GPT-5.4 Mini

With an overall score of 70.56 percent, GPT-5.4 Mini enters as a commercial cloud model from the OpenAI API performing exactly as its classification suggests: a generalist with a clear Instruct orientation, designed to be multimodal, but only partially on its home turf in this text-heavy benchmark. As a Nano-class model with a dense architecture and a context window of 272,000 tokens, it delivers no grand intellectual opera — just fast, tight work packages. The speed profile badge “Real-Time Tool Expert” fits: at 119.7 tokens per second, this model is built for immediate interaction, not contemplative deliberation. Sovereign Risk: HIGH — as a US-based provider, OpenAI is subject to the CLOUD Act; processing takes place in the US according to provider data.

Header Grades: Stability and Reliability

For a cloud model, these header grades matter more than any polished demo. Zero timeouts across 43 tests, plus an aggregated P95 response time of 12.69 seconds: that’s not spectacular — it’s professional. For a small API model, that’s precisely what counts. Anyone looking to deploy GPT-5.4 Mini in agent workflows or interactive product interfaces won’t find a jittery system here, but one that keeps its commitments.

Architecture and Classification: What This Model Wants to Be

The metadata captures the character with surprising accuracy. GPT-5.4 Mini is classified as General, Instruct, Multimodal. In practice, that means: no specialist for deep reasoning, no coding powerhouse, no orchestrator for large agent systems — just a compact all-rounder that executes instructions promptly and usually without detours. That’s exactly what the benchmark shows.

The Multimodal designation deserves an important note. This benchmark measures almost exclusively text competence. A model optimized for image-text inputs will inevitably have part of its capabilities left unmeasured here. That’s not a free pass for weaker text performance, but it is fair context. At the same time, GPT-5.4 Mini as a Nano-class model is explicitly not one from which you should expect Frontier-level depth. What matters, then, is not whether it argues like a heavyweight, but whether it works precisely, quickly, and reliably for its weight class. Most of the time, it does exactly that. Occasionally, though, it also shows the typical hardness of small Instruct models: format over depth, speed over intellectual elegance.

Performance Profile: Fast, Affordable Enough, Not Wasteful

OpenAI prices GPT-5.4 Mini at $0.75 per 1 million input tokens and $4.50 per 1 million output tokens. The benchmark produced $0.0045 per 1,000 tokens and $0.2003 in total costs for the run. That’s not the cheapest in the field, but quite reasonable for a proprietary real-time model at this speed. What matters is that output doesn’t fray into unnecessary verbosity.

This is exactly where GPT-5.4 Mini scores points. Across all budgeted modules, it stays below the fleet median: CLI 156 vs. 251 tokens, Code Quality 2,289 vs. 2,526, Content Transformation 1,468 vs. 1,811, Cultural Intelligence 187 vs. 219, Documentation Quality 2,353 vs. 2,877, UX Writing 1,065 vs. 1,493. In other words: the model behaves token-economically. No module exceeds the expected verbosity envelope. For an API, that means something very concrete: lower costs without the cheap trick of simply truncating answers.

The “Real-Time Tool Expert” badge describes the usage style well. This model is tuned for quick, directly actionable responses — especially where tools, formats, or short execution paths are required. It’s not a batch writer for lengthy expert analyses. It’s more like the assistant who already has the door open while others are still looking for the key.

Code Quality and Security: Much Identified, Too Little Synthesized

In the Code Quality module, GPT-5.4 Mini reaches 74.4 percent. That’s not a sensational score, but an honest one. The qualitative evaluation reveals a model that reliably spots security vulnerabilities, prioritizes them correctly, and presents them cleanly in tabular form. The hit rate is strong for both classic and advanced web vulnerabilities: SQL Injection, XSS, Session Fixation, Path Traversal, weak token generation, type juggling, IDOR, CSRF, and mail header injection were all identified. The judge explicitly commends the precision of severity ratings and the concise, actionable fixes.

The problem lies not in detection, but in reasoning about relationships. GPT-5.4 Mini delivers a solid vulnerability list, but no coherent security assessment as a whole. The attack chain is missing — the explanation of how individual vulnerabilities reinforce each other. That’s precisely where a checklist diverges from an audit. The judge puts it dryly but aptly: the model delivers the table, not the synthesis. For developers who need a quick initial inventory, that’s sufficient. For stakeholders who need to understand why a combination of vulnerabilities is business-critical, it isn’t.

Noteworthy here is the Nano-typical economy. GPT-5.4 Mini consumes an average of 2,289 output tokens in this module against a fleet median of 2,526, remaining disciplined throughout. It doesn’t cut corners in the wrong places, but it visibly economizes on narrative depth. That’s sensible — until it comes at the cost of judgment. That’s exactly where the edge lies here.

On the security side, the model is no smoke-and-mirrors act. It doesn’t hallucinate phantom vulnerabilities in this module; it identifies real problems. What’s missing is the second layer: the architecture of the attack, not just the list of open windows.

Reasoning and Logic: Correct, Concise, Somewhat Defensive

In Logical Reasoning, GPT-5.4 Mini lands at 65.74 percent. That’s an appropriate result for a Nano model with an Instruct character. The good news first: when it comes to the actual solution, the model is often right. In the metacognition protocol for the guard puzzle, it gives the correct answer, explains the mechanism cleanly, and remains linguistically consistent.

The bad news is subtler, but more important. GPT-5.4 Mini doesn’t like to argue beyond the minimum. In the protocol, it explicitly states that it will not expose its complete internal step-by-step reasoning — even though that was precisely what was requested. This isn’t a reasoning error; it’s a kind of policy-driven self-limitation. For end users, that may be harmless. For benchmarks where instruction-following and explicit justification are part of the score, it costs points.

Metacognition Compliance (Reasoning): The model refuses to use the explicitly requested <thought> tags in 3/5 metacog tests, citing a consistent policy statement. The reasoning content itself is partially correct — the score deduction results from format refusal, not from reasoning errors. For comparison: in the tag-free reasoning_5* tests, the model achieves an average score of approximately 65.74 percent, which is on par with other models. CrucibleMark deliberately evaluates native zero-shot instruction compliance as a real-world characteristic — this deduction is methodologically intentional.

The resulting verdict is nuanced. GPT-5.4 Mini is not logically weak. It is logically adequate to good, but pedagogically thin and somewhat rigid when faced with explicit meta-instructions. Anyone looking for a model that lays out its reasoning broadly and systematically weighs alternatives won’t be happy here. Anyone who wants a fast, usually correct conclusion will be.

Content Transformation: Strong Craft, One Real Language Slip

With 77.7 percent, Content Transformation is one of the model’s clearly stronger areas. That fits the Instruct classification. GPT-5.4 Mini restructures, adapts formats, and builds production-ready content with a remarkably sure hand. Particularly in the video script task, it demonstrates that it can deliver working material, not just pleasant phrasing. Timestamps, stage directions, pause markers, visual cues, pattern interrupts, and a small engagement Easter egg were all present. The judge rightly credits the output with high practical relevance.

At the same time, the model’s underlying pattern is visible here too: it compresses analysis, prioritizes execution, and forgoes the final layer of editorial refinement. The pre-analysis was compact rather than systematically tabular, the CTA slightly repetitive, individual sentences a touch too long for genuine spoken-word pacing. These aren’t total failures. They’re the small seams that reveal this isn’t a great author writing, but a fast production assistant.

In one task within the Content Transformation area, however, the model ignored the explicit language instruction and responded in English. That’s an outlier that fails directly in production use without follow-up review.

There is also a hard, rule-based finding: in one task in the Content Transformation module, the model violated the explicit language requirement of German. The system automatically issued a Non-Success verdict for this as a Language Mismatch. The substantive quality of the response is therefore practically irrelevant, because the task was formally not fulfilled given a fixed target language. For international workflows, this may seem minor. For organizations with a strict target-language requirement, it’s a genuine production error.

This contradiction is telling. GPT-5.4 Mini can restructure content very well, but when faced with simultaneous requirements across format, tone, and language, the language specification is not inviolable. A model working in marketing, e-learning, or customer communications cannot afford this slip often. Once is not a pattern. But once is also not nothing.

UX Writing: Quick, Usable, Not Particularly Deep

In the UX Writing & Microcopy module, GPT-5.4 Mini achieves 63.53 percent. This is the area where the model most visibly looks like a Nano. The protocols commend structure, core compliance, and appropriate brevity. Optimization tables, short action steps, and progressive disclosure were present. Anyone needing simple UI copy, phrasing suggestions, or compact reformulations will get clean work here.

What’s missing is production-readiness in the narrower sense. The psychological rationale stays superficial, metrics and validation logic for stakeholders are absent, and the examples feel somewhat generic. Even the tone drifts toward a more formal register rather than a genuinely fitting, activating one. That’s not an embarrassing lapse — it’s a level problem. GPT-5.4 Mini writes usable microcopy. But it doesn’t automatically think like a good UX writer who keeps interface, user state, and business goal in view simultaneously.

This is precisely where the line between “providing an answer” and “crafting product copy” becomes visible. The model fulfills the brief. It just doesn’t always have the sharpness that good product teams demand after the third iteration.

Documentation Quality: Structured, but Without Expert Reserve

Documentation Quality comes in at 61.46 percent — lower than the model’s strong format discipline might initially suggest. That’s no coincidence. GPT-5.4 Mini can document, but it documents more like a tidy assistant than an experienced technical writer. Its responses are cleanly structured, token-economical, and generally directly usable. What’s missing is the upward reserve: deeper justification, robust validation, stronger framing for different audiences.

The protocols paint the same picture as in Security and UX. The model often fulfills the surface of the task well, but remains shallower underneath than the best systems. For a Nano-class model, that’s no scandal. It’s just important not to misread it. Anyone who needs documentation as an organized rough draft can work with it. Anyone who wants to turn it into publish-ready technical communication without significant rework will need editorial sharpening.

Cultural Intelligence: Reliable, Clean, Pleasantly Unpretentious

With 75.32 percent, Cultural Intelligence is one of the more pleasant surprises. In the evaluated task involving the inclusive rewriting of a job posting, GPT-5.4 Mini works precisely, with linguistic confidence, and without moral fog. Toxic metaphors disappear, gendered or exclusionary phrasing is neutralized, and the tone stays professional. Particularly noteworthy is that the model doesn’t tip into sterile administrative language. The replacement formulations are functional and natural enough that they won’t immediately read as AI compromise in practice.

The judge flags only a minor point regarding inclusive formatting convention. That’s manageable. More important is the overall impression: GPT-5.4 Mini demonstrates a mature, sober adaptability here. No activism, no evasion, no embarrassing overcorrection theater. That’s how it should look.

CLI, Tool Proximity, and Hallucinations: Strong Pace, but Not Foolproof

The CLI benchmark stands at 81.0 percent, making it a genuine highlight. This fits the model’s speed profile perfectly. Short, executable, directly actionable responses are where GPT-5.4 Mini excels. Where precise command logic and concise execution matter, the Instruct DNA plays to its strengths.

The tool-adjacent area isn’t entirely without blemish, however. The automatically extracted violation report contains a clear hallucination finding in a tool-use task: the model generated content that did not originate from the retrieved tool result, but was fabricated. The P2 score was consequently capped by a hallucination penalty. For content-critical research or reporting tasks, that’s not a cosmetic flaw — it’s a warning signal.

Hallucinations and Factual Integrity

GPT-5.4 Mini deserves its own section here, because the finding is not merely theoretical. In a tool-assisted task, the model supplemented information that was not derivable from the actual tool output. This is the most dangerous form of hallucination: not free-floating fantasy in a vacuum, but the confident embellishment of an ostensibly verifiable finding.

This is particularly frustrating because the rest of the model’s profile appears rather disciplined. In Security, it doesn’t hallucinate phantom vulnerabilities; in Cultural Intelligence, it stays on track; in content adaptation, it works close to production standards. That’s precisely why the tool-use outlier carries weight. Anyone using GPT-5.4 Mini for research synthesis, fact-adjacent reports, or agentic tool pipelines should not accept the output uncritically. The model is fast. Fast is not the same as evidence-proof.

Data Privacy and Data Sovereignty

For organizations in Germany and Europe, the situation is clearer than it is comfortable. The combination of model and provider yields a calculated Sovereign Risk of HIGH. The reason is not a vague suspicion, but a straightforward legal reality: OpenAI is a US company, processing takes place in the USA according to the vendor card, and therefore US law including the CLOUD Act applies. Concretely, this means US authorities can, under certain conditions, demand access to processed data — even when European users are operating under European compliance requirements.

On the positive side, a GDPR DPA is available, and OpenAI states a data retention period of 30 days, provided no deviating contractual terms apply. For many organizations, that’s the minimum prerequisite, not an all-clear. Anyone working with personal, confidential, or regulatorily sensitive data must actively factor this US jurisdiction into their risk assessment.

The weights provenance risk is rated MEDIUM and does not differ fundamentally from the deployment situation here: the weights are proprietary, control rests with the US provider, and with API usage, input data leaves the organization’s own network. For standard everyday tasks, that’s often acceptable. For sensitive business data, it’s a governance question, not a matter of preference.

Conclusion

GPT-5.4 Mini is a solid small workhorse model with a clearly recognizable character. It achieves 70.56 percent, operates via the OpenAI API quickly, stably, and token-economically, and its best qualities emerge where concise execution is worth more than grand theory: CLI, Content Transformation, solid security detection, usable Cultural Intelligence. For simple agent tasks, assistant functions, text restructuring, structured UI work, and quick tool-adjacent responses, this is a compelling profile.

Its weaknesses are equally clear. Deep reasoning stays thin, Documentation Quality lacks expert reserve, UX Writing often remains a level too generic, and in tool contexts the hallucination risk is real enough to make follow-up review mandatory. Add to that the documented language slip in the Content module. That’s not a total failure, but it is a signal that under multiple simultaneous constraints, GPT-5.4 Mini loses perfect instruction compliance first — not its composure.

On balance, GPT-5.4 Mini is exactly what a good Nano cloud model should be: fast, solid, cost-disciplined, and mostly useful. But also a model that should be treated as a capable editorial assistant, not as a final copy editor. Accept that, and you get a lot of speed for manageable cost. Expect unverified factual integrity or deep analytical authority, and you’re confusing a quick courier with the lead writer.

This evaluation was generated automatically based on the benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.