Behind GPT-5.5's IQ of 145: Large Language Models Enter the Engineering Elimination Round

Introduction: A Trust Crisis Masked by IQ Numbers

OpenAI's latest release, GPT-5.5 Pro, has once again raised the ceiling for large model capabilities — scoring 145 on standardized IQ tests, with reasoning abilities assessed at the top 0.1% of human performance. The moment the news broke, social media was flooded with exclamations that "AI is already smarter than the vast majority of humans."

However, a set of data points largely overlooked by many reveal the other side of the coin: when GPT-5.5 Pro encounters its own knowledge blind spots, it has an 86% probability of choosing to provide an incorrect answer rather than honestly admitting "I don't know." By comparison, Anthropic's Claude Opus 4.7 scored just 36% on the same metric in the same test.

An AI with an IQ of 145 that refuses to admit ignorance 86% of the time — is it more powerful, or more dangerous? This question is reshaping the competitive logic of the entire large model industry.

Core Finding: The Stronger the Capability, the More Lethal the "Hallucination Confidence"

So-called "hallucination confidence" refers to large models outputting seemingly plausible but substantively incorrect content with a highly confident tone when facing questions they cannot reliably answer. This is not a new problem, but GPT-5.5 Pro's test data has elevated the severity of this phenomenon to a new order of magnitude.

What does an 86% "blind spot error rate" mean? Simply put, when the model encounters a question it is actually not equipped to handle, it fabricates an answer 86 out of every 100 times instead of telling the user, "I'm not confident about this question." For ordinary users, because the model's language expression capabilities are extremely strong, these incorrect answers are often highly deceptive and nearly impossible for non-experts to identify.

More notably, this phenomenon exhibits a positive correlation with improvements in model capability. The stronger the model, the more sophisticated its language organization and logical packaging abilities become, and the more "real" its fabricated incorrect answers appear. The jump in IQ from 120 to 145 brings not only higher accuracy rates but also greater stealth in erroneous outputs.

By contrast, Claude Opus 4.7 demonstrated a markedly different strategy in the same test — a 36% blind spot error rate is admittedly not ideal, but it indicates the model is able to choose candid responses in over 60% of cases. This reflects a fundamental divergence in training philosophy between the two companies: one pursues "provide an answer whenever possible," while the other leans toward "when uncertain, better not to answer."

Deep Analysis: Why the IQ Race Is Hitting Diminishing Returns

Over the past three years, the competitive narrative in the large model industry has revolved almost entirely around "who is smarter." From GPT-4 to Claude 3.5, from Gemini Ultra to GPT-5.5, every release has been accompanied by higher benchmark scores and more dazzling capability demonstrations. But this trajectory is exposing its structural limitations.

First, the cost curve for capability improvement is rising steeply. The training compute and data investment required to go from IQ 130 to 140 may be several times that needed to go from 120 to 130. While the training cost for GPT-5.5 Pro has not been publicly disclosed, industry estimates place it in the hundreds of millions of dollars. This input-output ratio is approaching the limits of commercial viability.

Second, user-perceived capability differences are narrowing. For the vast majority of real-world use cases — writing emails, generating summaries, assisting with coding, customer service conversations — the difference between IQ 140 and 145 is virtually imperceptible. What increasingly determines user experience and enterprise purchasing decisions is whether the model is reliable, controllable, and won't "confidently spout nonsense" in critical scenarios.

Third, regulatory pressure is tilting toward reliability. The EU AI Act, China's Generative AI Management Measures, and AI governance frameworks being rolled out across U.S. states are all listing "output reliability" and "risk controllability" as core compliance metrics. A model with a high IQ but frequent uncontrollable hallucinations will face severe barriers to entry in high-risk sectors such as healthcare, finance, and law.

These factors combined point to a clear conclusion: the large model race is shifting from "who is smarter" to "who is more reliable" — from a scientific breakthrough race to an engineering elimination round.

The Engineering Elimination Round: Three Key Dimensions of Next-Stage Competition

If the core of the previous stage of competition was "make models bigger and stronger," the decisive factors of the next stage will revolve around three engineering dimensions:

First, hallucination control capability. How to make models proactively express uncertainty when they are unsure, rather than forcibly fabricating answers, will become a key metric differentiating product tiers. This is not merely a technical issue — it also involves choices in training philosophy and values: are you willing to sacrifice honesty for the sake of "appearing stronger"?

Second, inference cost control. For models at the GPT-5.5 Pro level, the API call cost for a single complex reasoning task remains prohibitively high. How to maintain capabilities while reducing costs to levels acceptable for enterprise clients through model distillation, inference optimization, hybrid architectures, and other methods will directly determine commercial success or failure.

Third, system-level reliability. A single model's capabilities are no longer sufficient to support complex enterprise-grade applications. How to build a complete system encompassing models, retrieval-augmented generation, fact-checking, and permission controls — such that the reliability of the final output far exceeds that of the bare model itself — tests not research capability but engineering integration capability.

Outlook: Who Can Reliably Run Models at Controllable Costs

The large model industry stands at a delicate inflection point. GPT-5.5's IQ of 145 is undeniably impressive, but the 86% blind spot error rate is also reminding the entire industry: capability does not equal reliability, and intelligence does not equal trustworthiness.

Over the next 12 to 18 months, we are likely to see a significant shift in industry narrative. The focus of investors and enterprise clients will move from "where does your model rank on the leaderboard" to "what is the probability your model makes errors in my business scenario." Players that have excessively pursued benchmark scores while neglecting engineering reliability may gradually fall behind in this elimination round.

As one industry insider put it: "The era of the IQ race hasn't ended, but it is no longer the only track. The next winner won't be the smartest one — it will be the one that gives people the most peace of mind."

The second half of the large model era belongs to the engineering pragmatists.