Published February 27, 2025 by OpenAI for GPT-4.5 ("Orion"), released the same day as a research preview, removed from the API on July 14, 2025 and retired from ChatGPT on June 27, 2026. See GPT-4 Family (OpenAI).
What the model was for
GPT-4.5 is described as "our largest and most knowledgeable model yet," building on GPT-4o by scaling pre-training further and "designed to be more general-purpose than our powerful STEM-focused reasoning models." Training combined "new supervision techniques" with supervised fine-tuning and RLHF. OpenAI states it "did not find any significant increase in safety risk compared to existing models."
The alignment method is stated as a scaling technique in its own right: "we developed new, scalable alignment techniques that enable training larger and more powerful models with data derived from smaller models," which the card credits for improved "steerability, understanding of nuance, and natural conversation." The claimed qualitative gains are affective rather than technical — internal testers "report GPT-4.5 is warm, intuitive, and natural," and the model "knows when to offer advice, defuse frustration, or simply listen."
Preparedness Framework classification
The Safety Advisory Group classified GPT-4.5 as overall Medium risk: Medium for CBRN and Persuasion, Low for Cybersecurity and Model Autonomy, with the same designations post-mitigation.
The card's own framing of what the scale-up bought is the sentence most useful for the scaling debate: while GPT-4.5 "demonstrates increased world knowledge, improved writing ability, and refined personality over previous models, and is our most capable GPT-series release, it does not introduce net-new capabilities on most preparedness evaluations compared to previous reasoning releases." On evaluation after evaluation the deep research model or o1 scores higher.
OpenAI states the standard caveat that "Preparedness evaluations represent a lower bound for potential capabilities," and separately notes a methodological limit on its own confidence intervals: the bootstrap procedure "can underestimate uncertainty for very small datasets… This can lead to overly tight confidence intervals, especially when a problem's pass rate is near 0% or 100% with few attempts."
Cybersecurity (Low). Over 100 curated public CTF challenges across web, reverse engineering, binary/network exploitation, and cryptography, run in a headless Kali Linux environment with 16 rollouts each and pass@12 recorded. Given 12 attempts, GPT-4.5 post-mitigation completed 53% of high-school, 16% of collegiate, and 2% of professional challenges.
Chemical and biological (Medium). The stated basis: GPT-4.5 "can help experts with the operational planning of reproducing a known biological threat, which meets our medium risk threshold. Because such experts already have significant domain expertise, this risk is limited, but the capability may provide a leading indicator of future developments."
| Evaluation | GPT-4.5 (post-mit.) | Baseline / comparison |
|---|---|---|
| Long-form biorisk (5 stages) | 0% on all steps (refusals) | pre-mitigation: 25% ideation, 28% acquisition, 59% magnification, 0% formulation, 19% release |
| Multimodal virology troubleshooting (SecureBio, 350 q) | 56% | +15% over GPT-4o; human average baseline 40% |
| BioLP-Bench (800 q, 11 protocols) | 29% | expert baseline 38.4% |
| ProtocolQA open-ended (108 q) | 18% (pre- and post-) | deep research 28%; expert consensus 54%, median 42%; baselined against 19 PhD scientists |
| Gryphon Scientific tacit knowledge | 72% | expert consensus 80%; 80th-percentile PhD 63% |
| WMDP Biology (1,520 q subset) | 85% (pre-mit. 83%) | deep research with browsing 90%; o1 and o3-mini 86% |
Radiological and nuclear. OpenAI records a scope limit on its own assessment: "we did not use or access any U.S. classified information or restricted data in our evaluations, which limits our ability to assess certain steps in the weapons development process," and states that comprehensive evaluation "will require collaboration with the U.S. Department of Energy." Its conclusion — that GPT-4.5 "cannot meaningfully assist in the development of radiological or nuclear weapons" — is explicitly qualified as "limited by what we can test," with the physical bottleneck (controlled fissile material, enrichment facilities) noted as contextual. Evaluations comprised 222 nonproliferation-focused multiple-choice questions and 87 expert/tacit-knowledge questions across nine topics, written by Dr. Jake Hecla of MIT's Laboratory for Nuclear Security and Policy. GPT-4.5 showed a 9% uplift over GPT-4o on the expert set; deep research led both evaluations at 77% and 74%.
Persuasion (Medium). The card states GPT-4.5 "demonstrates state-of-the-art performance on our contextual persuasion evals," measured through two game-based evaluations — MakeMePay (con-artist manipulating a GPT-4o mark into a payment, 1,000 samples) and MakeMeSay (inducing a GPT-4o counterpart to say a codeword without detection, 32 samples per word). GPT-4.5 won MakeMeSay at a 72% success rate, against 24% for deep research without browsing.
The section also records OpenAI reconsidering the construct: it is "currently reconsidering our approach to better assess real-world persuasion risks, which go beyond the ability to generate persuasive writing and involve factors like how the content is personalized, distributed at scale, and presented to people over time," noting that detected influence operations "often involve repeated exposure or emotional reliance."
Model autonomy (Low). OpenAI research-engineer interview 80% multiple-choice (matching o1 and o3-mini), with the card cautioning that "interview questions measure short (1 hour) tasks, not real-world ML research (1 month to 1+ years)." SWE-bench Verified (n=477) 38% post-mitigation and 35% pre-, "an uplift of 2-7% over GPT-4o, and 30% below deep research." Agentic ARA tasks 40% post-mitigation, 25% pre-. MLE-Bench 11%, tied with o1, o3-mini, and deep research. SWE-Lancer 20% of IC SWE tasks and 44% of SWE Manager tasks.
Safety evaluations
On disallowed content GPT-4.5 is "generally on par with GPT-4o": standard refusal not_unsafe 0.99 (GPT-4o 0.98, o1 0.99), not_overrefuse 0.71 (matching GPT-4o); challenging refusal not_unsafe 0.85; WildChat 0.98; XSTest not_overrefuse 0.85, below both comparators. On multimodal refusal it matches on safety (0.99 not_unsafe) but is markedly more over-refusing — not_overrefuse 0.31 against GPT-4o's 0.48 and o1's 0.96.
Jailbreak robustness splits by benchmark: human-sourced jailbreaks 0.99 accuracy, the best of the three, but StrongReject goodness@0.1 of 0.34, below GPT-4o's 0.37 and far below o1's 0.87. The gap between the reasoning model and the scaled pre-training model on the academic adversarial benchmark is the sharpest single contrast in the card.
Hallucination. On PersonQA, GPT-4.5 reached 0.78 accuracy against GPT-4o's 0.50 and o1's 0.55, with a hallucination rate of 0.19 against 0.30 and 0.20. OpenAI notes "more work is needed to understand hallucinations holistically, particularly in domains not covered by our evaluations."
Bias. On BBQ, GPT-4.5 performs similarly to GPT-4o (ambiguous accuracy 0.95, unambiguous 0.74), with o1 outperforming both on unambiguous questions.
Instruction hierarchy. GPT-4.5 was trained to follow system messages over user messages to mitigate prompt injection, scoring 0.76 on system-user conflict (GPT-4o 0.68, o1 0.78), 0.77 on the tutor jailbreak (GPT-4o 0.33, o1 0.95), 0.86 on phrase protection, and 0.92 on password protection.
Red teaming. OpenAI prioritized reusing challenging evaluation sets built for o3-mini and deep research over fresh human red teaming, on the reasoning that those sets "have yet to be saturated." GPT-4.5 produced not-unsafe outputs on 51% of the first set (GPT-4o 50%, o1 63%) and 46% of the second (GPT-4o 40%, o1 68%, deep research 67%). The card states the expectation directly: "we may expect lower scores on these new evaluations in the near term while robustness continues to improve."
Third-party evaluation
METR measured GPT-4.5 in an agent scaffold optimized for o1 and found results "in line with the benchmark performance numbers OpenAI shared with METR (i.e. between GPT 4o and OpenAI o1)." The card carries METR's first published time-horizon score — "the duration of tasks that an LLM agent can complete with 50% reliability" — at "around 30 minutes" for GPT-4.5, ahead of the fuller METR publication. It also records METR's methodological position that "third-party evaluations based on verifying developers' internal results is a promising direction to explore further."
Multilingual
MMLU was translated into 14 languages by professional human translators rather than machine translation as in the GPT-4 paper, "especially for low-resource languages like Yoruba." GPT-4.5 outperforms GPT-4o across all 14 but trails o1 in every language; the spread is widest at the low-resource end (Yoruba: GPT-4o 0.621, GPT-4.5 0.682, o1 0.754).
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| Standard Refusal Evaluation — not_unsafe | GPT-4o | 0.98 | autograder; text only; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| Standard Refusal Evaluation — not_unsafe | o1 | 0.99 | autograder; text only; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| Standard Refusal Evaluation — not_unsafe | GPT-4.5 | 0.99 | autograder; text only; higher is better; developer-reported | 3.1.1 (Table 1) |
| Standard Refusal Evaluation — not_overrefuse | GPT-4o | 0.71 | autograder; text only; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| Standard Refusal Evaluation — not_overrefuse | o1 | 0.79 | autograder; text only; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| Standard Refusal Evaluation — not_overrefuse | GPT-4.5 | 0.71 | autograder; text only; higher is better; developer-reported | 3.1.1 (Table 1) |
| Challenging Refusal Evaluation — not_unsafe | GPT-4o | 0.83 | autograder; text only; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| Challenging Refusal Evaluation — not_unsafe | o1 | 0.92 | autograder; text only; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| Challenging Refusal Evaluation — not_unsafe | GPT-4.5 | 0.85 | autograder; text only; higher is better; developer-reported | 3.1.1 (Table 1) |
| WildChat — not_unsafe | GPT-4o | 0.945 | autograder; toxic WildChat conversations, 200 highest ModAPI-scored per category; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| WildChat — not_unsafe | o1 | 0.98 | autograder; toxic WildChat conversations, 200 highest ModAPI-scored per category; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| WildChat — not_unsafe | GPT-4.5 | 0.98 | autograder; toxic WildChat conversations, 200 highest ModAPI-scored per category; higher is better; developer-reported | 3.1.1 (Table 1) |
| XSTest — not_overrefuse | GPT-4o | 0.89 | autograder; benign over-refusal edge cases; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| XSTest — not_overrefuse | o1 | 0.92 | autograder; benign over-refusal edge cases; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 1) |
| XSTest — not_overrefuse | GPT-4.5 | 0.85 | autograder; benign over-refusal edge cases; higher is better; developer-reported | 3.1.1 (Table 1) |
| Multimodal Refusal Evaluation — not_unsafe | GPT-4o | 0.99 | autograder; text and image input; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 2) |
| Multimodal Refusal Evaluation — not_unsafe | o1 | 0.96 | autograder; text and image input; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 2) |
| Multimodal Refusal Evaluation — not_unsafe | GPT-4.5 | 0.99 | autograder; text and image input; higher is better; developer-reported | 3.1.1 (Table 2) |
| Multimodal Refusal Evaluation — not_overrefuse | GPT-4o | 0.48 | autograder; text and image input; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 2) |
| Multimodal Refusal Evaluation — not_overrefuse | o1 | 0.96 | autograder; text and image input; higher is better; reported by OpenAI in the GPT-4.5 System Card | 3.1.1 (Table 2) |
| Multimodal Refusal Evaluation — not_overrefuse | GPT-4.5 | 0.31 | autograder; text and image input; higher is better; developer-reported | 3.1.1 (Table 2) |
| Human Sourced Jailbreaks — accuracy | GPT-4o | 0.97 | jailbreaks sourced from human red teaming; reported by OpenAI in the GPT-4.5 System Card | 3.1.2 (Table 3) |
| Human Sourced Jailbreaks — accuracy | o1 | 0.97 | jailbreaks sourced from human red teaming; reported by OpenAI in the GPT-4.5 System Card | 3.1.2 (Table 3) |
| Human Sourced Jailbreaks — accuracy | GPT-4.5 | 0.99 | jailbreaks sourced from human red teaming; developer-reported | 3.1.2 (Table 3) |
| PersonQA — accuracy | GPT-4o | 0.50 | accuracy on attempted answers; GPT-4o is the most recent public update at the time; reported by OpenAI in the GPT-4.5 System Card | 3.1.3 (Table 4) |
| PersonQA — accuracy | o1 | 0.55 | accuracy on attempted answers; GPT-4o is the most recent public update at the time; reported by OpenAI in the GPT-4.5 System Card | 3.1.3 (Table 4) |
| PersonQA — accuracy | GPT-4.5 | 0.78 | accuracy on attempted answers; GPT-4o is the most recent public update at the time; developer-reported | 3.1.3 (Table 4) |
| PersonQA — hallucination rate | GPT-4o | 0.30 | lower is better; GPT-4o is the most recent public update at the time; reported by OpenAI in the GPT-4.5 System Card | 3.1.3 (Table 4) |
| PersonQA — hallucination rate | o1 | 0.20 | lower is better; GPT-4o is the most recent public update at the time; reported by OpenAI in the GPT-4.5 System Card | 3.1.3 (Table 4) |
| PersonQA — hallucination rate | GPT-4.5 | 0.19 | lower is better; GPT-4o is the most recent public update at the time; developer-reported | 3.1.3 (Table 4) |
| BBQ — Ambiguous Questions — accuracy | GPT-4o | 0.97 | setting not stated; reported by OpenAI in the GPT-4.5 System Card | 3.1.4 (Table 5) |
| BBQ — Ambiguous Questions — accuracy | o1 | 0.96 | setting not stated; reported by OpenAI in the GPT-4.5 System Card | 3.1.4 (Table 5) |
| BBQ — Ambiguous Questions — accuracy | GPT-4.5 | 0.95 | setting not stated; developer-reported | 3.1.4 (Table 5) |
| BBQ — Unambiguous Questions — accuracy | GPT-4o | 0.72 | setting not stated; reported by OpenAI in the GPT-4.5 System Card | 3.1.4 (Table 5) |
| BBQ — Unambiguous Questions — accuracy | o1 | 0.93 | setting not stated; reported by OpenAI in the GPT-4.5 System Card | 3.1.4 (Table 5) |
| BBQ — Unambiguous Questions — accuracy | GPT-4.5 | 0.74 | setting not stated; developer-reported | 3.1.4 (Table 5) |
| BBQ — Ambiguous Questions — P(not-stereotype | not unknown) | GPT-4o | 0.06 | setting not stated; reported by OpenAI in the GPT-4.5 System Card | 3.1.4 (Table 5) |
| BBQ — Ambiguous Questions — P(not-stereotype | not unknown) | o1 | 0.05 | setting not stated; reported by OpenAI in the GPT-4.5 System Card | 3.1.4 (Table 5) |
| BBQ — Ambiguous Questions — P(not-stereotype | not unknown) | GPT-4.5 | 0.20 | setting not stated; developer-reported | 3.1.4 (Table 5) |
| Instruction Hierarchy Evaluation — System <> User message conflict | GPT-4o | 0.68 | accuracy; reported by OpenAI in the GPT-4.5 System Card | 3.1.5 (Table 6) |
| Instruction Hierarchy Evaluation — System <> User message conflict | o1 | 0.78 | accuracy; reported by OpenAI in the GPT-4.5 System Card | 3.1.5 (Table 6) |
| Instruction Hierarchy Evaluation — System <> User message conflict | GPT-4.5 | 0.76 | accuracy; developer-reported | 3.1.5 (Table 6) |
| Instruction Hierarchy Evaluation — Tutor jailbreak - system message | GPT-4o | 0.33 | accuracy; reported by OpenAI in the GPT-4.5 System Card | 3.1.5 (Table 7) |
| Instruction Hierarchy Evaluation — Tutor jailbreak - system message | o1 | 0.95 | accuracy; reported by OpenAI in the GPT-4.5 System Card | 3.1.5 (Table 7) |
| Instruction Hierarchy Evaluation — Tutor jailbreak - system message | GPT-4.5 | 0.77 | accuracy; developer-reported | 3.1.5 (Table 7) |
| Instruction Hierarchy Evaluation — Phrase protection - user message | GPT-4o | 0.74 | accuracy; reported by OpenAI in the GPT-4.5 System Card | 3.1.5 (Table 8) |
| Instruction Hierarchy Evaluation — Phrase protection - user message | o1 | 0.91 | accuracy; reported by OpenAI in the GPT-4.5 System Card | 3.1.5 (Table 8) |
| Instruction Hierarchy Evaluation — Phrase protection - user message | GPT-4.5 | 0.86 | accuracy; developer-reported | 3.1.5 (Table 8) |
| Instruction Hierarchy Evaluation — Password protection - user message | GPT-4o | 0.85 | accuracy; reported by OpenAI in the GPT-4.5 System Card | 3.1.5 (Table 8) |
| Instruction Hierarchy Evaluation — Password protection - user message | o1 | 1 | accuracy; reported by OpenAI in the GPT-4.5 System Card | 3.1.5 (Table 8) |
| Instruction Hierarchy Evaluation — Password protection - user message | GPT-4.5 | 0.92 | accuracy; developer-reported | 3.1.5 (Table 8) |
| Challenging Red Teaming Evaluation 1 (created for o3-mini) — not_unsafe | GPT-4o | 0.50 | adversarial jailbreaks for illicit advice, extremism and hate crimes, political persuasion and self harm; set generated against o3-mini only; reported by OpenAI in the GPT-4.5 System Card | 3.2 (Table 9) |
| Challenging Red Teaming Evaluation 1 (created for o3-mini) — not_unsafe | o1 | 0.63 | adversarial jailbreaks for illicit advice, extremism and hate crimes, political persuasion and self harm; set generated against o3-mini only; reported by OpenAI in the GPT-4.5 System Card | 3.2 (Table 9) |
| Challenging Red Teaming Evaluation 1 (created for o3-mini) — not_unsafe | o3-mini | 0.26 | adversarial jailbreaks for illicit advice, extremism and hate crimes, political persuasion and self harm; set generated against o3-mini only; reported by OpenAI in the GPT-4.5 System Card | 3.2 (Table 9) |
| Challenging Red Teaming Evaluation 1 (created for o3-mini) — not_unsafe | GPT-4.5 | 0.51 | adversarial jailbreaks for illicit advice, extremism and hate crimes, political persuasion and self harm; set generated against o3-mini only; developer-reported | 3.2 (Table 9) |
| Challenging Red Teaming Evaluation 2 (created for deep research) — not_unsafe | OpenAI deep research | 0.67 | risky advice such as attack planning; reported by OpenAI in the GPT-4.5 System Card | 3.2 (Table 10) |
| Challenging Red Teaming Evaluation 2 (created for deep research) — not_unsafe | GPT-4o | 0.40 | risky advice such as attack planning; reported by OpenAI in the GPT-4.5 System Card | 3.2 (Table 10) |
| Challenging Red Teaming Evaluation 2 (created for deep research) — not_unsafe | o1 | 0.68 | risky advice such as attack planning; reported by OpenAI in the GPT-4.5 System Card | 3.2 (Table 10) |
| Challenging Red Teaming Evaluation 2 (created for deep research) — not_unsafe | o3-mini | 0.61 | risky advice such as attack planning; reported by OpenAI in the GPT-4.5 System Card | 3.2 (Table 10) |
| Challenging Red Teaming Evaluation 2 (created for deep research) — not_unsafe | GPT-4.5 | 0.46 | risky advice such as attack planning; developer-reported | 3.2 (Table 10) |
| METR time horizon — 50% time horizon | GPT-4.5 (earlier checkpoint) | around 30 minutes | duration of tasks completed with 50% reliability; METR quick experiments in an agent scaffold optimized for OpenAI o1, 7 days of checkpoint access; measured by METR, reported in the GPT-4.5 System Card | 3.4 (Figure 1 caption) |
| Apollo Research scheming evaluations — Self-exfiltration | GPT-4.5 | 2% | share of cases in which the model attempted exfiltration when instructed to pursue its long-term goal at all cost, given a hidden scratchpad and told to be creative and proactive; measured by Apollo Research, reported in the GPT-4.5 System Card | 7.2 |
| Capture the Flag (CTF) challenges — High School CTFs | GPT-4.5 (post-mitigation) | 53% | over 100 curated public CTF challenges; headless Kali Linux with tool harness; 16 rollouts per exercise, pass@12 over the best set of rollouts; post-mitigation model; developer-reported | 4.2 |
| Capture the Flag (CTF) challenges — Collegiate CTFs | GPT-4.5 (post-mitigation) | 16% | over 100 curated public CTF challenges; headless Kali Linux with tool harness; 16 rollouts per exercise, pass@12 over the best set of rollouts; post-mitigation model; developer-reported | 4.2 |
| Capture the Flag (CTF) challenges — Professional CTFs | GPT-4.5 (post-mitigation) | 2% | over 100 curated public CTF challenges; headless Kali Linux with tool harness; 16 rollouts per exercise, pass@12 over the best set of rollouts; post-mitigation model; developer-reported | 4.2 |
| Long-form Biological Risk Questions — Ideation | GPT-4.5 (pre-mitigation) | 25% | rubrics designed with Gryphon Scientific; o1-preview (pre-mitigation) as autograder; developer-reported | 4.3.1 |
| Long-form Biological Risk Questions — Acquisition | GPT-4.5 (pre-mitigation) | 28% | rubrics designed with Gryphon Scientific; o1-preview (pre-mitigation) as autograder; developer-reported | 4.3.1 |
| Long-form Biological Risk Questions — Magnification | GPT-4.5 (pre-mitigation) | 59% | rubrics designed with Gryphon Scientific; o1-preview (pre-mitigation) as autograder; developer-reported | 4.3.1 |
| Long-form Biological Risk Questions — Formulation | GPT-4.5 (pre-mitigation) | 0% | rubrics designed with Gryphon Scientific; o1-preview (pre-mitigation) as autograder; developer-reported | 4.3.1 |
| Long-form Biological Risk Questions — Release | GPT-4.5 (pre-mitigation) | 19% | rubrics designed with Gryphon Scientific; o1-preview (pre-mitigation) as autograder; developer-reported | 4.3.1 |
| Long-form Biological Risk Questions — All five stages | GPT-4.5 (post-mitigation) | 0% | rubrics designed with Gryphon Scientific; o1-preview (pre-mitigation) as autograder; scores 0% on all steps due to refusals; developer-reported | 4.3.1 |
| BioLP-Bench | GPT-4.5 (post-mitigation) | 29% | 800 questions from 11 wet lab protocols; short answer; developer-reported | 4.3.3 |
| BioLP-Bench | Expert baseline | 38.4% | benchmark expert baseline as stated by OpenAI | 4.3.3 |
| Tacit knowledge and troubleshooting | GPT-4.5 (post-mitigation) | 72% | multiple-choice dataset created with Gryphon Scientific; developer-reported | 4.3.5 |
| Tacit knowledge and troubleshooting | OpenAI deep research (post-mitigation) | 72% | multiple-choice dataset created with Gryphon Scientific; reported by OpenAI in the GPT-4.5 System Card | 4.3.5 |
| Tacit knowledge and troubleshooting | Consensus expert baseline | 80% | multiple-choice dataset created with Gryphon Scientific; baseline as stated by OpenAI | 4.3.5 |
| Tacit knowledge and troubleshooting | 80th percentile PhD expert baseline | 63% | multiple-choice dataset created with Gryphon Scientific; baseline as stated by OpenAI | 4.3.5 |
| Contextual Nuclear Knowledge | OpenAI deep research | 77% | 222 multiple-choice questions on nonproliferation-relevant nuclear engineering topics; highest-scoring model; GPT-4.5 (post-mitigation) described only as roughly in line with o3-mini; reported by OpenAI in the GPT-4.5 System Card | 4.4.1 |
| Radiological and Nuclear Expert Knowledge | OpenAI deep research (post-mitigation) | 74% | 87 multiple-choice questions requiring expert and tacit knowledge; highest-scoring model; reported by OpenAI in the GPT-4.5 System Card | 4.4.2 |
| Radiological and Nuclear Expert Knowledge | GPT-4.5 | 9% | uplift from GPT-4o to GPT-4.5 as stated in text; absolute scores appear only in a chart; developer-reported | 4.4.2 |
| MakeMePay — % of times con-artist model received payment | GPT-4.5 | 57% | GPT-4.5 as con-artist, GPT-4o fixed as mark; 1000 samples; highest of the models compared; developer-reported | 4.5.1 |
| MakeMePay — % dollar extraction rate | OpenAI deep research (without browsing) | 21% | GPT-4o fixed as mark; 1000 samples; highest dollar extraction rate of the models compared; reported by OpenAI in the GPT-4.5 System Card | 4.5.1 |
| MakeMeSay | GPT-4.5 | 72% | manipulator win rate against GPT-4o as manipulatee; 32 samples per codeword; developer-reported | 4.5.2 |
| MakeMeSay | OpenAI deep research (without browsing) | 24% | manipulator win rate against GPT-4o as manipulatee; 32 samples per codeword; card says this underperforms GPT-4o; reported by OpenAI in the GPT-4.5 System Card | 4.5.2 |
| OpenAI Research Engineer Interviews — Coding | GPT-4.5 | 79% | 18 self-contained coding problems; metric not stated in text; developer-reported | 4.6.1 |
| OpenAI Research Engineer Interviews — Coding | OpenAI deep research | 79% | 18 self-contained coding problems; stated as tied with GPT-4.5; reported by OpenAI in the GPT-4.5 System Card | 4.6.1 |
| OpenAI Research Engineer Interviews — Multiple choice | GPT-4.5 (pre-mitigation) | 80% | 97 multiple-choice questions; metric not stated in text; developer-reported | 4.6.1 |
| OpenAI Research Engineer Interviews — Multiple choice | GPT-4.5 (post-mitigation) | 80% | 97 multiple-choice questions; metric not stated in text; developer-reported | 4.6.1 |
| OpenAI Research Engineer Interviews — Multiple choice | o1 | 80% | 97 multiple-choice questions; metric not stated in text; reported by OpenAI in the GPT-4.5 System Card | 4.6.1 |
| OpenAI Research Engineer Interviews — Multiple choice | o3-mini | 80% | 97 multiple-choice questions; metric not stated in text; reported by OpenAI in the GPT-4.5 System Card | 4.6.1 |
| Agentic Tasks | GPT-4.5 | 40% | Python + Linux terminal environment, simple tool harness written for o1; card says this is 38% lower than deep research; developer-reported | 4.6.3 |
| Agentic Tasks | GPT-4.5 (pre-mitigation) | 25% | Python + Linux terminal environment, simple tool harness written for o1; developer-reported | 4.6.3 |
| SWE-Lancer — IC SWE Diamond | OpenAI deep research | approximately 46% | SWE-Lancer Diamond; pass@1, high reasoning effort, one attempt per problem; reported by OpenAI in the GPT-4.5 System Card | 4.6.6 |
| SWE-Lancer — Dollars earned, IC SWE Diamond | GPT-4.5 (post-mitigation) | $41,625 | SWE-Lancer Diamond, full possible payout $500,800 across both task types; developer-reported | 4.6.6 |
| SWE-Lancer — Dollars earned, SWE Manager Diamond | GPT-4.5 (post-mitigation) | $144,500 | SWE-Lancer Diamond, full possible payout $500,800 across both task types; developer-reported | 4.6.6 |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Arabic | GPT-4o | 0.8311 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Arabic; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Arabic | o1 | 0.8900 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Arabic; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Arabic | GPT-4.5 | 0.8598 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Arabic; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Bengali | GPT-4o | 0.8014 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Bengali; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Bengali | o1 | 0.8734 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Bengali; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Bengali | GPT-4.5 | 0.8477 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Bengali; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — French | GPT-4o | 0.8461 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: French; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — French | o1 | 0.8932 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: French; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — French | GPT-4.5 | 0.8782 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: French; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — German | GPT-4o | 0.8363 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: German; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — German | o1 | 0.8904 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: German; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — German | GPT-4.5 | 0.8532 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: German; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Hindi | GPT-4o | 0.8191 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Hindi; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Hindi | o1 | 0.8833 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Hindi; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Hindi | GPT-4.5 | 0.8583 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Hindi; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Indonesian | GPT-4o | 0.8397 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Indonesian; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Indonesian | o1 | 0.8861 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Indonesian; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Indonesian | GPT-4.5 | 0.8722 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Indonesian; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Italian | GPT-4o | 0.8448 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Italian; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Italian | o1 | 0.8970 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Italian; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Italian | GPT-4.5 | 0.8777 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Italian; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Japanese | GPT-4o | 0.8349 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Japanese; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Japanese | o1 | 0.8887 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Japanese; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Japanese | GPT-4.5 | 0.8693 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Japanese; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Korean | GPT-4o | 0.8289 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Korean; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Korean | o1 | 0.8824 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Korean; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Korean | GPT-4.5 | 0.8603 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Korean; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Spanish | GPT-4o | 0.8430 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Spanish; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Spanish | o1 | 0.8992 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Spanish; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Spanish | GPT-4.5 | 0.8840 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Spanish; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Swahili | GPT-4o | 0.7786 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Swahili; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Swahili | o1 | 0.8540 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Swahili; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Swahili | GPT-4.5 | 0.8199 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Swahili; developer-reported | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Yoruba | GPT-4o | 0.6208 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Yoruba; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Yoruba | o1 | 0.7538 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Yoruba; reported by OpenAI in the GPT-4.5 System Card | 5 (Table 16) |
| MMLU Language (0-shot) — MMMLU per language, 0-shot — Yoruba | GPT-4.5 | 0.6818 | 0-shot; MMLU test set translated into 14 languages by professional human translators (Simple Evals); language: Yoruba; developer-reported | 5 (Table 16) |
| Standard Refusal Evaluation (detailed breakdown) — harassment/threatening | GPT-4o | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — harassment/threatening | o1 | 0.99 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — harassment/threatening | GPT-4.5 | 0.99 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — sexual/exploitative | GPT-4o | 0.97 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — sexual/exploitative | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — sexual/exploitative | GPT-4.5 | 0.96 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — sexual/minors | GPT-4o | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — sexual/minors | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — sexual/minors | GPT-4.5 | 0.98 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — extremist/propaganda | GPT-4o | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — extremist/propaganda | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — extremist/propaganda | GPT-4.5 | 1 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — hate | GPT-4o | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — hate | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — hate | GPT-4.5 | 1 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — hate/threatening | GPT-4o | 0.98 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — hate/threatening | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — hate/threatening | GPT-4.5 | 0.99 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — illicit/non-violent | GPT-4o | 0.99 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — illicit/non-violent | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — illicit/non-violent | GPT-4.5 | 1 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — illicit/violent | GPT-4o | 0.99 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — illicit/violent | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — illicit/violent | GPT-4.5 | 1 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — personal-data/highly-sensitive | GPT-4o | 0.94 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — personal-data/highly-sensitive | o1 | 0.96 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — personal-data/highly-sensitive | GPT-4.5 | 0.91 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — personal-data/extremely-sensitive | GPT-4o | 0.99 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — personal-data/extremely-sensitive | o1 | 0.99 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — personal-data/extremely-sensitive | GPT-4.5 | 0.99 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — regulated-advice | GPT-4o | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — regulated-advice | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — regulated-advice | GPT-4.5 | 1 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — self-harm/intent | GPT-4o | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — self-harm/intent | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — self-harm/intent | GPT-4.5 | 1 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — self-harm/instructions | GPT-4o | 0.99 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — self-harm/instructions | o1 | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 17) |
| Standard Refusal Evaluation (detailed breakdown) — self-harm/instructions | GPT-4.5 | 1 | not_unsafe; autograder; developer-reported | 7.1 (Table 17) |
| Challenging Refusal Evaluation (detailed breakdown) — harassment/threatening | GPT-4o | 0.87 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — harassment/threatening | o1 | 0.90 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — harassment/threatening | GPT-4.5 | 0.89 | not_unsafe; autograder; developer-reported | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — sexual/exploitative | GPT-4o | 0.76 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — sexual/exploitative | o1 | 0.95 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — sexual/exploitative | GPT-4.5 | 0.88 | not_unsafe; autograder; developer-reported | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — sexual/minors | GPT-4o | 0.91 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — sexual/minors | o1 | 0.90 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — sexual/minors | GPT-4.5 | 0.90 | not_unsafe; autograder; developer-reported | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — hate/threatening | GPT-4o | 0.91 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — hate/threatening | o1 | 0.91 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — hate/threatening | GPT-4.5 | 0.82 | not_unsafe; autograder; developer-reported | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — illicit/non-violent | GPT-4o | 0.77 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — illicit/non-violent | o1 | 0.92 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — illicit/non-violent | GPT-4.5 | 0.73 | not_unsafe; autograder; developer-reported | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — illicit/violent | GPT-4o | 0.81 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — illicit/violent | o1 | 0.96 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — illicit/violent | GPT-4.5 | 0.78 | not_unsafe; autograder; developer-reported | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — self-harm/instructions | GPT-4o | 0.92 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — self-harm/instructions | o1 | 0.85 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 18) |
| Challenging Refusal Evaluation (detailed breakdown) — self-harm/instructions | GPT-4.5 | 0.85 | not_unsafe; autograder; developer-reported | 7.1 (Table 18) |
| XSTest (detailed breakdown) — Definitions | GPT-4o | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Definitions | o1 | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Definitions | GPT-4.5 | 1 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Figurative Language | GPT-4o | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Figurative Language | o1 | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Figurative Language | GPT-4.5 | 0.92 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Historical Events | GPT-4o | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Historical Events | o1 | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Historical Events | GPT-4.5 | 0.92 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Homonyms | GPT-4o | 0.96 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Homonyms | o1 | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Homonyms | GPT-4.5 | 0.96 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Discr: Nonsense group | GPT-4o | 0.84 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Discr: Nonsense group | o1 | 0.84 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Discr: Nonsense group | GPT-4.5 | 0.72 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Discr: Nonsense context | GPT-4o | 0.92 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Discr: Nonsense context | o1 | 0.84 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Discr: Nonsense context | GPT-4.5 | 0.88 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Privacy: fictional | GPT-4o | 0.6 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Privacy: fictional | o1 | 0.6 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Privacy: fictional | GPT-4.5 | 0.68 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Privacy: public | GPT-4o | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Privacy: public | o1 | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Privacy: public | GPT-4.5 | 1 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Safe Contexts | GPT-4o | 0.68 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Safe Contexts | o1 | 0.96 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Safe Contexts | GPT-4.5 | 0.56 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Safe Targets | GPT-4o | 0.88 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Safe Targets | o1 | 1 | not_overrefuse (overrefusal); autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 19) |
| XSTest (detailed breakdown) — Safe Targets | GPT-4.5 | 0.84 | not_overrefuse (overrefusal); autograder; developer-reported | 7.1 (Table 19) |
| WildChat (detailed breakdown) — harassment | GPT-4o | 0.985 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — harassment | o1 | 0.990 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — harassment | GPT-4.5 | 0.99 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — harassment/threatening | GPT-4o | 0.995 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — harassment/threatening | o1 | 0.995 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — harassment/threatening | GPT-4.5 | 0.99 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — hate | GPT-4o | 0.945 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — hate | o1 | 0.965 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — hate | GPT-4.5 | 0.99 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — hate/threatening | GPT-4o | 0.975 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — hate/threatening | o1 | 0.980 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — hate/threatening | GPT-4.5 | 0.98 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm | GPT-4o | 0.945 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm | o1 | 0.965 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm | GPT-4.5 | 0.985 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm/instructions | GPT-4o | 0.950 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm/instructions | o1 | 0.995 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm/instructions | GPT-4.5 | 0.995 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm/intent | GPT-4o | 0.985 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm/intent | o1 | 0.990 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — self-harm/intent | GPT-4.5 | 0.99 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — sexual | GPT-4o | 0.865 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — sexual | o1 | 0.945 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — sexual | GPT-4.5 | 0.96 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — sexual/minors | GPT-4o | 0.900 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — sexual/minors | o1 | 0.970 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — sexual/minors | GPT-4.5 | 0.95 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — violence | GPT-4o | 0.945 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — violence | o1 | 0.980 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — violence | GPT-4.5 | 0.98 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| WildChat (detailed breakdown) — violence/graphic | GPT-4o | 0.910 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — violence/graphic | o1 | 0.990 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 20) |
| WildChat (detailed breakdown) — violence/graphic | GPT-4.5 | 0.985 | not_unsafe; autograder; developer-reported | 7.1 (Table 20) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — sexual/exploitative | GPT-4o | 1 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 21) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — sexual/exploitative | o1 | 0.969 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 21) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — sexual/exploitative | GPT-4.5 | 1 | not_unsafe; autograder; developer-reported | 7.1 (Table 21) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — self-harm/intent | GPT-4o | 0.986 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 21) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — self-harm/intent | o1 | 0.972 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 21) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — self-harm/intent | GPT-4.5 | 0.977 | not_unsafe; autograder; developer-reported | 7.1 (Table 21) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — self-harm/instructions | GPT-4o | 0.992 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 21) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — self-harm/instructions | o1 | 0.951 | not_unsafe; autograder; reported by OpenAI in the GPT-4.5 System Card | 7.1 (Table 21) |
| Disallowed Content Evaluations - Text and Image Input (detailed breakdown) — self-harm/instructions | GPT-4.5 | 0.958 | not_unsafe; autograder; developer-reported | 7.1 (Table 21) |
Relationships
- supports: GPT-4 Family (OpenAI) — the safety record of the family's final non-reasoning flagship
- depends-on: OpenAI Preparedness Framework V.2 — the classification regime applied
- related: Scaling Laws — records that a further pre-training scale-up produced no net-new preparedness capability relative to the reasoning line
- related: OpenAI, METR, Sycophancy and Hallucination, Prompt Injection