GPT-5.6 in Kiro: cost per finished task, not per token

On August 24, AWS and OpenAI put the full GPT-5.6 family — Sol, Terra, and Luna — inside Kiro, AWS's spec-driven coding agent, across its IDE, CLI, and web surfaces. The number in the announcement is the one worth reading: on joint Terminal-Bench 2.1 testing, GPT-5.6 Terra completed successful tasks at roughly 82% lower cost. Not 82% cheaper tokens. Cheaper finished work.

What actually happened

Kiro is not a chat box with a code button. It converts a requirement into a written implementation plan first, runs the model through multi-step execution against that plan, stops at checkpoints where a human reviews the diff before it lands, and then verifies the result with property-based testing. The GPT-5.6 tiers slot into that loop at different price points: Kiro applies credit multipliers of 2.4x for Sol, 1.2x for Terra, and 0.6x for Luna, with a 272K context window across all three. Rollout is to Kiro Pro, Pro+, Pro Max, and Power customers in AWS US-East-1 and Europe (Frankfurt), with cross-region inference.

Read the 82% claim carefully. It is a vendor-run benchmark, on a benchmark suite, comparing cost per successful task. A mid-tier model that fails half its attempts is not cheap, and a top-tier model that one-shots everything can be. The structure around the model — plan, checkpoint, test — is what converts a cheaper tier into a lower bill instead of a longer retry loop. Developer Tech's writeup has the same framing from AWS's side.

Why cost per finished task matters for your business

Per-token pricing tells you almost nothing. You do not buy tokens. You buy merged pull requests. The unit that belongs in your spreadsheet is dollars per accepted change, and it includes every failed attempt, every re-prompt, and the engineer's time reading output that was wrong.

The harness is the variable you control. You cannot make Sol cheaper. You can make the loop around it waste less: a written spec before execution, a checkpoint before the write, a test that proves the thing works. Those three moves are why a 1.2x-multiplier model can beat a 2.4x one on total cost — and they are portable to whatever tool you are already using.

Vendor benchmarks are a hypothesis, not a result. Terminal-Bench is not your codebase. Run the same task through two tiers on your own repo, count accepted diffs, and divide. That measurement takes an afternoon and beats every leaderboard for your purposes.

Model choice is now a routing decision. Three tiers with a 4x credit spread means the interesting question is which tier handles which class of work — boilerplate to Luna, feature work to Terra, thorny debugging to Sol. That is a policy you write once, not a preference you set forever.

Sources: Kiro, OpenAI, Developer Tech.