kiro
Kiro was built on Anthropic’s Claude models from the start, and Claude options — including Opus 4.8 and Sonnet 4.6 — remain available in the platform’s model selector. GPT-5.6 is being added as an additional choice, not a replacement, giving developers the ability to route different tasks to different models depending on cost and performance needs, either manually or through Kiro’s “Auto” model-routing option.
The benchmark data suggests neither vendor is uniformly ahead. GPT-5.6 leads on the Artificial Analysis Coding Agent Index and Terminal-Bench 2.1; Claude models lead on SWE-Bench Pro. For developers evaluating the choice, the practical difference may come down less to which model tests best on a given benchmark and more to cost per task, credit consumption, and how each model performs on their specific codebase.
Limitations and Open Questions
A few practical constraints are worth keeping in mind. Availability is currently capped at two AWS regions, and only paid Kiro subscribers can use GPT-5.6 — free-tier users are excluded entirely. The hidden reasoning process means developers can’t inspect how the model reached a given output, which may matter for teams that want to audit or debug agent behavior closely. And the headline cost-reduction figure OpenAI and AWS have publicized lacks a disclosed baseline, making it difficult to independently verify how meaningful the savings are in practice.
It’s also unclear whether Kiro plans to extend GPT-5.6 access to additional regions, to its free tier, or whether it will formally support version-control integrations beyond GitHub. Kiro supports GitHub-based workflows, including assigning tasks to its agent through issue labels, but no equivalent GitLab integration has been documented.
Frequently Asked Questions
Is GPT-5.6 replacing Claude models in Kiro? No. Claude models remain available in Kiro alongside GPT-5.6. The addition gives developers more model choice rather than replacing the platform’s existing options.
How much does GPT-5.6 cost to use in Kiro? Kiro uses a subscription-plus-credit system rather than per-token billing. Sol consumes credits at a 2.4x multiplier, Terra at 1.2x, and Luna at 0.6x, on top of Kiro’s standard plans ($20–$200 per month) and $0.04-per-credit overage pricing.
What’s the difference between GPT-5.6 Sol, Terra, and Luna? Sol is the flagship model for the most demanding, multi-step engineering work. Terra is a balanced option for everyday tasks. Luna is the fastest and most cost-efficient tier, intended for high-frequency, simpler work.
Which regions currently support GPT-5.6 in Kiro? As of this announcement, GPT-5.6 is available in Kiro only in AWS’s US East (Northern Virginia) and Europe (Frankfurt) regions, with cross-region inference support.
Is GPT-5.6 better than Claude for coding tasks? It depends on the benchmark. GPT-5.6 Sol leads on the Artificial Analysis Coding Agent Index and Terminal-Bench 2.1, while Claude models score higher on SWE-Bench Pro. Neither leads across every measure.
Can Kiro’s free tier access GPT-5.6? No. GPT-5.6 is currently limited to Kiro’s four paid tiers — Pro, Pro+, Pro Max, and Power.
What does “property-based testing” mean in Kiro? It’s a testing method that runs generated tests hundreds of times against randomly varied inputs to check whether an implementation holds up against a specification, rather than checking only a fixed set of predefined cases.