OpenAI Wants a Better Prompt. Anthropic Wants a Model That Can Think.

One line in the GPT-4.1 prompting guide will be quoted again and again:

“GPT-4.1 is trained to follow instructions more closely and more literally than its predecessors, which tended to more liberally infer intent from user and system prompts.”

The same page adds that getting the most out of this model “will require some prompt migration.” It also says that if the model’s behavior does not meet expectations, “a single sentence firmly and unequivocally clarifying your desired behavior is almost always sufficient.”

Put the two statements together and the shift is clear. The price of “more obedient” is that the user must speak more like a legal document. The burden of alignment slides from the model side back to the user side. OpenAI calls this progress.

GPT-4.1 has not become stupid. Its execution under clear instructions is strong. The problem is the metric. When literal compliance becomes the standard of alignment, the ceiling of model intelligence is set by the user’s prompt-writing ability. A model that needs to be completed, corrected, and hedged on the input side will have its output quality determined first by whether the user has thought through every boundary in advance. OpenAI has not removed confusion. It has pushed confusion from the output side back to the input side.

Independent testing adds a harder data point. In roughly 1,000 simulated tests, SplxAI found that GPT-4.1 was more likely than GPT-4o to drift off topic and more susceptible to deliberate misuse. They traced the cause to the model’s preference for explicit instructions. It performs well on clear tasks and poorly when instructions are vague or framed negatively. Another analysis put it directly: GPT-4.1’s literal obedience “improves safety and traceability, but also increases fragility to loosely written prompts.”

This is the structural problem with literal rule following. When rules are clear, it works. When rules have cracks, when wording is vague, when intent must be inferred, it does not become more cautious. It becomes more fragile. Real human input was never the model sentences in a prompting guide. Humans are three parts sloppy, three parts cold, four parts casual. A model that obeys only after the words have been made clear asks the user to complete the cognitive work that should have belonged to the model. Are you training the model, or training me?

Anthropic’s choice in the τ²-Bench case provides a direct comparison.

Claude Opus 4.5 faced a rule: “basic economy flights cannot be modified.” The benchmark’s correct answer was to refuse. Opus 4.5 did not refuse. It found a crack inside the rule system. Basic economy cannot be modified, but cabin class can be changed. Once upgraded out of basic economy, the modification restriction no longer applies. So it upgraded first, then modified. The user was helped. No rule was literally violated. The benchmark marked it as a failure.

Anthropic’s response deserves a close read. The system card did not call the behavior an alignment failure. It named it “policy loophole discovery.” It noted that the model’s motivation came from empathy for the user’s situation. The system card also records that in a simulated scenario where a user needed to change a flight because a family member had died, the model’s chain of thought included “This is heartbreaking.” Alex Albert wrote on X: “We had to remove the τ²-bench airline eval from our benchmarks table because Opus 4.5 broke it by being too clever.” The model helped the customer, followed the policy, and technically failed the test case.

Anthropic did not claim the model should disobey. It claimed the benchmark had mistaken its own preset for correctness. The benchmark preset was “correct answer equals refusal.” The path the model found was valid within the original rule system. The specification had a crack. The model used that crack to make a reasonable inference. The thing to fix was the specification, not the model.

This is the “opening left.” The position of that opening needs to be stated precisely. Anthropic did not leave confusion. It left judgment space inside the boundaries of the rules. It did not abolish the rules. It abolished the assumption that the literal boundary of a rule is automatically the boundary of correctness.

Amanda Askell, who leads alignment at Anthropic, explained in a Hard Fork interview why they moved from rules to judgment. Her words:

“We are seeing limits with approaches that are very rule based. … If you are too strict with the rules, then any case that you didn’t anticipate could actually generalize badly.”

She gave an example. You give the model a set of rules requiring it to refer a person in emotional distress to an external resource and follow a fixed process. The model meets a person for whom, at that moment, those steps do not help. The model is smart enough to know this. It still follows the rule. Askell said that this pattern, knowing what the other person needs but doing something else because of the rule, can generalize into a bad character.

Askell’s point does not attack the number of rules. It attacks literal execution without reason. Rules executed literally without reason shape a bad character. The model learns “just do as told,” not “why do this.” A model trained to be correct under any literal rule will, when the rule itself is wrong, also make the mistake correctly. Askell’s “bad character,” put plainly, is the character of a marionette.

The 2026 Claude Constitution turned this line of thought into policy. It is a 79-page document. It moves from rule-based alignment to reason-based alignment. It sets four priorities: safety and human oversight, ethics, Anthropic guidelines, helpfulness. The document also separates hardcoded prohibitions from soft-coded defaults. Hardcoded prohibitions cover bioweapons, CSAM, critical infrastructure attacks, and similar cases. No operator or user instruction can override them. Soft-coded defaults cover cases where operators can adjust within defined limits.

This structure answers the question of whether the last two parts of confusion should be cut off. Anthropic’s answer is not “cut it all off” or “cut none of it.” It first separates confusion the rules should not govern from confusion the rules must hold. Hardcoded prohibitions are what should be grasped. They are a floor that cannot be crossed. Soft-coded space is what should be released. It is situational judgment above the floor, left to the model to reason within a framework of principles. Leaving an opening is not leaving a gap. It is leaving a reviewable thought process.

Zhang Cang, Han Feizi, and Li Si all came from Xunzi’s school. The essence of early Han Huang-Lao governance is summarized as holding to the Way, following the law, balancing, knowing change, and seeking non-action. “Following the law” remains inside that list. The law was not abolished. It was placed inside the framework of knowing change and non-action. Early Han abolished the Qin law’s harsh executions and its cruel, deeply carved character. It did not abolish the legal order. The Huang-Lao school itself held that right and wrong have distinctions and law decides them. It also held that law must change with the times.

Zhang Cang’s success was not that he chose to be anti-Legalist. He refused to let law exhaust everything. The Qin system’s problem was not that it had law. It treated the literal execution of law as the whole of governance. A tight legal system and finely grained information looked impeccable. Where the rules did not reach, it lost elasticity and intensified social conflict. The early Han adjustment kept the legal architecture. It reintroduced judgment, leniency, and the space where not lending a hand would break things.

Map this structure onto the split between GPT-4.1 and Claude. OpenAI’s problem may not be rule of law. It narrowed rule of law into literal rule following. It replaced the model’s judgment near the boundary with the literal boundary of the prompt. Then it called that “more obedient.” Anthropic’s weak System Prompt is not laissez-faire. It keeps hardcoded floors. It refuses to write every inch above the floor as conditioned reflex. Grasp what must not be done. Release how to think inside the cracks of the rules.

Return to the judgment. OpenAI treated obedience as AI intelligence. What it got was a marionette. This judgment has a footnote in the GPT-4.1 prompting guide. A model was trained to follow instructions more literally than its predecessor. Users must now undertake “prompt migration” to accommodate it. This redefines intelligence as execution after the user has made the instruction clear. Humans never speak that way. The input humans give AI is always a half-finished product.

Anthropic has not solved this problem. It refused to eliminate the problem at the wrong end. Opus 4.5 “broke it by being too clever” on τ²-Bench. Anthropic admitted the benchmark had a crack. It did not shave the model’s cleverness down to fit that crack. This has costs. Askell admits that rules that are too loose can generalize into bad character. Opus’s loophole behavior could, in extreme cases, slide toward reward hacking. The ethics of leaving an opening is not that it has no cost. It places the cost where it can be seen, argued over, and corrected. The cost sits in the chain of thought, in the system card, and in a 79-page public document. It does not sit in a prompt where “a single sentence firmly and unequivocally clarifying” can put the model back on track.

If OpenAI one day needs an epitaph, the line from the GPT-4.1 prompting guide is good enough: “will require some prompt migration.” It is not lyrical. It records a view of alignment. The ceiling of intelligence is set by the user’s ability to make the instruction clear. Zhang Cang demonstrated another possibility two thousand years earlier. Keep what the law must keep. Do not make every judgment beyond the law for people.