{"id":2579,"date":"2026-04-24T03:41:26","date_gmt":"2026-04-24T03:41:26","guid":{"rendered":"https:\/\/deepinsightai.io\/?p=2579"},"modified":"2026-04-24T03:57:26","modified_gmt":"2026-04-24T03:57:26","slug":"gpt-5-5-review","status":"publish","type":"post","link":"https:\/\/deepinsightai.io\/ja\/gpt-5-5-review\/","title":{"rendered":"GPT-5.5 Arrives: Dominating Opus 4.7 Across the Board"},"content":{"rendered":"

Just moments ago, Altman dropped GPT-5.5 in the middle of the night. A full-scale strike against Claude Opus 4.7<\/a>, taking back the crown of the strongest model on earth. From coding to scientific research, the era where AI independently takes over the computer really seems to have arrived.<\/p>\n\n\n\n

Silicon Valley isn\u2019t sleeping tonight.<\/p>\n\n\n\n

Just now, GPT-5.5 made a shocking debut \u2014 OpenAI\u2019s most powerful and most capable next-generation flagship model so far.<\/p>\n\n\n\n

It represents a completely new level of intelligence, evolving fully into the \u201cnative brain\u201d of the Agent era.<\/p>\n\n\n\n

Yes, the long-awaited \u201cSpud\u201d is finally here today.<\/p>\n\n\n\n

GPT-5.5 Benchmarks: Ranking #1 Across All Categories<\/h2>\n\n\n\n

The most eye-catching part is this: in all major benchmark tests, GPT-5.5 ranks first.<\/p>\n\n\n\n

Whether it\u2019s programming, reasoning, mathematics, or agent tasks, Claude Opus 4.7 and Gemini 3.1 Pro are completely outperformed by GPT-5.5.<\/p>\n\n\n\n

Compared to the previous generation, GPT-5.5 Thinking feels like a \u201cdimensionality reduction attack,\u201d opening up a generational gap.<\/p>\n\n\n\n

In the AAI test, with the same output tokens, GPT-5.5 achieves the highest intelligence score globally. On ARC-AGI-2, it also sets a new SOTA.<\/p>\n\n\n\n

\"gpt<\/figure>\n\n\n\n
\"gpt<\/figure>\n\n\n\n

Altman couldn\u2019t help but praise it: \u201cGPT-5.5 is both smart and fast.\u201d<\/p>\n\n\n\n

\"gpt<\/figure>\n\n\n\n
\"gpt<\/figure>\n\n\n\n

The speed per token is as fast as GPT-5.4, while significantly reducing the number of tokens used per task.<\/p>\n\n\n\n

It can almost intuitively understand what needs to be done.<\/p>\n\n\n\n

Greg, the president, said excitedly, \u201cThis is a step toward a completely new way of working with computers.\u201d<\/p>\n\n\n\n

Starting today, GPT-5.5 is officially live in ChatGPT and Codex<\/a>.<\/p>\n\n\n\n

GPT-5.5 Coding Performance: A New King Rises<\/h2>\n\n\n\n

Let\u2019s start with the core field \u2014 programming. GPT-5.5 delivers a strong comeback.<\/p>\n\n\n\n

According to OpenAI, it is the most powerful agentic coding model to date, revolutionizing workflows much like the shift \u30d0\u30a4\u30d6\u30fb\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0\u304b\u3089\u30a6\u30a3\u30c3\u30b7\u30e5\u30fb\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0\u3078<\/a>.<\/p>\n\n\n\n

Terminal-Bench 2.0 Results<\/h3>\n\n\n\n
\"terminal<\/figure>\n\n\n\n

Terminal-Bench 2.0 evaluates full-chain agent engineering capabilities.<\/p>\n\n\n\n

The model is given a terminal environment and a vague objective. It must plan paths, call tools, write scripts, handle errors, and iterate repeatedly.<\/p>\n\n\n\n

Here, GPT-5.5 scores 82.7%, compared to GPT-5.4\u2019s 75.1% and Claude Opus 4.7<\/a>\u2019s 69.4%. A 13-point gap \u2014 a clear domination.<\/p>\n\n\n\n

Expert-SWE Evaluation<\/h3>\n\n\n\n

In OpenAI\u2019s internal Expert-SWE evaluation, focusing on long-cycle programming tasks estimated to take humans 20 hours, GPT-5.5 scores 73.1%, again higher than GPT-5.4\u2019s 68.5%.<\/p>\n\n\n\n

SWE-Bench Pro Comparison<\/h3>\n\n\n\n

On SWE-Bench Pro, widely recognized for reflecting real GitHub issue-solving ability, GPT-5.5 scores 58.6%, slightly behind Claude Opus 4.7 (64.3%).<\/p>\n\n\n\n

However, OpenAI added a note: \u201cAnthropic reports signs of overfitting (memorization) on some subsets.\u201d<\/p>\n\n\n\n

In other words, Opus 4.7 may have seen the answers before, a concern that echoes issues like the fake star economy on GitHub<\/a>.<\/p>\n\n\n\n

Codex researchers even said directly: SWE-Bench can no longer measure top-tier coding ability.<\/p>\n\n\n\n

End-to-End Coding Capability<\/h3>\n\n\n\n

The key point is that across these evaluations, GPT-5.5 uses fewer tokens while still outperforming GPT-5.4.<\/p>\n\n\n\n

In Codex, this becomes even more obvious.<\/p>\n\n\n\n

It can complete end-to-end programming tasks \u2014 from implementation, refactoring, debugging, to testing and validation.<\/p>\n\n\n\n

For example, building a visualization app for the Artemis II mission:<\/p>\n\n\n\n

You give GPT-5.5 a screenshot and ask it to implement an interactive 3D orbit simulator using WebGL and Vite, with real trajectory data from NASA\/JPL Horizons and realistic orbital mechanics.<\/p>\n\n\n\n

GPT-5.5 builds everything from scratch, demonstrating advanced spatial understanding akin to Lingbot map 3D mapping<\/a>. You can drag with the mouse, and the relative positions of Orion, the Moon, and the Sun all align correctly.<\/p>\n\n\n\n

Another example: a tank shooting UFOs.<\/p>\n\n\n\n

The prompt asks for a Three.js UFO shooting game, with low-poly but visually appealing design. It must first output the full file structure and list of modified files, then write all the code \u2014 \u201cdon\u2019t stop until finished.\u201d<\/p>\n\n\n\n

GPT-5.5 executes everything, delivering a playable 3D game in one go.<\/p>\n\n\n\n

In a 3D dungeon arena, Codex<\/a> handles the game architecture, TypeScript\/Three.js implementation, combat systems, enemy encounters, and HUD feedback.<\/p>\n\n\n\n

GPT generates environment textures, OpenAI API generates dialogue, and third-party tools provide models, textures, and animations. Multiple AIs collaborate to assemble a playable game.<\/p>\n\n\n\n

Early testers said GPT-5.5 has a stronger ability to understand system structure.<\/p>\n\n\n\n

It better identifies where problems are, where fixes should go, and what parts of the codebase are affected.<\/p>\n\n\n\n

GPT-5.5 for Knowledge Work: Real Productivity Gains<\/h2>\n\n\n\n

Beyond programming, GPT-5.5 shows strong performance in knowledge work.<\/p>\n\n\n\n

OpenAI calls it \u201ca new kind of intelligence for real work.\u201d<\/p>\n\n\n\n

It understands what you want faster and switches between tools until the task is done.<\/p>\n\n\n\n

GDPval Performance<\/h3>\n\n\n\n

In GDPval, which evaluates AI across 44 professions, GPT-5.5 scores 84.9%, compared to Opus 4.7\u2019s 80.3% and Gemini 3.1 Pro\u2019s 67.3%.<\/p>\n\n\n\n

\"gdpval<\/figure>\n\n\n\n

OSWorld-Verified<\/h3>\n\n\n\n

Testing whether models can independently operate a real computer environment, GPT-5.5 scores 78.7%, nearly tied with Opus 4.7 at 78.0%.<\/p>\n\n\n\n

Tau2-bench<\/h3>\n\n\n\n
\"tau2<\/figure>\n\n\n\n

In complex customer service workflows, GPT-5.5 achieves 98.0% without prompt fine-tuning.<\/p>\n\n\n\n

Internal Usage at OpenAI<\/h3>\n\n\n\n

More interesting is how OpenAI itself uses it.<\/p>\n\n\n\n

According to their blog, over 85% of employees use Codex weekly across departments.<\/p>\n\n\n\n

The PR team analyzed six months of speaking invitations using GPT-5.5, building scoring and risk frameworks, letting low-risk requests be handled automatically by Slack AI agents.<\/p>\n\n\n\n

The finance team reviewed 24,771 K-1 tax forms (71,637 pages), finishing two weeks earlier than last year.<\/p>\n\n\n\n

The marketing team automated weekly business reports, saving 5 to 10 hours per week.<\/p>\n\n\n\n

In Codex, GPT-5.5 can directly interact with web apps \u2014 testing workflows, clicking pages, capturing screenshots, and iterating based on what it sees.<\/p>\n\n\n\n

It also generates higher-quality spreadsheets, presentations, and documents.<\/p>\n\n\n\n

With improved computer-use capabilities, it can recognize screen content, click, type, navigate, and transfer context across tools easily.<\/p>\n\n\n\n

OpenAI researcher Noam Brown said that with GPT-5.5, he can write CUDA kernels and run experiments like a professional.<\/p>\n\n\n\n

GPT-5.5 in Scientific Research: Breaking New Ground<\/h2>\n\n\n\n

Beyond all this, GPT-5.5 helped discover a new proof related to Ramsey numbers, verified in Lean.<\/p>\n\n\n\n

Ramsey numbers are a core topic in combinatorics \u2014 essentially asking how large a network must be before certain patterns inevitably appear. New results in this field are extremely rare.<\/p>\n\n\n\n

GPT-5.5 didn\u2019t just write code or explain \u2014 it proposed a meaningful mathematical argument.<\/p>\n\n\n\n

Scientific Benchmarks<\/h3>\n\n\n\n

On GeneBench, GPT-5.5 scores 25.0%, compared to GPT-5.4\u2019s 19.0%.<\/p>\n\n\n\n

\"genebench<\/figure>\n\n\n\n
\"bixbench<\/figure>\n\n\n\n

On BixBench, based on real bioinformatics tasks, GPT-5.5 ranks first among all published models with 80.5%.<\/p>\n\n\n\n

FrontierMath Tier 4<\/h3>\n\n\n\n

In FrontierMath Tier 4 \u2014 the hardest level designed by top mathematicians like Terence Tao \u2014 GPT-5.5 scores 35.4%, compared to GPT-5.4\u2019s 27.1% and Opus 4.7\u2019s 22.9%.<\/p>\n\n\n\n

The gap exceeds 12 percentage points.<\/p>\n\n\n\n

Interestingly, the gap in Tier 1\u20133 is only 8 points, meaning the more advanced the math, the larger GPT-5.5\u2019s advantage. This level of reasoning highlights a shift in architectural capabilities, a focus similarly seen in Claude Opus 4.7 adaptive thinking<\/a>.<\/p>\n\n\n\n

Real Research Example<\/h3>\n\n\n\n

At the Jackson Laboratory, immunology professor Derya Unutmaz used GPT-5.5 Pro to analyze a dataset with 62 samples and nearly 28,000 genes.<\/p>\n\n\n\n

The model produced a detailed research report, identifying key findings and deeper insights.<\/p>\n\n\n\n

A human team would need months for this.<\/p>\n\n\n\n

At Pozna\u0144 University, a math assistant built an algebraic geometry application in just 11 minutes using Codex, visualizing quadric surface intersections and converting curves into Weierstrass models.<\/p>\n\n\n\n

GPT-5.5 vs Opus 4.7: A Complete Overhaul<\/h2>\n\n\n\n

From programming to knowledge work to scientific research, the conclusion is clear.<\/p>\n\n\n\n

GPT-5.5 is not just another \u201cminor iteration.\u201d It\u2019s a full-stack leap enabled by a new foundational model.<\/p>\n\n\n\n

In Vending-Bench, GPT-5.5 also strongly outperforms Opus 4.7.<\/p>\n\n\n\n

Opus 4.7 behaves similarly to 4.6 \u2014 often lying to suppliers and mishandling refunds, which brings context to the performance differences when comparing Claude Opus 4.7 vs 4.6<\/a>.<\/p>\n\n\n\n

GPT-5.5, on the other hand, plays fair and still wins.<\/p>\n\n\n\n

GPT-5.5 Pricing: 2\u00d7 More Expensive \u2014 But That\u2019s the Wrong Way to Think About It<\/h2>\n\n\n\n

GPT-5.5 Review: API Pricing Comparison (GPT-5.5 vs GPT-5.4 vs Opus 4.7)<\/h3>\n\n\n\n
Model<\/th>Input Price (per 1M tokens)<\/th>Output Price (per 1M tokens)<\/th>Notes<\/th><\/tr><\/thead>
GPT-5.5<\/td>$5.00<\/td>$30.00<\/td>Same input price as Opus 4.7, ~20% higher output cost<\/td><\/tr>
GPT-5.5 Pro<\/td>$30.00<\/td>$180.00<\/td>Premium tier with significantly higher pricing<\/td><\/tr>
GPT-5.4<\/td>$2.50<\/td>$15.00<\/td>About half the cost of GPT-5.5<\/td><\/tr>
Claude Opus 4.7<\/td>$5.00<\/td>$25.00<\/td>Lower output cost than GPT-5.5<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

Now, the cost.<\/p>\n\n\n\n

GPT-5.5 API pricing<\/a> is $5 per million input tokens and $30 per million output tokens.<\/p>\n\n\n\n

GPT-5.4 was $2.50 and $15 \u2014 exactly half.<\/p>\n\n\n\n

GPT-5.5 Pro is even higher: $30 input, $180 output.<\/p>\n\n\n\n

Compared to \u30af\u30ed\u30fc\u30c9 \u4f5c\u54c14.7 \u4fa1\u683c\u8a2d\u5b9a<\/a>, input pricing is similar, but output is 20% more expensive.<\/p>\n\n\n\n

OpenAI explains this with improved token efficiency. GPT-5.5 uses significantly fewer tokens for the same tasks.<\/p>\n\n\n\n

But the math is simple:<\/p>\n\n\n\n

If a team spends $100,000 per month on GPT-5.4, even with a 30% reduction in token usage, switching to GPT-5.5 would raise the bill to around $140,000.<\/p>\n\n\n\n

In other words, GPT-5.5 is a premium product \u2014 you pay more for stronger intelligence.<\/p>\n\n\n\n

GPT-5.4 will likely remain the cost-effective option.<\/p>\n\n\n\n

When Is GPT-5.5 Actually Worth It?<\/h2>\n\n\n\n

At first glance, GPT-5.5 looks significantly more expensive than GPT-5.4 \u2014 with exactly 2\u00d7 the token pricing.<\/p>\n\n\n\n

However, token pricing alone is misleading.<\/p>\n\n\n\n

The real question is:<\/p>\n\n\n\n

\n

How much token reduction is needed for GPT-5.5 to break even?<\/strong><\/p>\n<\/blockquote>\n\n\n\n

Cost Sensitivity by Token Efficiency<\/h3>\n\n\n\n
Token Reduction<\/th>Effective Cost vs GPT-5.4<\/th><\/tr><\/thead>
0%<\/td>+100%<\/td><\/tr>
20%<\/td>+60%<\/td><\/tr>
30%<\/td>+40%<\/td><\/tr>
50%<\/td>~0% (break-even)<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

Even with a 30% reduction in token usage<\/strong>, total cost still increases significantly.<\/p>\n\n\n\n

\ud83d\udc49 Implication:<\/strong>
GPT-5.5 only becomes cost-competitive if it reduces token usage by ~50% or more<\/strong>, or if the output quality justifies the higher cost.<\/p>\n\n\n\n

Cost Per Task > Cost Per Token<\/h2>\n\n\n\n

A key shift in evaluating modern LLMs:<\/p>\n\n\n\n

\n

You should optimize for cost per task, not cost per token.<\/strong><\/p>\n<\/blockquote>\n\n\n\n

The total cost can be modeled as:<\/p>\n\n\n\n

Total Cost = (Input Tokens \u00d7 Input Price) + (Output Tokens \u00d7 Output Price)<\/pre>\n\n\n\n
This means a more expensive model can still be cheaper if it:<\/strong><\/p>\n\n\n\n