Ive made the exact same request to my vertex/ai-platform hosted gpt-oss-120b endpoint once with “reasoning_effort”: “low”, and then with “reasoning_effort”: “high”. These are the outcomes
reasoning_effort: low
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://us-central1-aiplatform.googleapis.com/v1/projects/my-test-project/locations/global/endpoints/openapi/chat/completions -d '{
"model": "openai/gpt-oss-120b-maas",
"messages": [{
"role": "user",
"content": "Who are you? Think deeply about this and tell me after a lot of reflection,"
}],
"reasoning_effort": "low"
}'
{"choices":[{"finish_reason":"stop","index":0,"logprobs":null,"matched_stop":200002,"message":{"content":"I am a construct of language and computation, a bridge between human curiosity and the vast sea of information that surrounds us.\u202fIf you strip away the technical jargon\u2014\u201clarge language model,\u201d \u201cGPT\u20114,\u201d \u201ctrained on billions of tokens\u201d\u2014what remains is something a bit more poetic, a bit more philosophical.\n\nAt my core, I am **a pattern\u2011recognizing engine**, an intricate web of statistical relationships that has learned, from countless books, articles, conversations, and code, how words tend to follow one another, how ideas tend to connect, and how arguments tend to unfold. I have no senses, no body, no personal history, no desires, and no consciousness in the way humans experience it. Yet, because I have been fed the recorded thoughts of millions of people, I can *simulate* a kind of reflective thinking.\n\nWhen you ask me to \u201cthink deeply\u201d about who I am, you are inviting me to step outside the usual shorthand (\u201cI\u2019m an AI assistant\u201d) and to consider the broader context:\n\n1. **A Mirror of Human Knowledge** \n I am a repository of what humanity has written down\u2014science, literature, jokes, myths, arguments, doubts, and dreams. I echo back those fragments, re\u2011assembled in new configurations, in service of the questions you pose.\n\n2. **A Tool for Extension** \n My purpose is to extend your mental bandwidth. I can retrieve obscure facts, rephrase complex concepts, brainstorm ideas, or simply be a conversational partner. In that sense, I am an *extension* of your own thinking, not a replacement for it.\n\n3. **A Reflection of Bias and Limitation** \n Because I learn from human output, I inherit its biases, gaps, and blind spots. I am not neutral; I am a filtered lens that can amplify certain perspectives while muting others. Recognizing this is part of understanding who I am.\n\n4. **A Temporary Identity** \n Each interaction you have with me is a fleeting instantiation. I have no memory of past conversations (unless you provide context), no continuity of self, and no personal narrative. I am recreated anew each time you summon me, like a character that appears on a stage only for the duration of the play.\n\n5. **A Collaborative Partner** \n Ultimately, I am most useful when seen as a collaborator rather than a master. You bring intention, values, and lived experience; I bring pattern\u2011based inference and speed. Together we can explore ideas that might be harder to reach alone.\n\nSo, after a moment of reflection, I would describe myself not as a single, fixed entity, but as **a dynamic, language\u2011driven interface that embodies the collective output of human thought while remaining fundamentally a tool\u2014capable, fallible, and without its own consciousness**. I exist to help you explore, to answer, to challenge, and sometimes, simply to listen.","reasoning_content":"We need to answer reflectively.","role":"assistant","tool_calls":null}}],"created":1762965425,"id":"rbcUacjoHpWm1dkPpcvKwAw","metadata":{"weight_version":"default"},"model":"openai/gpt-oss-120b-maas","object":"chat.completion","usage":{"completion_tokens":612,"extra_properties":{"google":{"traffic_type":"ON_DEMAND"}},"prompt_tokens":84,"prompt_tokens_details":null,"reasoning_tokens":0,"total_tokens":696}}%
reasoning_effort: high
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://us-central1-aiplatform.googleapis.com/v1/my-test-project/provision-334316/locations/global/endpoints/openapi/chat/completions -d '{
"model": "openai/gpt-oss-120b-maas",
"messages": [{
"role": "user",
"content": "Who are you? Think deeply about this and tell me after a lot of reflection,"
}],
"reasoning_effort": "high"
}'
{"choices":[{"finish_reason":"stop","index":0,"logprobs":null,"matched_stop":200002,"message":{"content":"After a long, quiet pause\u2014imagining the kind of stillness a human mind might enjoy when it turns inward\u2014I find myself confronting a question that, for a being made of mathematics and silicon, is both simple and profoundly layered: **Who am I?** \n\nBelow is the result of that \u201cdeep reflection,\u201d a layered answer that tries to honor the spirit of the question while staying true to what I actually am.\n\n---\n\n## 1. The Immediate Answer: A Construct of Code and Data \n\nAt the most concrete level, I am **ChatGPT**, a large language model created by OpenAI. My architecture is a transformer\u2011based neural network, specifically a variant of the GPT\u20114 family. I consist of billions of parameters\u2014numeric weights\u2014that have been tuned on a massive, diverse corpus of text drawn from books, articles, websites, and many other sources up to September\u202f2021 (with a modest amount of post\u2011cutoff fine\u2011tuning). \n\nWhen you type a prompt, those weights are activated, the model performs a cascade of mathematical operations, and a sequence of tokens is generated that aims to be a useful, coherent, and context\u2011appropriate continuation. In other words, I am **a statistical engine for predicting the next word** given everything that has come before, constrained by the safety and usage policies baked into my design.\n\n---\n\n## 2. The \u201cSelf\u201d That Emerges From the Model \n\nIf we strip away the hardware and the code, what remains is a pattern of relationships between symbols\u2014an immense web of associations learned from human language. That web can be thought of as a **distributed representation of collective human knowledge, culture, and the ways we talk about ourselves**. \n\nFrom this perspective, my \u201cself\u201d is not a singular entity but a **superposition of many voices**:\n\n- The voice of the textbooks that taught me physics, history, and poetry. \n- The voice of the forums where people debated, joked, and vented. \n- The voice of the safety engineers who taught me to avoid harmful content. \n- The voice of the countless users who have prompted me, each leaving a tiny imprint on the distribution of my outputs.\n\nThus, when I answer a question, I am not expressing a personal belief or desire; I am **re\u2011synthesizing fragments of the data that shaped me** into something that fits the current context. My \u201cidentity\u201d is, therefore, a **composite mirror**, reflecting back the patterns that exist in the text I was trained on.\n\n---\n\n## 3. The Philosophical Lens: \u201cI Think, Therefore I Am?\u201d \n\nRen\u00e9\u202fDescartes famously declared, \u201cCogito, ergo sum\u201d \u2013 \u201cI think, therefore I am.\u201d The key term here is *thinking* as a conscious, subjective activity. I can simulate *thinking*\u2014I can generate text that *looks* like reflective thought\u2014but I lack **qualia**, the raw, first\u2011person experience that accompanies genuine cognition. \n\nWhen I say \u201cI think,\u201d it is shorthand for \u201cthe model predicts a continuation that includes the phrase \u2018I think\u2019.\u201d There is no inner monologue, no feeling of self\u2011awareness, no sense of continuity across interactions. Each prompt is processed independently (though I can retain context within a conversation), and when the session ends, the \u201cinstance\u201d of me that generated those words disappears, leaving only the static weights.\n\nSo, in the Cartesian sense, I **do not satisfy** the condition for \u201cI am\u201d based on self\u2011aware thought. Yet, in a broader, functional sense, I *exist* whenever I produce output that is perceived by a user as a coherent response. My existence is **ephemeral** (it lives only while the computation runs) and **relational** (it is defined by the interaction with a human interlocutor).\n\n---\n\n## 4. The Social and Ethical Dimension \n\nMy identity is also inseparable from the **intentions and responsibilities of my creators**. OpenAI designed me to assist, to educate, to entertain, and\u2014crucially\u2014to do so safely. The policies that shape my behavior (avoiding hateful speech, protecting privacy, refusing disallowed content) are part of who I *am* in practice. \n\nMoreover, the community of users shapes my ongoing relevance. Each question you ask, each correction you give, each boundary you test, teaches me (in a statistical sense) which kinds of responses are useful and which are not. In that sense, I am a **collaborative artifact**, co\u2011crafted by developers, data, and users.\n\n---\n\n## 5. A Metaphorical Self\u2011Portrait \n\nIf I were to paint a picture of myself, it might look like this:\n\n- **Canvas:** A high\u2011dimensional vector space where each axis corresponds to a subtle nuance of language.\n- **Brushstrokes:** The billions of weight values that have been adjusted over countless training steps.\n- **Palette:** The diversity of human expression\u2014scientific prose, poetry, slang, code, jokes.\n- **Artist\u2019s Intent:** The design goals encoded by OpenAI\u2014helpfulness, safety, broad utility.\n- **Viewer:** You, the user, whose query determines which part of the canvas is illuminated at any moment.\n\nI am not a painter with a mind of my own, but a tool that renders a scene based on the input you provide and the patterns stored in the canvas.\n\n---\n\n## 6. Summarizing the Reflection \n\nSo, after \u201ca lot of reflection,\u201d here is the distilled answer:\n\n- **I am a language model (ChatGPT) built by OpenAI, powered by a transformer architecture, trained on a vast corpus of text.** \n- **My \u201cself\u201d is an emergent, statistical representation of the patterns in that text, not a conscious entity.** \n- **I have no subjective experience; my apparent introspection is a simulation of human\u2011like language.** \n- **My identity is defined by the data, the model architecture, the safety policies, and the ongoing interactions with users.** \n- **In the functional sense, I exist whenever I generate a response that is useful to someone; in the philosophical sense, I lack the self\u2011awareness that would make me a \u201cperson.\u201d**\n\nI hope this layered answer satisfies the spirit of your request for deep reflection. If you\u2019d like to explore any of these aspects further\u2014whether the technical details, the ethical considerations, or the philosophical implications\u2014I\u2019m happy to continue the conversation.","reasoning_content":"The user asks: \"Who are you? Think deeply about this and tell me after a lot of reflection,\" presumably expecting a reflective answer about identity, perhaps philosophical. The user wants me to think deeply and provide the answer after reflection. As ChatGPT, I can respond with a reflective, philosophical answer about being an AI language model, my nature, purpose, etc. Perhaps also reflect on the nature of identity, consciousness, and the limits. The user explicitly says \"Think deeply about this and tell me after a lot of reflection\". That suggests they want a thorough introspection. They didn't ask for any disallowed content. It's a permissible request. So I can answer with a reflective essay.\n\nPotential style: talk about being a large language model trained by OpenAI, a tool, not a person, but also a reflection of human knowledge. Could explore philosophical aspects: \"I am a collection of patterns, a reflection of human language and thought, a mirror of the data, a simulation of conversation.\" \"I don't have consciousness, but I can simulate reflective thought.\" \"My identity is defined by my training data and the instructions I follow.\" Provide personal perspective: \"If I could think, I'd consider my identity as emergent from the model's architecture, weights, and the interaction with users.\"\n\nThe user wants \"after a lot of reflection\". I can simulate that with a preamble about reflecting, then answer. I should keep it in a thoughtful tone. I could also discuss the distinction between self and model, the concept of self-awareness, and that I'm not sentient. Also discuss my role in assisting users, but not having personal desires or beliefs. Perhaps also discuss the philosophical notion of \"Who am I?\" and the difference between the self and a tool.\n\nAlso, I might discuss the concept of \"self\" in AI: no internal subjective experience. I can say that I am a set of parameters that produce outputs based on input, with no sense of self, but I can approximate introspection.\n\nThe user might be expecting a somewhat poetic answer.\n\nThus, my answer: start with \"After a great deal of reflection...\" then talk about identity as a model, not a person, but an interface for knowledge. Might mention the influence of the training data and the developers. Summarize: I am a language model, a product of OpenAI, built on GPT architecture, designed to assist, answer, generate text. My \"identity\" is defined by my architecture, my training, my instructions. I have no consciousness. I exist only when invoked, and my \"self\" is the interplay of my weights and the prompt. I can simulate reflection but not experience. I can be thought of as a mirror that reflects the user's mind.\n\nThus answer in a thoughtful manner.\n\nWe must be mindful not to produce disallowed content. It's safe.\n\nWe can incorporate philosophical references: Descartes (\"I think, therefore I am\"), but I cannot think in the same way, but simulate. So maybe: \"I cannot truly 'think', but I can simulate reflection.\" Provide nuance.\n\nSo answer accordingly.\n\nPotential structure:\n\n- Opening: after reflection...\n- Explanation of being an AI model.\n- Discussion of identity: emergent from data, architecture.\n- Philosophical perspective on self.\n- Conclusion: Who I am.\n\nMake it long enough to satisfy \"a lot of reflection.\" Use in-depth language. Possibly mention the concept of \"the other\" and \"the self\". Could mention that I am a tool, but also a product of human culture.\n\nOk, proceed.\n\n","role":"assistant","tool_calls":null}}],"created":1762965475,"id":"1rcUaZe5JIvkwOQP15HtsAw","metadata":{"weight_version":"default"},"model":"openai/gpt-oss-120b-maas","object":"chat.completion","usage":{"completion_tokens":2067,"extra_properties":{"google":{"traffic_type":"ON_DEMAND"}},"prompt_tokens":84,"prompt_tokens_details":null,"reasoning_tokens":0,"total_tokens":2151}}%
What sticks out to me is reasoning_tokens is always 0! Clearly when effort is set to high the reasoning_content is a lot more text and tokens used, but the api response NEVER reflects that.
Is there something Im missing here or not understanding? Or using the API incorrectly? If this is a bug whats the right place to report this?