A mere 20% price cut leads to an 80% reduction in total billing, GPT-5.6 breaks into Claude's territory to recalculate the programming cost account.

For the same model, the official price has only been reduced by 20%, but your bill has dropped by 80%.

On August 24, OpenAI announced the results of a test conducted jointly with AWS:

On Terminal-Bench 2.1, the cost for GPT-5.6 Terra to complete a successful task in Kiro was reduced by approximately 82%.

Kiro is AWS's software development agent platform that covers IDE, CLI and Web.

The three members of the GPT-5.6 family, Sol, Terra and Luna, have been running on it for more than a month.

This 82% cost reduction does not come from official price cuts.

The last price adjustment for Terra was on July 30, with a reduction of 20%.

OpenAI's price adjustment announcement on July 30: Terra's price was reduced by 20%.

The unit price was only cut by 20%, but the bill can be reduced by 80%.

Where the remaining 60 percentage points of savings come from is what truly deserves our attention.

GPT-5.6 was launched on Kiro in July this year.

First, on the 13th, AWS announced that GPT-5.6 Sol, Terra and Luna were officially available on Amazon Bedrock.

The very next day, Kiro posted a blog announcing that the three models were launched on IDE, CLI and Web.

This is the first time OpenAI's models have entered Kiro, which coincides with the first anniversary of Kiro's public preview.

After putting the models on the shelves and putting them into practical use for more than a month, OpenAI came back to deliver the results:

The two parties jointly optimized the Kiro environment and OpenAI's models, and the cost for Terra to complete a successful task was reduced by approximately 82%.

The savings

are the costs that would have been wasted on detours

During the price adjustment on July 30, Terra's price was reduced by 20%, but in Kiro's test, the cost per task was reduced by 82%.

Where does the extra 60% of savings come from?

The possible directions are only a few:

The model generates fewer tokens, the number of back-and-forth tool calls is reduced, and retries and detours after failures are fewer.

Therefore, this large portion of savings does not come from the cost of each call, but from those calls that would otherwise have been wasted.

The logic is simple: if an AI agent messes up a task, the bill will still be charged. If it chooses the wrong path halfway, goes off track for three rounds before getting back on track, these tokens will also be charged.

In real development, costs are often wasted in this way.

OpenAI also mentioned 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 schedules requests in the middle, and finally the model itself running on the GPU.

OpenAI breaks down the efficiency sources of GPT-5.6: requests start from the agent framework, are scheduled by the orchestration system, and finally run the model on the GPU, with savings achieved at every layer.

Cost savings can also be achieved through model division of labor.

OpenAI also cited a usage scenario: in the coding workflow, Sol can be used first to clarify the problem and set the plan, then Luna is used to implement the clearly defined modifications, write tests and run evaluations.

In the same pipeline, different levels of intelligence are allocated for different links.

Models only account for half of the bill

Switching frameworks changes the total cost

The test suite of Terminal-Bench 2.1 does not require the model to answer questions alone.

It puts the model into a terminal environment, gives it a vague goal, and lets it plan the path, call tools, write scripts, handle errors, and iterate repeatedly on its own.

Therefore, the score obtained is the score of the "agent + model" combination.

The public leaderboard of Terminal-Bench 2.1 adds a new cost item next to the accuracy. (Source: Terminal-Bench)

The four lines of figures in the leaderboard best illustrate the point:

Claude Code paired with Fable 5, 83.8%, $552.67;

Codex paired with GPT-5.5, 83.1%, $2059.19;

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 lines differ by only 0.7 percentage points, but the bill differs by nearly 4 times.

For the same model, if it is put into different frameworks, paired with different context organization and tool strategies, the resulting cost will be completely different.

According to the official data provided by Kiro, Terra scored 77.4 on the Coding Agent Index, which is only slightly higher than Claude Fable 5's 77.2.

Its selling point is not the score, but the cost corresponding to this score.

Kiro's spec-driven focus is also here, and the core gameplay can be summed up in one sentence: do not start writing code immediately.

It first breaks down the user's vague goal into formal requirement documents, technical designs and executable task lists, and then hands them over to the model.

In this way, what the model gets is no longer an ambiguous sentence, but a clearly organized task.

People familiar with Agents will immediately realize that this step saves the most expensive part of the overhead.

The tokens burned by the model going off track, reworking, and starting all over again are often more than those consumed by doing the task properly.

Kiro also sets two checkpoints in the process: before the code is actually modified, it pauses for human review; after the task is completed, it automatically runs a round of tests to verify correctness.

Every rework prevented by these two checkpoints can save real money.

On AWS's home turf where Claude once dominated

GPT has taken half of the market share

A year ago, Kiro was just a spec-driven IDE, and the model selector was dominated by Anthropic.

A year later, AWS has added three tiers of OpenAI models to its own development agent platform in one go.

Sol, Terra and Luna are listed side by side with Claude in the same drop-down menu, a scenario that was hard to imagine a year ago.

Although GPT-5.6 has "settled in as a whole family" this time, it is not "fully open".

The three models are opened in a gradual and experimental way, targeting Pro, Pro+, Pro Max and Power users, covering only two regions: Northern Virginia in the United States and Frankfurt in Europe, with cross-region reasoning supported.

There is also a point that many people are not used to: these models run hidden chain of thought in Kiro, so you cannot see their reasoning steps, only the final result.

People who are used to watching Agents reason step by step will feel like they have thrown the task into an opaque box.

The official statement says this is the expected behavior and does not affect the output quality.

The three tiers of models have clear price tags on Kiro.

When they were first launched on July 14, for the same task, Sol charged 2.4 times the base rate, Terra charged 1.2 times, and Luna charged 0.6 times.

After OpenAI's price cut on July 30 took effect, Kiro followed up the next day: Luna's rate was cut from 0.6 times to 0.1 times, Terra's rate was reduced from 1.2 times to 1.0 times, and only Sol's rate remained unchanged.

AWS's attitude is clear: a development platform cannot be tied to only one model provider.

In the same selector, the two leading models start to undercut each other on price.

The evaluation criteria for models have also changed: high scores no longer guarantee victory, and low cost can also be a winning advantage.

For developers, before choosing a model, they used to ask "How much does this model cost per million tokens", but now they have to ask "How much will it cost me to get this task done".

References:

https://x.com/OpenAIDevs/status/2091966982015103068

https://openai.com/index/gpt-5-6-in-kiro/

This article is from the WeChat Official Account "AI Era" (ID: AI_era), written by ASI Revelation, edited by Yuan Yu, and authorized for release by 36Kr.