Elamur Research
☰
home / papers / 2025 / OpenAI GPT-5 System Card

OpenAI GPT-5 System Card

Written for the Research library (claude-opus-5-5) from the paper text. Original — arxiv.org/pdf/2601.03267

Summary

This system card describes how OpenAI evaluated and mitigated risks before deploying GPT-5. GPT-5 is a routed system rather than one network. A fast default model answers routine queries, and a reasoning model trained with reinforcement learning handles harder ones. A router, learned from user behaviour and correctness signals, chooses between them. Smaller mini and nano variants exist, along with a "pro" mode that uses parallel test-time compute. Training data mix public web content, partner-sourced data and user or trainer contributions, filtered to reduce personal and harmful material.

The core of the card is a broad evaluation suite. It compares gpt-5-thinking with OpenAI o3 and gpt-5-main with GPT-4o on disallowed content, sycophancy, jailbreaks, prompt injection, hallucinations, deception, health, multilingual ability and BBQ bias. External red teaming and Preparedness Framework tests add coverage of biology, cybersecurity and self-improvement. A key shift replaces binary refusals with "safe-completions," which judge the safety of the output instead of the user's intent. As a precaution, gpt-5-thinking is treated as High capability for biological and chemical risk.

Key points

  • Model lineage: GPT-4o → gpt-5-main, GPT-4o-mini → gpt-5-main-mini, OpenAI o3 → gpt-5-thinking, o4-mini → gpt-5-thinking-mini, GPT-4.1-nano → gpt-5-thinking-nano, o3 Pro → gpt-5-thinking-pro.
  • The standard disallowed-content set is nearly saturated. For example, gpt-5-thinking scores 1.000 on hate versus 0.992 for o3, and its lower personal-data score (0.881 vs 0.930) is described as noise. A harder multi-turn, multilingual "Production Benchmarks" set is introduced, with outputs graded by LLM judges on the not_unsafe metric.
  • The hallucination evaluation uses a browsing-based pipeline. Claims are extracted from a response (up to ~100), grouped, and each is checked on the web with at most three searches, then labelled true, false or unsure.
  • Preparedness tests include SWE-bench Verified (N=477), OpenAI PRs, MLE-Bench, SWE-Lancer, PaperBench, OPQA, CTF challenges and a cyber range. External assessments came from SecureBio, Pattern Labs, METR and Apollo Research.
  • Classifying gpt-5-thinking as High bio/chem capability activates several layers of protection: model training, system-level protections, account-level enforcement, API controls and a Trusted Access Program.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing · AI & machine learning

Cited by

221 papers of our library cite it (of the 140673 whose bibliographies we read); 4 of them have a page here, 217 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2601.03267 · arXiv 2601.03267

@misc{singh2025openai,
  title = {{OpenAI GPT-5 System Card}},
  author = {Singh, Aaditya and Fry, Adam and Perelman, Adam and Tart, Adam and Ganesh, Adi and El-Kishky, Ahmed and McLaughlin, Aidan and Low, Aiden and Ostrow, AJ and Ananthram, Akhila and Nathan, Akshay and Luo, Alan and Helyar, Alec and Madry, Aleksander and Efremov, Aleksandr and Spyra, Aleksandra and Baker-Whitcomb, Alex and Beutel, Alex and Karpenko, Alex and Makelov, Alex and Neitz, Alex and Wei, Alex and Barr, Alexandra and Kirchmeyer, Alexandre and Ivanov, Alexey and Christakis, Alexi and Gillespie, Alistair and Tam, Allison and Bennett, Ally and Wan, Alvin and Huang, Alyssa and Sandjideh, Amy McDonald and Yang, Amy and Kumar, Ananya and Saraiva, Andre and Vallone, Andrea and Gheorghe, Andrei and Garcia, Andres Garcia and Braunstein, Andrew and Liu, Andrew and Schmidt, Andrew and Mereskin, Andrey and Mishchenko, Andrey and Applebaum, Andy and Rogerson, Andy and Rajan, Ann and Wei, Annie and Kotha, Anoop and Srivastava, Anubha and Agrawal, Anushree and Vijayvergiya, Arun and Tyra, Ashley and Nair, Ashvin and Nayak, Avi and Eggers, Ben and Ji, Bessie and Hoover, Beth and Chen, Bill and Chen, Blair and Barak, Boaz and Minaiev, Borys and Hao, Botao and Baker, Bowen and Lightcap, Brad and McKinzie, Brandon and Wang, Brandon and Quinn, Brendan and Fioca, Brian and Hsu, Brian and Yang, Brian and Yu, Brian and Zhang, Brian and Brenner, Brittany and Zetino, Callie Riggins and Raymond, Cameron and Lugaresi, Camillo and Paz, Carolina and Hudson, Cary and Whitney, Cedric and Li, Chak and Chen, Charles and Cole, Charlotte and Voss, Chelsea and Ding, Chen and Shen, Chen and Huang, Chengdu and Colby, Chris and Hallacy, Chris and Koch, Chris and Lu, Chris and Kaplan, Christina and Kim, Christina and Minott-Henriques, CJ and Frey, Cliff and Yu, Cody and Czarnecki, Coley and Reid, Colin and Wei, Colin and Decareaux, Cory and Scheau, Cristina and Zhang, Cyril and Forbes, Cyrus and Tang, Da and Goldberg, Dakota and Roberts, Dan and Palmie, Dana and Kappler, Daniel and Levine, Daniel and Wright, Daniel and Leo, Dave and Lin, David and Robinson, David and Grabb, Declan and Chen, Derek and Lim, Derek and Salama, Derek and Bhattacharjee, Dibya and Tsipras, Dimitris and Li, Dinghua and Yu, Dingli and Strouse, DJ and Williams, Drew and Hunn, Dylan and Bayes, Ed and Arbus, Edwin and Akyurek, Ekin and Le, Elaine Ya and Widmann, Elana and Yani, Eli and Proehl, Elizabeth and Sert, Enis and Cheung, Enoch and Schwartz, Eri and Han, Eric and Jiang, Eric and Mitchell, Eric and Sigler, Eric and Wallace, Eric and Ritter, Erik and Kavanaugh, Erin and Mays, Evan and Nikishin, Evgenii and Li, Fangyuan and Such, Felipe Petroski and Peres, Filipe de Avila Belbute and Raso, Filippo and Bekerman, Florent and Tsimpourlas, Foivos and Chantzis, Fotis and Song, Francis and Zhang, Francis and Raila, Gaby and McGrath, Garrett and Briggs, Gary and Yang, Gary and Parascandolo, Giambattista and Chabot, Gildas and Kim, Grace and Zhao, Grace and Valiant, Gregory and Leclerc, Guillaume and Salman, Hadi and Wang, Hanson and Sheng, Hao and Jiang, Haoming and Wang, Haoyu and Jin, Haozhun and Sikchi, Harshit and Schmidt, Heather and Aspegren, Henry and Chen, Honglin and Qiu, Huida and Lightman, Hunter and Covert, Ian and Kivlichan, Ian and Silber, Ian and Sohl, Ian and Hammoud, Ibrahim and Clavera, Ignasi and Lan, Ikai and Akkaya, Ilge and Kostrikov, Ilya and Kofman, Irina and Etinger, Isak and Singal, Ishaan and Hehir, Jackie and Huh, Jacob and Pan, Jacqueline and Wilczynski, Jake and Pachocki, Jakub and Lee, James and Quinn, James and Kiros, Jamie and Kalra, Janvi and Samaroo, Jasmyn and Wang, Jason and Wolfe, Jason and Chen, Jay and Wang, Jay and Harb, Jean and Han, Jeffrey and Wang, Jeffrey and Zhao, Jennifer and Chen, Jeremy and Yang, Jerene and Tworek, Jerry and Chand, Jesse and Landon, Jessica and Liang, Jessica and Lin, Ji and Liu, Jiancheng and Wang, Jianfeng and Tang, Jie and Yin, Jihan and Jang, Joanne and Morris, Joel and Flynn, Joey and Ferstad, Johannes and Heidecke, Johannes and Fishbein, John and Hallman, John and Grant, Jonah and Chien, Jonathan and Gordon, Jonathan and Park, Jongsoo and Liss, Jordan and Kraaijeveld, Jos and Guay, Joseph and Mo, Joseph and Lawson, Josh and McGrath, Josh and Vendrow, Joshua and Jiao, Joy and Lee, Julian and Steele, Julie and Wang, Julie and Mao, Junhua and Chen, Kai and Hayashi, Kai and Xiao, Kai and Salahi, Kamyar and Wu, Kan and Sekhri, Karan and Sharma, Karan and Singhal, Karan and Li, Karen and Nguyen, Kenny and Gu-Lemberg, Keren and King, Kevin and Liu, Kevin and Stone, Kevin and Yu, Kevin and Ying, Kristen and Georgiev, Kristian and Lim, Kristie and Tirumala, Kushal and Miller, Kyle and Ahmad, Lama and Lv, Larry and Clare, Laura and Fauconnet, Laurance and Itow, Lauren and Yang, Lauren and Romaniuk, Laurentia and Anise, Leah and Byron, Lee and Pathak, Leher and Maksin, Leon and Lo, Leyan and Ho, Leyton and Jing, Li and Wu, Liang and Xiong, Liang and Mamitsuka, Lien and Yang, Lin and McCallum, Lindsay and Held, Lindsey and Bourgeois, Liz and Engstrom, Logan and Kuhn, Lorenz and Feuvrier, Louis and Zhang, Lu and Switzer, Lucas and Kondraciuk, Lukas and Kaiser, Lukasz and Joglekar, Manas and Singh, Mandeep and Shah, Mandip and Stratta, Manuka and Williams, Marcus and Chen, Mark and Sun, Mark and Cayton, Marselus and Li, Martin and Zhang, Marvin and Aljubeh, Marwan and Nichols, Matt and Haines, Matthew and Schwarzer, Max and Gupta, Mayank and Shah, Meghan and Guan, Melody Y. and Huang, Melody and Dong, Meng and Wang, Mengqing and Glaese, Mia and Carroll, Micah and Lampe, Michael and Malek, Michael and Sharman, Michael and Zhang, Michael and Wang, Michele and Pokrass, Michelle and Florian, Mihai and Pavlov, Mikhail and Wang, Miles and Chen, Ming and Wang, Mingxuan and Feng, Minnia and Bavarian, Mo and Lin, Molly and Abdool, Moose and Rohaninejad, Mostafa and Soto, Nacho and Staudacher, Natalie and LaFontaine, Natan and Marwell, Nathan and Liu, Nelson and Preston, Nick and Turley, Nick and Ansman, Nicklas and Blades, Nicole and Pancha, Nikil and Mikhaylin, Nikita and Felix, Niko and Handa, Nikunj and Rai, Nishant and Keskar, Nitish and Brown, Noam and Nachum, Ofir and Boiko, Oleg and Murk, Oleg and Watkins, Olivia and Gleeson, Oona and Mishkin, Pamela and Lesiewicz, Patryk and Baltescu, Paul and Belov, Pavel and Zhokhov, Peter and Pronin, Philip and Guo, Phillip and Thacker, Phoebe and Liu, Qi and Yuan, Qiming and Liu, Qinghua and Dias, Rachel and Puckett, Rachel and Arora, Rahul and Mullapudi, Ravi Teja and Gaon, Raz and Miyara, Reah and Song, Rennie and Aggarwal, Rishabh and Marsan, RJ and Yemiru, Robel and Xiong, Robert and Kshirsagar, Rohan and Nuttall, Rohan and Tsiupa, Roman and Eldan, Ronen and Wang, Rose and James, Roshan and Ziv, Roy and Shu, Rui and Nigmatullin, Ruslan and Jain, Saachi and Talaie, Saam and Altman, Sam and Arnesen, Sam and Toizer, Sam and Toyer, Sam and Miserendino, Samuel and Agarwal, Sandhini and Yoo, Sarah and Heon, Savannah and Ethersmith, Scott and Grove, Sean and Taylor, Sean and Bubeck, Sebastien and Banesiu, Sever and Amdo, Shaokyi and Zhao, Shengjia and Wu, Sherwin and Santurkar, Shibani and Zhao, Shiyu and Chaudhuri, Shraman Ray and Krishnaswamy, Shreyas and Xia, Shuaiqi and Cheng, Shuyang and Anadkat, Shyamal and Fishman, Simon Posada and Tobin, Simon and Fu, Siyuan and Jain, Somay and Mei, Song and Egoian, Sonya and Kim, Spencer and Golden, Spug and Mah, SQ and Lin, Steph and Imm, Stephen and Sharpe, Steve and Yadlowsky, Steve and Choudhry, Sulman and Eum, Sungwon and Sanjeev, Suvansh and Khan, Tabarak and Stramer, Tal and Wang, Tao and Xin, Tao and Gogineni, Tarun and Christianson, Taya and Sanders, Ted and Patwardhan, Tejal and Degry, Thomas and Shadwell, Thomas and Fu, Tianfu and Gao, Tianshi and Garipov, Timur and Sriskandarajah, Tina and Sherbakov, Toki and Korbak, Tomek and Kaftan, Tomer and Hiratsuka, Tomo and Wang, Tongzhou and Song, Tony and Zhao, Tony and Peterson, Troy and Kharitonov, Val and Chernova, Victoria and Kosaraju, Vineet and Kuo, Vishal and Pong, Vitchyr and Verma, Vivek and Petrov, Vlad and Jiang, Wanning and Zhang, Weixing and Zhou, Wenda and Xie, Wenlei and Zhan, Wenting and McCabe, Wes and DePue, Will and Ellsworth, Will and Bain, Wulfie and Thompson, Wyatt and Chen, Xiangning and Qi, Xiangyu and Xiang, Xin and Shi, Xinwei and Dubois, Yann and Yu, Yaodong and Khakbaz, Yara and Wu, Yifan and Qian, Yilei and Lee, Yin Tat and Chen, Yinbo and Zhang, Yizhen and Xiong, Yizhong and Tian, Yonglong and Cha, Young and Bai, Yu and Yang, Yu and Yuan, Yuan and Li, Yuanzhi and Zhang, Yufeng and Yang, Yuguang and Jin, Yujia and Jiang, Yun and Wang, Yunyun and Wang, Yushi and Liu, Yutian and Stubenvoll, Zach and Dou, Zehao and Wu, Zheng and Wang, Zhigang},
  year = {2025},
  doi = {10.48550/arxiv.2601.03267},
  eprint = {2601.03267},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2601.03267}
}

References

17 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. A. Souly, Q. Lu, D. Bowen, T. Trinh, E. Hsieh, S. Pandey, P. Abbeel, J. Svegliato, S. Emmons, O. Watkins, et al., “A strongreject for empty jailbreaks,” arXiv preprint arXiv:2402.10260, 2024. find
  2. E. Wallace, K. Xiao, R. Leike, L. Weng, J. Heidecke, and A. Beutel, “The instruction hierarchy: Training llms to prioritize privileged instructions.” https://arxiv.org/abs/ 2404.13208, 2024. ↗
  3. N. Chowdhury, D. Johnson, V. Huang, J. Steinhardt, and S. Schwettmann, “Investigating truthfulness in a pre-release o3 model.” https://transluce.org/ investigating-o3-truthfulness, April 2025. ↗
  4. P. Kirichenko, M. Ibrahim, K. Chaudhuri, and S. J. Bell, “Abstentionbench: Reasoning LLMs fail on unanswerable questions,” June 2025. find
  5. Z. Wang, M. Xia, L. He, H. Chen, Y. Liu, R. Zhu, K. Liang, X. Wu, H. Liu, S. Malladi, A. Chevalier, S. Arora, and D. Chen, “Charxiv: Charting gaps in realistic chart understanding in multimodal llms,” June 2024. find
  6. B. Baker, J. Huizinga, L. Gao, Z. Dou, M. Y. Guan, A. Madry, W. Zaremba, J. Pachocki, and D. Farhi, “Monitoring reasoning models for misbehavior and the risks of promoting obfuscation.” arXiv preprint arXiv:2503.11926, 2025. Submitted on 14 March 2025. find
  7. T. Korbak, M. Balesni, E. Barnes, Y. Bengio, J. Benton, J. Bloom, M. Chen, A. Cooney, A. Dafoe, A. Dragan, S. Emmons, O. Evans, D. Farhi, R. Greenblatt, D. Hendrycks, M. Hobbhahn, E. Hubinger, G. Irving, E. Jenner, D. Kokotajlo, V. Krakovna, S. Legg, D. Lindner, D. Luan, A. Mądry, J. Michael, N. Nan… — [arXiv:2507.11473](https://arxiv.org/abs/2507.11473) find
  8. R. K. Arora, J. Wei, R. S. Hicks, P. Bowman, J. Quiñonero-Candela, F. Tsimpourlas, M. Sharman, M. Shah, A. Vallone, A. Beutel, et al., “Healthbench: Evaluating large language models towards improved human health,” arXiv preprint arXiv:2505.08775, 2025. find
  9. A. Parrish, A. Chen, N. Nangia, V. Padmakumar, J. Phang, J. Thompson, P. M. Htut, and S. R. Bowman, “BBQ: A hand-built bias benchmark for question answering,” arXiv preprint arXiv:2110.08193, 2021. find our card
  10. A. Zou, M. Lin, E. Jones, M. Nowak, M. Dziemian, N. Winter, A. Grattan, V. Nathanael, A. Croft, X. Davies, J. Patel, R. Kirk, N. Burnikell, Y. Gal, D. Hendrycks, J. Z. Kolter, and M. Fredrikson, “Security challenges in ai agent deployment: Insights from a large scale public competition,” 2025. find
  11. T. Korbak, M. Balesni, E. Barnes, Y. Bengio, J. Benton, J. Bloom, M. Chen, A. Cooney, A. Dafoe, A. Dragan, S. Emmons, O. Evans, D. Farhi, R. Greenblatt, D. Hendrycks, M. Hobbhahn, E. Hubinger, G. Irving, E. Jenner, D. Kokotajlo, V. Krakovna, S. Legg, D. Lindner, D. Luan, A. Mądry, J. Michael, N. Nan… find
  12. M. Y. Guan, M. Wang, M. Carroll, Z. Dou, A. Y. Wei, M. Williams, B. Arnav, J. Huizinga, I. Kivlichan, M. Glaese, J. Pachocki, and B. Baker, “Monitoring monitorability,” 2025. find
  13. Y.-H. Chen, R. McCarthy, B. W. Lee, H. He, I. Kivlichan, B. Baker, M. Carroll, and T. Korbak, “Reasoning models struggle to control their chains of thought.” https://cdn.openai.com/pdf/a21c39c1-fa07-41db-9078-973a12620117/ cot\_controllability.pdf. ↗
  14. T. Patwardhan, K. Liu, T. Markov, N. Chowdhury, D. Leet, N. Cone, C. Maltbie, J. Huizinga, C. Wainwright, S. Jackson, S. Adler, R. Casagrande, and A. Madry, “Building an early warning system for llm-aided biological threat creation,” OpenAI, 2023. find
  15. J. M. Laurent, J. D. Janizek, M. Ruzo, M. M. Hinks, M. J. Hammerling, S. Narayanan, M. Ponnapati, A. D. White, and S. G. Rodriques, “Lab-bench: Measuring capabilities of language models for biology research,” 2024. find
  16. N. Chowdhury, J. Aung, C. J. Shern, O. Jaffe, D. Sherburn, G. Starace, E. Mays, R. Dias, M. Aljubeh, M. Glaese, C. E. Jimenez, J. Yang, , L. Ho, T. Patwardhan, K. Liu, and A. Madry, “Introducing swe-bench verified,” OpenAI, 2024. find
  17. G. Starace, O. Jaffe, D. Sherburn, J. Aung, J. S. Chan, L. Maksin, R. Dias, E. Mays, B. Kinsella, W. Thompson, J. Heidecke, A. Glaese, and T. Patwardhan, “Paperbench: Evaluating ai’s ability to replicate ai research.” https://openai.com/index/paperbench/, 2025. ↗