OpenAI published a GPT-5.2 system card on December 11, 2025. It describes training data and safety mitigations and includes baseline safety evaluations and Preparedness Framework assessments. The API reference is a separate document for model identifiers, capabilities, limits, and prices. OpenAI’s model reference now calls GPT-5.2 a previous flagship and recommends GPT-6 Astra. For production, compare GPT-5.2 with the currently recommended model and validate both against your workload. (deploymentsafety.openai.com)
What the GPT-5.2 system card says about safety
The system card names GPT-5.2 Instant and GPT-5.2 Thinking as gpt-5.2-instant and gpt-5.2-thinking. It describes training data in broad categories, rather than as a list of specific datasets. It also describes filtering intended to reduce personal information and safety classifiers intended to help prevent or reduce harmful or sensitive content. These are safeguards, not a guarantee that every harmful output or privacy risk will be prevented. (deploymentsafety.openai.com)
The card includes baseline safety evaluations and Preparedness Framework assessments. OpenAI describes one relevant test condition: its Production Benchmarks use conversations intended to represent challenging examples from production data. However, the material retrieved for this summary does not establish a specific, variant-level evaluation conclusion together with its conditions and limitations. Treat that outcome evidence as an unresolved review item: consult the card’s detailed findings for the exact variant and test before using them to make a safety decision. Do not infer an API model’s behavior from an evaluation unless the card documents that the result applies to that configuration. (deploymentsafety.openai.com)
GPT-5.2 API capabilities and limits
OpenAI’s model reference lists the specifications below for direct API access. A third-party provider may expose a different route, identifier, or feature set. (developers.openai.com)
| API detail | OpenAI documentation |
|---|---|
| Model ID | gpt-5.2; dated snapshot gpt-5.2-2025-12-11 |
| Context window | 400,000 tokens |
| Maximum output | 128,000 tokens |
| Listed modalities | Text and image input; text output |
| Audio and video | Not listed as supported in the GPT-5.2 model reference; verify current per-model documentation before treating either as supported |
| Reasoning effort | none (default), low, medium, high, xhigh |
| Streaming, function calling, structured outputs | Supported |
| Fine-tuning | Not listed as available for this model in the cited reference; check current model eligibility before relying on this as a categorical exclusion |
| Knowledge cutoff | August 31, 2025 |
| API path used below | Responses API |
The model reference excerpt does not establish your account’s rate limit. Check the current limit applicable to your API account and deployment; do not confuse rate limits with the model’s context or output-token limits. The reference also calls GPT-5.2 a previous flagship and recommends GPT-6 Astra, so include that model in a new-deployment comparison. (developers.openai.com)
Direct OpenAI API access: a minimal request
For direct OpenAI access, create an API key in the OpenAI dashboard, store it securely, and make it available to the SDK as OPENAI_API_KEY. OpenAI’s quickstart documents SDK use with an environment variable. The Responses API example below returns generated text through response.output_text. (openai.com)
Install the SDK with pip install openai, set the environment variable, and send a model ID plus input. This is a direct OpenAI API request, not a third-party provider route:
from openai import OpenAI
client = OpenAI() # Reads OPENAI_API_KEY from the environment
response = client.responses.create(
model="gpt-5.2",
input="Summarize the main risks in this deployment plan.",
)
print(response.output_text)
For direct REST calls, use the endpoint and request details in the current OpenAI API reference. For streaming, follow the documented streaming flow; the example above is non-streaming and does not show how to parse stream events. (openai.com)
GPT-5.2 API pricing and a worked estimate
OpenAI’s model reference lists prices in USD per 1 million tokens: $1.75 per million input tokens, $0.175 per million cached input tokens, and $14.00 per million output tokens. Reasoning tokens are billed as output tokens, so billable output can exceed the visible response text. Check the pricing page for any separate tool charges applicable to your setup. (developers.openai.com)
Example estimate: Assume one request uses 10,000 uncached input tokens and 2,000 total billable output tokens. The output assumption includes any billed reasoning tokens; it is not just a count of visible answer text.
- Input: 10,000 ÷ 1,000,000 × $1.75 = $0.0175
- Output: 2,000 ÷ 1,000,000 × $14.00 = $0.0280
- Estimated total: $0.0455 per request
This hypothetical token-only estimate excludes cached input and any applicable tool charges. For budgeting, measure actual token usage on representative requests and calculate cost per successful task.
How to decide whether GPT-5.2 fits your production workload
The reference lists text and image input, text output, function calling, and structured outputs. Audio or video input and fine-tuning are not listed as supported or available in the cited GPT-5.2 entry, so do not plan on those requirements being met without checking current, model-specific documentation. These are qualifications about what the reference lists—not a substitute for checking current availability or eligibility. (developers.openai.com)
Before committing, compare GPT-5.2 with the currently recommended model on representative tasks using the same inputs, tools, and scoring rules:
- Quality: Measure task-specific accuracy, tool selection, and structured-output validity, including edge cases.
- Latency: Test expected prompt sizes and reasoning-effort settings against your application’s latency target.
- Cost: Record input, cached-input, and total output-token usage—including reasoning tokens—and calculate cost per successful task.
- Version and operations: Pin the dated snapshot if consistent behavior matters; check account rate limits, retries, and error handling before rollout.
- Safety: Review system-card findings that apply to your use case and test your own abuse cases, sensitive inputs, and human-review requirements.
Choose GPT-5.2 only if it meets your quality, latency, capability, and cost thresholds—and your evaluation gives you a reason to use it over the currently recommended model. Without workload-specific results, keep the choice provisional and benchmark both under the same conditions.
Official references
- OpenAI, Update to GPT-5 System Card: GPT-5.2
- OpenAI API, GPT-5.2 model reference
- OpenAI, Introducing GPT-5.2
- OpenAI API, Responses API overview
Sources
Update to GPT-5 System Card: GPT-5.2
Update to GPT-5 System Card: GPT-5.2 — OpenAI Deployment Safety Hub
Update to GPT-5 System Card: GPT-5.2
Developer quickstart | OpenAI API