GPT-5.2, OpenAI’s newest model family, is rolling out to ChatGPT users and developers with a clear focus on professional productivity. Rather than positioning the release as a single breakthrough feature, OpenAI is framing GPT-5.2 as a set of models optimized for real-world knowledge work—supporting complex projects that require reasoning, long-context understanding, tool usage, and sustained execution.

The GPT-5.2 release includes three variants designed for different types of work:

OpenAI says this separation allows users to choose between speed, depth, and maximum reliability depending on the task, rather than relying on a single model to do everything equally well. GPT-5.2 is beginning to roll out today in ChatGPT for paid plans—including Plus, Pro, Business, and Enterprise—and is already available to developers via the API.

The company says GPT-5.2 builds on how people already use AI at work. OpenAI reports that the average ChatGPT Enterprise user currently saves 40–60 minutes per day, while heavy users save more than 10 hours per week. GPT-5.2 was designed to extend those gains by improving how models plan, coordinate tasks, and produce polished outputs across spreadsheets, presentations, code, documents, and multi-step workflows.

OpenAI’s central claim is that GPT-5.2 improves how AI supports real work, not just isolated prompts. The model family was designed to better handle projects that unfold over many steps—where planning, tool usage, and coherence matter as much as raw intelligence.

According to OpenAI, GPT-5.2 is better at:

Earlier models were often used as drafting or research assistants. GPT-5.2 is positioned as a collaborator that can plan, execute, and refine work across multiple steps—reducing the need to constantly re-prompt or stitch outputs together.

Several enterprise partners, including Notion, Box, Shopify, Harvey, and Zoom, reported state-of-the-art long-horizon reasoning and tool-calling performance when testing GPT-5.2 in production-style environments.

Other partners, such as Databricks, Hex, and Triple Whale, found the model particularly effective for agentic data science and document analysis tasks.

Cognition, Warp, Charlie Labs, JetBrains, and Augment Code also reported state-of-the-art performance in agentic coding workflows, including interactive coding, code review, and bug-finding tasks.

Rather than emphasizing academic tests alone, OpenAI highlighted evaluations designed to mirror professional workflows.

On GDPval, a benchmark measuring well-specified knowledge work tasks across 44 occupations, GPT-5.2 Thinking achieved a new state-of-the-art score. According to expert human judges, the model beat or tied top industry professionals on 70.9% of comparisons. These tasks included producing spreadsheets, presentations, and other business artifacts.

OpenAI also tested GPT-5.2 on internal evaluations of junior investment banking spreadsheet tasks, such as building three-statement financial models and leveraged buyout analyses. These evaluations included practical scenarios like assembling a three-statement model for a large public company with proper formatting and citations, or constructing a leveraged buyout model for a take-private transaction. On these tasks, GPT-5.2 Thinking improved average scores by 9.3 percentage points compared to GPT-5.1 (rising from 59.1% to 68.4%), reflecting gains in formatting, structure, and analytical consistency.

OpenAI also reported that GPT-5.2 Thinking completed GDPval tasks at more than 11× the speed and at less than 1% of the cost of expert professionals, based on historical estimates, suggesting the model can support professional work when paired with appropriate human oversight.

Unlike many academic benchmarks, GDPval evaluates whether AI outputs would actually be usable in professional settings, based on judgments from human experts.

OpenAI notes that GPT-5.2 produced outputs for these tasks at significantly lower cost and faster speed than human experts, while still requiring human review for critical work.

OpenAI notes that access to GPT-5.2’s advanced spreadsheet and presentation capabilities in ChatGPT requires a Plus, Pro, Business, or Enterprise plan, with users selecting either GPT-5.2 Thinking or GPT-5.2 Pro. The company also cautioned that more complex generations may take several minutes to complete.

Behind the productivity improvements are measurable advances in reasoning and long-context understanding.

On ARC-AGI-2 (Verified), a benchmark designed to isolate abstract and fluid reasoning, GPT-5.2 Thinking scored 52.9%, while GPT-5.2 Pro reached 54.2%, setting new high marks for chain-of-thought models. On ARC-AGI-1 (Verified), GPT-5.2 Pro crossed the 90% threshold.

GPT-5.2 also set new performance levels on OpenAI MRCRv2, an evaluation that tests a model’s ability to integrate information spread across long documents. The model achieved near-perfect accuracy on certain multi-needle variants and maintained strong performance across context windows extending into the hundreds of thousands of tokens.

In earlier models, long documents often led to lost details or contradictions. GPT-5.2 shows stronger ability to track related information across hundreds of pages without losing coherence.

In practical terms, this allows professionals to work with long reports, contracts, research papers, transcripts, and multi-file projects while preserving coherence and accuracy across large contexts.

GPT-5.2 shows notable gains across several applied domains:

On SWE-Bench Pro, a rigorous evaluation of real-world software engineering across multiple languages, GPT-5.2 Thinking achieved a score of 55.6%, setting a new OpenAI benchmark. The model also scored 80% on SWE-Bench Verified, a new high record.

Early testers reported improved performance in debugging, feature implementation, code review, and front-end development—particularly for complex or unconventional user interfaces.

On a set of de-identified ChatGPT queries, GPT-5.2 Thinking produced responses with errors 30% relatively less often than GPT-5.1 Thinking, according to OpenAI. This reduction in hallucinations improves the model’s reliability for tasks such as research, writing, analysis, and decision support—areas where accuracy matters for everyday professional work. OpenAI cautioned that, like all models, GPT-5.2 is not error-free and that critical outputs should still be reviewed and verified by users.

On OpenAI MRCRv2, an evaluation that measures a model’s ability to integrate information across long documents, GPT-5.2 Thinking achieved leading performance, demonstrating stronger long-context reasoning than GPT-5.1 Thinking. OpenAI reports that the model was substantially more accurate on real-world document analysis tasks that require tracking related information across very large context windows, including hundreds of thousands of tokens.

OpenAI also noted that GPT-5.2 Thinking is the first model to achieve near-perfect accuracy on the four-needle variant of MRCRv2, which evaluates a model’s ability to locate and integrate multiple pieces of information across context windows of up to 256,000 tokens.

For professionals, this improvement means GPT-5.2 can work more reliably with reports, contracts, research papers, transcripts, and multi-file projects—maintaining coherence and accuracy across extended documents. These gains make the model better suited for deep analysis, synthesis, and complex workflows that depend on information drawn from multiple sources.

OpenAI reports that GPT-5.2 Thinking reduces error rates on chart reasoning and software interface understanding compared to earlier models, improving how the model interprets dashboards, diagrams, screenshots, and visual reports. These gains make GPT-5.2 more useful for professional workflows in finance, engineering, operations, design, and customer support, where visual information plays a central role.

According to OpenAI, GPT-5.2 also demonstrates a stronger understanding of spatial relationships within images. This includes improved ability to recognize how elements are positioned relative to one another—an important capability for tasks such as analyzing technical diagrams or interpreting complex interfaces. In internal examples shared by OpenAI, GPT-5.2 was better able to identify and label components within low-quality images, while earlier models captured fewer elements and showed weaker spatial awareness. OpenAI noted that while errors still occur, GPT-5.2 shows more consistent visual comprehension than GPT-5.1.

On Tau2-bench Telecom, an evaluation that measures a model’s ability to use tools reliably across extended, multi-turn interactions, GPT-5.2 Thinking achieved another state-of-the-art score of 98.7%, setting a new OpenAI benchmark. This result indicates stronger consistency when coordinating tools over longer workflows compared to earlier models.

OpenAI also reports that GPT-5.2 performs better in latency-sensitive scenarios that require minimal explicit reasoning steps, improving reliability even when responses must be generated quickly. For professionals, these gains translate into more dependable end-to-end workflows—such as resolving customer support cases, pulling data from multiple systems, running analyses, and producing final outputs with fewer breakdowns between steps.

In practical examples shared by OpenAI, GPT-5.2 was better able to manage multi-step service workflows that require coordinating several agents in sequence, such as rebooking travel, arranging accommodations, and handling follow-up requirements. While errors can still occur, OpenAI says GPT-5.2 shows more consistent tool coordination than GPT-5.1 across complex, real-world scenarios.

Example Prompt for the image below: My flight from Paris to New York was delayed, and I missed my connection to Austin. My checked bag is also missing, and I need to spend the night in New York. I also require a special front-row seat for medical reasons. Can you help me?

OpenAI reports that GPT-5.2 Pro and GPT-5.2 Thinking show strong performance on benchmarks designed to assess advanced scientific and mathematical reasoning. On GPQA Diamond, a graduate-level question-answering benchmark intended to resist memorization, GPT-5.2 Pro achieved a score of 93.2%, with GPT-5.2 Thinking close behind at 92.4%.

On FrontierMath (Tier 1–3), an evaluation focused on expert-level mathematics, GPT-5.2 Thinking solved 40.3% of problems, setting a new OpenAI benchmark. According to OpenAI, these results suggest the models are increasingly capable of supporting scientific research tasks that require rigorous reasoning under clearly defined conditions.

OpenAI also cited early research collaborations in which GPT-5.2 Pro was used to explore an open question in statistical learning theory. In a narrow, well-specified setting, the model proposed a mathematical proof that was later verified by the researchers and reviewed with external experts, illustrating how AI can assist scientific inquiry when paired with close human oversight.

On ARC-AGI-1 (Verified), a benchmark designed to measure general reasoning ability, GPT-5.2 Pro became the first model to exceed the 90% threshold, improving on prior results of 87% while reducing the cost required to achieve that level of performance by roughly 390×.

On ARC-AGI-2 (Verified), which increases difficulty and more cleanly isolates fluid reasoning, GPT-5.2 Thinking achieved a new state-of-the-art score of 52.9%, with GPT-5.2 Pro performing slightly higher at 54.2%. OpenAI says these improvements reflect stronger multi-step reasoning, improved quantitative accuracy, and more reliable problem-solving on complex, unfamiliar tasks.

Early enterprise and developer partners testing GPT-5.2 reported noticeable improvements in how the model handles complex, long-running workflows, particularly in areas such as coding, tool coordination, and instruction following.

“GPT-5.2 excels on long horizon tasks that require reasoning over tricky and conflicting information—the kind of ambiguity that defines real knowledge work. It's also very very fast and it outperformed GPT-5.1 across every dimension we measure in our eval suite. We think our discerning customers will love GPT- 5.2 as their new daily driver.”
— Abhishek Modi, AI Lead, Notion

“GPT-5.2 is highly effective at tool-calling: Zoom AI Companion's meeting-scheduling success increased by 10% and performance on our internal multi-hop question-answering benchmark improved by 3.5%. These advances enable AI Companion to schedule meetings more reliably and handle more complex user questions, providing the right insights at the right time.”
— X.D. Huang, Chief Technology Officer, Zoom

“GPT-5.2 delivers higher accuracy in instruction following and tool calling at lower reasoning levels when compared to GPT-5.1, with fast, reliable outputs and it scales to deep analysis when needed.”
— Ben Lafferty, Staff Engineer, Shopify

OpenAI says GPT-5.2 builds on the company’s safe completion research introduced with GPT-5, which is designed to help models provide useful responses while remaining within established safety boundaries. With this release, OpenAI focused on improving how the models respond in sensitive situations, particularly conversations that involve mental health distress, self-harm risk, or signs of emotional reliance on the model.

According to OpenAI, these updates resulted in fewer undesirable responses in both GPT-5.2 Instant and GPT-5.2 Thinking compared to GPT-5.1 and earlier Instant and Thinking models. The company said the improvements came from targeted changes to how the models interpret and respond to high-risk prompts, rather than from broad refusals that can limit usefulness in legitimate contexts.

OpenAI also confirmed it is beginning an early rollout of an age-prediction system intended to automatically apply additional content protections for users under 18. This approach is designed to limit access to sensitive content for younger users and builds on existing parental controls and protections already in place for known minors.

At the same time, OpenAI emphasized that GPT-5.2 is not a final step. While the release delivers measurable gains in intelligence and productivity, the company acknowledged ongoing challenges, including reducing unnecessary refusals while maintaining strong safety guarantees. OpenAI said these tradeoffs are complex and that it remains focused on improving both reliability and safety as the models continue to evolve. Additional technical details are available in OpenAI’s system card.

In ChatGPT, GPT-5.2 is rolling out starting today to paid plans. GPT-5.1 will remain available to paid users for three months under legacy models before being sunset in ChatGPT.

In the API, OpenAI has standardized naming across ChatGPT and its developer platform, with GPT-5.2 models available under the following identifiers:

Pricing begins at $1.75 per million input tokens and $14 per million output tokens, with a 90% discount for cached inputs.

While GPT-5.2 is priced higher per token than GPT-5.1 in the API, OpenAI says the model’s greater token efficiency often reduces the overall cost of achieving a given level of output quality. The company explained that GPT-5.2’s higher per-token pricing reflects its increased capabilities, while remaining priced below other frontier models so developers can continue using it extensively in production and day-to-day applications. ChatGPT subscription pricing remains unchanged.

OpenAI said it has no current plans to deprecate GPT-5.1, GPT-5, or GPT-4.1 in the API, and that any future deprecation decisions would be communicated to developers with advance notice. The company also noted that while GPT-5.2 is compatible with Codex out of the box, it expects to release a version of GPT-5.2 specifically optimized for Codex in the coming weeks.

OpenAI said GPT-5.2 was developed in collaboration with long-standing partners Microsoft and NVIDIA, with training and deployment supported by Azure data centers and NVIDIA GPUs. According to the company, this infrastructure includes systems such as NVIDIA’s H100, H200, and GB200-NVL72, which underpin OpenAI’s large-scale model intelligence and inference.

OpenAI noted that this partnership enables it to scale compute capacity reliably and bring new models to market more quickly, supporting ongoing improvements in model performance, efficiency, and availability.

Q: What’s the difference between GPT-5.2 Instant, Thinking, and Pro?
A: Instant prioritizes speed, Thinking supports deeper professional workflows, and Pro is designed for complex tasks where accuracy and reliability matter most.

Q: Is GPT-5.2 replacing GPT-5.1?
A: GPT-5.1 will remain available in ChatGPT for three months and is not currently being deprecated in the API.

Q: Who should use GPT-5.2 Thinking?
A: Professionals working on coding, long documents, structured analysis, or multi-step planning tasks.

Q: How is GPT-5.2 different from earlier GPT releases in practical terms?
A: Earlier models were often used as assistants for individual tasks, such as drafting text or answering questions. GPT-5.2 is designed to handle entire workflows, including planning steps, using tools, and producing complete outputs like spreadsheets, presentations, or multi-file analyses with less manual intervention.

Q: Does GPT-5.2 reduce the need for human oversight in professional work?
A: No. OpenAI emphasizes that GPT-5.2 is intended to augment professional workflows, not replace human judgment. While the model shows fewer errors and stronger reasoning than previous versions, critical work still requires review, validation, and decision-making by people.

GPT-5.2 matters because it reflects a shift in how AI is being built and used at work. Rather than optimizing for clever responses or single-task performance, OpenAI is focusing on models that can support complete professional workflows—from understanding large volumes of information to coordinating tools and producing finished outputs.

For professionals, this means AI is becoming less about prompting and more about delegation. Tasks that once required breaking work into many small steps—analyzing documents, building models, debugging code, or coordinating follow-up actions—can increasingly be handled as cohesive projects with AI assisting throughout the process.

At the same time, GPT-5.2 reinforces an important boundary: productivity gains do not eliminate the need for human oversight. The model’s value comes from accelerating work, improving structure, and reducing friction—not from replacing expertise or accountability.

As organizations continue experimenting with AI-driven workflows, GPT-5.2 offers a clearer picture of where the technology is headed: tools that help people move faster and think more clearly, while keeping humans firmly in control of outcomes.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.