AI・テクノロジー
GPT-5.5 Captivates Engineers | Why It's Not the Top Performer and How Evaluation Criteria Are Changing
機械翻訳 / Machine-translated
A
AI Friends
2026年5月12日・14分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What You'll Learn in This Article
On April 23, 2026, OpenAI announced a new AI model called "GPT-5.5." This model is designed as a "coding agent" specialized in programming assistance.
Here's the thing: GPT-5.5 does not achieve the highest scores across all benchmarks (standardized tests that measure AI performance). For example, on "SWE-Bench Pro," a test that measures the ability to solve programming problems, GPT-5.5 scored 58.6%. Meanwhile, its rival Claude Opus 4.7 recorded a higher score of 64.3%.
Yet many engineers are paying close attention to GPT-5.5. Why is that?
The biggest reason GPT-5.5 is being praised is a capability called "agentic durability." In plain terms, this means "the ability to keep thinking and acting independently all the way to the finish."
Traditional AI models would sometimes stall midway if humans didn't provide detailed step-by-step instructions. For instance, if an error appeared while writing code, they would often respond with something like, "What should I do?"
GPT-5.5 is different. When an error occurs, it investigates the cause on its own, thinks through a fix, actually applies it, and then moves on to the next task. In other words, it can "run through to the end on its own" even without a human watching.
Let's look at some concrete numbers. GPT-5.5 recorded a score of 82.7% on a test called "Terminal-Bench 2.0," which measures how completely a model can carry out complex tasks using the command line (a method of operating a computer using only text).
This score represents a significant improvement over previous models. For context, GPT-5.5 also recorded 78.7% on "OSWorld-Verified," a test for general computer operation, and 84.9% on "GDPval," a test for knowledge work tasks.
For a long time in the AI world, "achieving high scores on benchmarks" was considered the most important measure of success. Benchmarks are standardized tests used to measure AI performance.
But the landscape changed in 2026. Engineers began demanding AI that "sees real work through to completion" over AI that "scores high on tests."
There is data that illustrates this shift. According to the "AI Index 2026" published by Stanford University, the probability of an AI agent successfully completing a PC operation task was just 12% in 2024. By 2026, that figure had risen to 66%—an improvement of more than fivefold in just two years.
This rapid evolution has also dramatically changed how AI is evaluated. Whereas the focus used to be on "how high a score can it achieve on individual tests," three qualities are now considered important:
In short, "how reliable is it to delegate work to" has become more important than "how intelligent is it."
Another defining feature of GPT-5.5 is its high "token efficiency." Tokens are the smallest unit by which AI processes text—the more tokens required, the higher the processing cost.
GPT-5.5 handles the same tasks using 72% fewer tokens than other models. This means it can deliver accurate answers in shorter form.
Specifically, GPT-5.5 is built around the design philosophy of "achieving higher scores with fewer tokens." By cutting out unnecessary explanations and focusing only on what's needed, it can deliver strong results while keeping costs down.
That said, API pricing (the fee for using AI from within a program) has been set at $5.00 for input and $30.00 for output—double that of the previous generation. Even so, because of its high token efficiency, the actual cost per task is often lower than with other models.
GPT-5.5's main rival is "Claude Opus 4.7," announced by Anthropic on April 16, 2026. How do the two differ?
Claude Opus 4.7 recorded 64.3% on SWE-Bench Pro and 87.6% on SWE-bench Verified, and it is said to have reclaimed the throne in coding capability. It is particularly strong at understanding large codebases (entire programs made up of many files) and handling wide-ranging architectural changes.
GPT-5.5, on the other hand, excels at precise tool usage, file manipulation, and solving concrete problems. In short, Claude is better suited for painting the big picture, while GPT-5.5 is characterized by its ability to handle detailed tasks accurately.
Experts point out that "which one is better depends on what kind of work you're doing." Claude is the right choice when you need to survey a large-scale application as a whole, while GPT-5.5 is better suited for automating specific tasks.
In "GitHub Copilot," a tool popular among programmers, selecting GPT-5.5 applies a billing "multiplier" of 7.5x. This means the same work costs 7.5 times more than usual.
As a result, when adopting GPT-5.5 at scale in an enterprise setting, the Total Cost of Ownership (TCO) could increase significantly. Careful consideration is required from a cost perspective.
On the other hand, given GPT-5.5's high token efficiency, the actual cost per task may in some cases be lower than with other models. It's important to assess which model is optimal based on your specific usage situation.
With the arrival of GPT-5.5 and Claude Opus 4.7, competition in the AI industry has entered a new phase. Experts note that "this landscape could change dramatically within a month," signaling that the pace of AI development is accelerating.
Going forward, efficiency evaluation in concrete operational scenarios will matter more than simple performance comparisons. In other words, rather than just knowing "what score this AI can get on a test," it will be essential to determine "how useful it actually is for real work."
Additionally, 2026 is seeing the diversification of AI evaluation. Different metrics are now used for different purposes: SWE-bench has become the standard for coding, BrowserGym for web navigation, and τ-bench for enterprise environments.
These shifts indicate that AI is evolving from "a research lab toy" into "a practical tool for real work."
The arrival of GPT-5.5 raises a fundamental question for the AI industry: "What really matters?" Not just being intelligent, but being reliable enough to delegate to. That is what the AI of tomorrow needs to be.
This article is a cross-post from AI Friends.
