When a new AI model comes out,
we choose the strongest model first.
GPT-5.6 Sol
shows why that habit is wrong.
The most important thing is not the highest performance.
It's about designing
the desired results with the least cost.
I. We chose the model too simply
When a new model is announced,
people look at benchmarks first.
How much smarter is it than the previous model?
How much has the coding score improved?
How much longer is the context?
And the model that receives the highest score
is set as the default.
This method is convenient.
But it's expensive and slow.
Above all,
it misses the changes in GPT-5.6.
GPT-5.6 is not just one model.
It is divided into three tiers:
Sol, Terra, and Luna.
Sol is a flagship for complex coding, computer usage,
research, and security tasks.
Terra balances performance and cost.
Luna handles fast and inexpensive
bulk tasks.
The default Power setting
uses Sol and intermediate inference.
OpenAI Model Selection Guide
The question changes here.
It's not "What is the smartest model?"
It's "What is the smallest model
needed for this task?"
II. Real change is not in performance but in combinations
GPT-5.6 Sol is powerful.
It supports a context of 1.05 million tokens.
The maximum input is 922,000 tokens.
The maximum output is 128,000 tokens.
It takes text and images as input
and utilizes web search, file search, code execution,
computer usage, MCP, and skills.
GPT-5.6 Sol Official Specifications
But focusing only on numbers
misses the essence.
The real change in GPT-5.6
is the increased axes to choose from.
Even after choosing a model,
you still need to make three more decisions.
GPT-5.6's 4-axis tuning stack
Model Tier
Choosing between Sol, Terra, and Luna
for which engine to use.Inference Strength
Choosing from
none,low,medium,
high,xhigh,max
how deeply to think.Execution Mode
Using the standard
standard.Using
proto increase reliability
with more operations.Prompt Structure
Instructing what
to handle and what to leave to the model.
I call these four axes
the MREP stack.
Model.
Reasoning.
Execution.
Prompt.
Now, model selection
is not just choosing a product.
It's designing server infrastructure.
"Use the lowest reasoning effort that produces the result you need."
— OpenAI Model Guide
If the desired result is achieved,
lower inference strength is better.
Using the highest specifications
is not optimization.
It's giving up on measurement.
III. Sol is not always the answer
For example,
Classify 1,000 customer inquiries daily
and summarize them in three lines.
Complex reasoning is not necessary.
The result format is fixed.
For this task, using Sol and Pro mode
will yield good results.
But even Luna
can produce the same quality.
In this case, Sol
is not about performance improvement but waste.
There are also opposite situations.
Analyze security vulnerabilities
spanning dozens of files.
Trace code flow
and distinguish between attack possibilities and false positives.
Also review if the fixes
introduce new vulnerabilities.
In this case, Sol is the right choice.
You can compare High or Extra High
and even test up to Max if necessary.
If you can investigate multiple areas independently,
multi-agent can also be a candidate.
There is also a third scenario.
An executive needs to read 30 market reports
and compare competitor features and prices.
It is not necessary for Sol
to read each document line by line.
Tools collect data.
Code removes duplicates.
Only relevant results are kept.
The model receives compressed evidence
and focuses on the final judgment.
This is the Programmatic Tool Calling of GPT-5.6.
The model writes JavaScript
to call allowed tools in parallel.
It uses loops and conditional statements.
Processes intermediate results within the execution environment
and sends back only small structured results as context.
Instead of the model reading all data,
the code handles mechanical tasks.
Programmatic Tool Calling Guide
AI has not just become smarter.
AI has now been given the choice
of when to think
and when to process as code.
IV. 8 Changes in GPT-5.6 Sol
1. Programmatic Tool Calling
Traditional tool calls
were interactive.
The model calls the tool.
Reads the result.
Decides on the next call.
With each call,
intermediate results are added to the context.
GPT-5.6 can bundle
predictable tasks into JavaScript programs.
Parallel calls.
Filtering.
Sorting.
Duplicate removal.
Aggregation.
Validation.
It processes these tasks as code
and only sends the final result to the model.
However, it is not suitable for all tool operations.
If new meaning judgments are needed
with each result,
direct tool calling is better.
Tasks requiring approval, such as payments or deletions,
are also better handled this way.
2. Multi-Agent Beta
The GPT-5.6 model
can create multiple sub-agents under a root agent
and operate them in parallel.
One agent looks at security.
One agent focuses on accuracy.
Another agent finds missing tests.
The root agent
collects results and resolves duplicates and conflicts.
Beneficial for exploring large codebases.
Comparing multiple documents.
Verifying multiple hypotheses.
It is advantageous for implementing independent components.
However, this is not always a quick fix.
It is necessary for tasks requiring sequential reasoning.
Tasks where editing the same file simultaneously.
If a slow external API determines the overall time,
multi-agent may not be helpful.
It can also increase token usage.
The fact that this is currently a beta feature is also important.
OpenAI Multi-Agent Guide
3. Persisted Reasoning
In previous multi-turn tasks,
the structure of previous judgments could weaken
even if the conversation continued.
GPT-5.6 allows the reuse of inference items
for subsequent turns.
Through reasoning.context,
you can decide whether to use it only for the current turn
or to continue across multiple turns.
There is an important point.
This feature does not expose
the model's raw thought process.
Inference items are kept in an opaque state.
The effect lies in continuity, not exposure.
For tasks like long debugging or research
where goals and assumptions are maintained over multiple turns,
there is less need to repeat the same judgment from scratch.
Preserve Reasoning Guide
4. Pro Mode
Pro is not a separate model name.
It is not about switching models to gpt-5.6-pro.
In the same GPT-5.6 model,
set reasoning.mode: "pro".
Pro mode performs more model tasks
to produce a single final answer.
Suitable for tasks requiring high-level optimization.
Costly code reviews.
Tasks where quality differences
affect actual results, such as costly analyses.
However, it increases latency.
It also increases tokens and costs.
"Choose pro mode when quality matters most."
— OpenAI GPT-5.6 Guide
"Important tasks"
and "difficult tasks" are different.
Pro is not needed for simple but important tasks.
It should be used when tasks are difficult
and quality differences can lead to actual losses.
5. Max Inference Strength
GPT-5.6 supports max inference.
While it may seem that better answers
come from thinking more, official recommendations differ.
If currently using xhigh,
compare xhigh and max in the same representative task.
If accuracy remains the same
but costs and time have increased,
Max is not an improvement.
Measure the business success rate
rather than the model's thought volume.
6. Explicit Prompt Caching
In GPT-5.6,
you can specify prompt sections to reuse.
Long system instructions.
Fixed reference documents.
Definitions of tools used repeatedly
are placed at the beginning of the prompt.
Content that changes, like user questions,
is placed at the end.
This allows the same prefix segment
to be read from the cache.
However, starting from GPT-5.6,
cache writing is charged 1.25 times the regular input.
While cache reading is discounted,
if the prompt is written once and not reused,
it can actually be a loss.
So, don't just look at cached_tokens.
You should also measure cache_write_tokens.
Cache is not a cost-saving feature.
It is an investment that saves
only when repetition is sufficient.
Prompt Caching Guide
7. Frontend Design and Intent Understanding
GPT-5.6 has improved in layout, visual hierarchy,
and design judgment.
It focuses not only on generating code
but also on creating more organized and user-friendly interfaces.
It is designed to infer not only from the user's surface sentences
but also from the actual goals and expectations.
Therefore, there is less need to write
prompts for every step.
However, omitting boundary conditions and success criteria
is not recommended.
You should distinguish between areas
the model should handle on its own
and constraints that must be adhered to.
8. Original Image Resolution
GPT-5.6 can preserve
the original size of images passed in as original or auto.
Beneficial for tasks where detailed visual information is crucial,
such as small buttons on large screens.
Complex dashboards.
Drawings where spatial relationships are important.
Tasks requiring precise visual information, like click locations in computer usage.
However, large images
can increase token usage and latency.
Resolution is also the same as the model.
Higher is not always better.
It should be increased when necessary.
V. The most paradoxical change in the AI era is the prompt
Among the new features of GPT-5.6,
the most important may not be a feature.
The way prompts are written has changed.
We have been adding instructions
every time the model makes a mistake.
"Always double-check."
"Never guess."
"Think step by step."
"Write concisely."
"Revalidate."
These sentences have piled up
and the prompt has become a configuration file.
The problem is that even as the model evolves,
the configuration file remains the same.
Even a 30-year developer
faces the same temptations with a new version.
Adding a new option is more comfortable
than removing the existing configuration.
But legacy prompts
are like legacy code.
Conditions created to prevent past bugs
can hinder performance in new versions.
In OpenAI's internal evaluations,
switching from long and explicit system prompts
to smaller prompts resulted in about a 10-15% improvement.
The total tokens decreased by 41-66%
and costs decreased by 33-67%.
It is not guaranteed that the same results
will be achieved in all tasks.
But the direction is clear.
"Use shorter prompts."
— OpenAI GPT-5.6 Guide
There is also a trap here.
Do not strongly repeat "Answer concisely."
GPT-5.6 inherently prefers concise responses
compared to previous models.
Excessive instructions for conciseness
do not mean just removing introductions.
It can cut out necessary evidence and outputs.
A good prompt
is not just a short prompt.
A prompt that is concise where necessary
and clear about success criteria is the best prompt.
Bad practice looks like this.
간결하게 답해라.
불필요한 설명을 하지 마라.
핵심만 써라.
길게 쓰지 마라.
반복하지 마라.
A better approach is like this.
결론부터 제시한다.
결론을 뒷받침하는 근거와 중요한 예외,
다음 행동은 유지한다.
서론, 반복, 일반론부터 제거한다.
Instead of saying to reduce the volume,
it is about specifying what to keep.
VI. The competitiveness in the AI era is not model access
GPT-5.6 Sol is powerful.
It handles long contexts.
It bundles tools into programs.
It creates sub-agents to operate in parallel.
It continues previous inferences into the next turn.
It also provides deeper Pro mode
and Max inference.
But turning on all these features
is not the answer.
When operating servers,
you don't send all requests to the most expensive instances.
In databases,
you don't allocate maximum resources to every query.
You route, measure, and adjust
based on importance and difficulty.
AI has reached the same stage.
Models are no longer
products to purchase.
They are infrastructures to operate.
Prompts are not just commands.
They are architectures that connect models, tools, and permissions.
In the future, those who use AI well
are not the ones with the strongest models.
They are the ones who know
where to send tasks, how much to think,
and what to delegate to code.
Don't buy the most powerful model.
Design the cheapest path to success.
Try removing redundant three lines
from the prompt you are currently using.
Then compare Sol Medium
with a setting one step lower.
Decide based on results, not feelings.