The same model's official price dropped by only 20%, but your bill decreased by eighty percent.
On August 24, OpenAI released test results conducted with AWS:
On Terminal-Bench 2.1, the cost of completing a successful task with GPT-5.6 Terra in Kiro has decreased by approximately 82%.
Kiro is AWS's software development agent platform, covering IDE, CLI, and Web.
The three brothers—Sol, Terra, and Luna—from the GPT-5.6 family have been running inside for over a month.
This 82% is not an official price reduction.
Terra's most recent price adjustment was on July 30, with a 20% change.
On July 30, OpenAI announced a price adjustment, with Terra reduced by 20%.
A 20% reduction in unit price can cut your bill by up to 80%.
The sixty percentage points in between are what truly deserve our attention.
GPT-5.6 logged into Kiro in July this year.
On the 13th, AWS announced that GPT-5.6 Sol, Terra, and Luna are now available on Amazon Bedrock.
The next day, Kiro posted a blog announcing the launch of three models on IDE, CLI, and Web.
This is the first time the OpenAI model has entered Kiro, coinciding with the first anniversary of Kiro's public preview.
First, put the model on the shelf, then send it off to work; over a month later, OpenAI returns with its assignment:
The two teams jointly fine-tuned the Kiro environment with the OpenAI model, reducing Terra’s cost to complete a successful task by approximately 82%.
Saved
Money spent on taking a detour
On the July 30 price adjustment, Terra decreased by 20%, but in Kiro's testing, the cost per task dropped by 82%.
Where did the additional 60% savings come from?
There are only a few directions:
The model generates fewer tokens, calls tools fewer times, and requires fewer retries or detours after failures.
So the significant savings aren't per-call costs, but rather the costs of calls that would have otherwise been wasted.
It’s simple: if an AI agent messes up a task once, you still get billed. If it takes the wrong path and goes off track for three rounds before correcting itself, you’re still charged for those tokens.
In real development, money often slips away this way.
OpenAI also highlighted the same logic in its official blog: efficiency comes from three layers:
The agent framework that initiates requests and organizes context, the orchestration system that mediates requests in between, and finally the model itself running on the GPU.
OpenAI breaks down the efficiency sources of GPT-5.6: from the agent framework, through orchestration system scheduling, to GPU model execution—each layer is optimized for savings.
Save money even by optimizing model allocation.
OpenAI also provided an example usage: the coding workflow can first use Sol to clarify the problem and establish a plan, then switch to Luna to implement the well-defined changes, write tests, and run evaluations.
The same production line, with different levels of intelligence assigned to each stage.
The model accounts for only half of the bill.
Change the framework, change the price.
This set of questions for Terminal-Bench 2.1 is not intended for the model to answer individually.
It feeds the model into a terminal environment, gives it a vague goal, and lets it autonomously plan its path, invoke tools, write scripts, handle errors, and iterate repeatedly.
Therefore, the results achieved are those of the combination of "agent + model."
The Terminal-Bench 2.1 public leaderboard includes an additional cost metric to the right of accuracy. (Source: Terminal-Bench)
The four rows of numbers on the leaderboard best illustrate the point:
Claude Code with Fable 5, 83.8%, $552.67;
Codex paired with GPT-5.5, 83.1%, $2,059.19;
Codex paired with GPT-5.6 Terra, 78.4%, $421.15;
Codex paired with GPT-5.6 Luna, 75.7%, $241.45.
The scores of the first two rows differed by only 0.7 percentage points, but the bills differed by nearly four times.
The same model, plugged into different frameworks with varying context organization and tool strategies, can yield completely different results.
According to Kiro's official data, Terra scores 77.4 on the Coding Agent Index, just slightly higher than Claude Fable 5's 77.2.
Its appeal lies not in the score itself, but in the price corresponding to that score.
Kiro’s approach with spec-driven development also centers here: the core rule is simple—don’t start writing code right away.
It first breaks down the user’s vague goal into a formal requirements document, technical design, and an actionable task list, then passes it on to the model.
This way, instead of receiving a vague statement, the model gets a clearly defined task.
Those familiar with agents will immediately recognize that this step eliminates the most expensive part of the cost.
Models going off track, requiring rework, or being completely restarted often consume more tokens than actually doing the work.
Kiro still has two checkpoints in the process: pause before actually modifying the code to have someone review it, and automatically run a round of tests after completion to verify correctness.
Each rework prevented by these two checkpoints saves real money.
Claude on Amazon's turf
GPT took half.
A year ago, Kiro was merely a spec-driven IDE, with the model selector firmly in Anthropic's domain.
A year later, AWS introduced three OpenAI models on its own developed agent platform.
A year ago, it would have been hard to imagine Sol, Terra, Luna, and Claude all listed together in the same dropdown menu.
Although GPT-5.6 is now "fully integrated," it has not been "fully opened."
The three models are being gradually and experimentally released to Pro, Pro+, Pro Max, and Power users, with availability limited to two regions: Northern Virginia in the United States and Frankfurt in Europe, supporting cross-region inference.
Another aspect that many users find unfamiliar: these models use hidden reasoning chains in Kiro—you can’t see their step-by-step reasoning, only the final result.
People accustomed to watching an agent reason step by step may feel as if they’re throwing a task into an opaque box.
According to official statements, this is expected behavior and does not affect output quality.
The three-tier model is clearly priced in Kiro.
When it launched on July 14, the same task deducted 2.4 times for Sol, 1.2 times for Terra, and 0.6 times for Luna.
On July 30, OpenAI implemented its recent price reduction; the next day, Kiro followed suit: Luna's multiplier was slashed from 0.6x to 0.1x, Terra's dropped from 1.2x to 1.0x, while Sol remained unchanged.
AWS has made its position clear: a development platform cannot be tied to just one model.
Within the same selector, two leading models begin to undercut each other.
The evaluation criteria for the model have also changed: a high score no longer guarantees victory, and spending less can also lead to winning.
For developers, what used to be asked as "How much does this model cost per million tokens?" is now being asked as "How much will it actually cost me to get this task done?"
Reference materials:
https://x.com/OpenAIDevs/status/2091966982015103068
https://openai.com/index/gpt-5-6-in-kiro/
This article is from the WeChat public account "New Intelligence Yuan" (ID: AI_era), author: ASI Revelation, editor: Yuan Yu.