GPT-5 Pro
Last reviewed
May 18, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,661 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 18, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,661 words
Add missing citations, update stale details, or suggest a clearer explanation.
GPT-5 Pro is a large language model developed by OpenAI and the highest-capability variant of the GPT-5 model family. It is positioned above the standard GPT-5 and GPT-5 Thinking modes through the use of scaled, parallel test-time compute, which OpenAI describes as enabling the model to "think for ever longer" to produce more comprehensive and accurate answers.[^1] GPT-5 Pro first became available to subscribers of the ChatGPT Pro tier when GPT-5 launched on August 7, 2025,[^2][^3] and was added to the OpenAI API as the model gpt-5-pro on October 6, 2025, during OpenAI's DevDay 2025 developer conference in San Francisco.[^4][^5] The API release marked the first time OpenAI offered direct programmatic access to its most advanced production reasoning model.[^6]
The model is targeted at high-stakes, accuracy-sensitive workflows in domains such as finance, legal analysis, healthcare, scientific research, and complex software engineering, where the additional inference cost and longer response times are acceptable trade-offs for higher reliability.[^4][^7] In OpenAI's internal evaluations on more than 1,000 economically valuable real-world reasoning prompts, external expert raters preferred GPT-5 Pro responses to those produced by GPT-5 Thinking 67.8% of the time, and the model produced 22% fewer major errors than GPT-5 Thinking.[^8][^9]
OpenAI introduced the broader GPT-5 model family on August 7, 2025, presenting it as a unified system that combines the company's prior reasoning-focused models (the o-series, including OpenAI o3) and its general-purpose ChatGPT models into a single product hierarchy.[^2][^10] Under this hierarchy, GPT-5 in ChatGPT is exposed through an automatic router that selects between a fast non-reasoning model, a reasoning model ("GPT-5 Thinking"), and the high-capability "GPT-5 Pro" mode, which is reserved for ChatGPT Pro subscribers.[^11]
OpenAI has stated that GPT-5 Pro is intended to replace and surpass OpenAI o3-pro, the parallel-compute "pro" tier of the o3 reasoning model that was made available to ChatGPT Pro users earlier in 2025.[^8][^12] The August 2025 release made GPT-5 Pro available only inside the ChatGPT product, with no public API; this was identified at the time as a notable gap for developers who wanted programmatic access.[^11]
The API release on October 6, 2025, was announced by chief executive Sam Altman during the keynote of OpenAI's DevDay 2025, held at Fort Mason in San Francisco.[^4][^5] At the same event, OpenAI announced Sora 2 in the API, the Apps SDK for building applications inside ChatGPT, and the AgentKit platform for agent development.[^4][^13] Altman positioned GPT-5 Pro at DevDay as being designed for applications requiring "high accuracy and depth of reasoning," and singled out finance, legal services, and healthcare as flagship sectors.[^14]
OpenAI has published only limited technical detail about GPT-5 Pro's internal architecture. Public documentation and the GPT-5 system card describe the broader GPT-5 family as a unified reasoning system; GPT-5 Pro is specifically characterized as a configuration of the underlying GPT-5 reasoning model that uses scaled but efficient parallel test-time compute to evaluate multiple reasoning paths in parallel before producing a final answer.[^8][^9]
Within ChatGPT, the configuration corresponding to GPT-5 Pro is referred to internally as gpt-5-thinking-pro, and is described in OpenAI's GPT-5 launch materials as "a setting that makes use of parallel test-time compute."[^8] Independent technical write-ups describe the system as running multiple reasoning threads simultaneously and converging on a single answer based on best-of-many or related selection strategies, in contrast to the linear chain-of-thought used by the standard GPT-5 Thinking mode.[^9][^15]
In the API, the model is exposed as gpt-5-pro and defaults to (and only supports) the reasoning.effort: high setting, meaning developers cannot reduce its reasoning budget below the maximum.[^6][^16] OpenAI's API documentation notes that, because of this high reasoning effort, requests "may take several minutes to finish" and recommends that developers use the background mode of the Responses API to avoid request timeouts.[^6]
According to OpenAI's API documentation, GPT-5 Pro accepts text and image inputs and produces text-only outputs; it does not produce audio or video output and does not provide image generation.[^16] The model supports streaming and function calling, but does not support fine-tuning, predicted outputs, or the code interpreter tool, and it is offered only through OpenAI's Responses API rather than the older Chat Completions endpoint.[^6][^16] OpenAI states that the Responses-only deployment is intended to enable multi-turn model interactions and other advanced API features.[^6]
OpenAI's published specifications for the gpt-5-pro API model are as follows:[^16]
The 400,000-token context window and 272,000-token maximum output are larger than the limits originally exposed for GPT-5 in the API at launch (which used a smaller maximum output of 128,000 tokens).[^17] Within ChatGPT, the user-visible context window for GPT-5 Pro mode is smaller than the API limit, reported by OpenAI's help documentation as up to 128,000 tokens.[^11]
When made available in the API on October 6, 2025, GPT-5 Pro was priced at $15.00 per million input tokens and $120.00 per million output tokens.[^16][^4] At those rates, the model is approximately 10 to 12 times more expensive on input tokens than the standard GPT-5 model and roughly 12 times more expensive on output tokens.[^18] This made GPT-5 Pro the most expensive generally available text model in OpenAI's catalog at launch.[^18]
The model is available through OpenAI's standard, priority, batch, and flex service tiers, with priority processing charged at higher rates in exchange for reduced and more consistent latency.[^19]
GPT-5 Pro mode is restricted to subscribers of the ChatGPT Pro plan, which OpenAI prices at $200 per month.[^20][^11] Pro subscribers receive what OpenAI describes as unlimited access to standard GPT-5 and access to GPT-5 Pro mode; the lower-priced ChatGPT Plus and ChatGPT Free tiers do not include GPT-5 Pro mode.[^20][^11] ChatGPT Business, Enterprise, and Edu plans also expose GPT-5 Pro for eligible users in those tiers.[^11]
OpenAI reported GPT-5 Pro benchmark scores in the GPT-5 launch materials published on August 7, 2025, framing them as the model's "with extended reasoning" results.[^8] Selected scores from OpenAI's own reporting include:[^8][^21]
In OpenAI's own preference evaluations on more than 1,000 real-world reasoning prompts spanning health, science, mathematics, coding, and other domains, external experts preferred GPT-5 Pro outputs to those of GPT-5 Thinking 67.8% of the time, and GPT-5 Pro produced 22% fewer "major errors" than GPT-5 Thinking on those evaluations.[^8]
The third-party benchmark platform Artificial Analysis lists GPT-5 (high effort, the configuration most comparable to GPT-5 Pro in their methodology) with a 400,000-token context window versus 200,000 for o3-pro, and identifies GPT-5 (high) as competitive with or ahead of o3-pro across the ten benchmarks it evaluates, including GPQA Diamond, Humanity's Last Exam, and SciCode.[^23] Independent benchmark aggregators have also reported that GPT-5 Pro and its variants achieve higher scores on the AIME 2025, GPQA Diamond, SWE-bench Verified, and LiveCodeBench benchmarks than standalone o3 and o3-pro.[^24][^15]
OpenAI has cautioned that AIME results obtained with Python tool access "should not be compared directly to the performance of models without tool access," and frames those numbers as a demonstration of tool integration rather than a like-for-like accuracy comparison.[^21]
The GPT-5 system card, published by OpenAI in August 2025, describes safety evaluations for the GPT-5 model family. According to the document, hallucination evaluations placed gpt-5-thinking at a hallucination rate roughly 65% lower than OpenAI o3 on relevant test prompts, and approximately eight times fewer hallucinations than o3 on challenging conversations.[^25] OpenAI states that GPT-5 Pro is built on top of gpt-5-thinking with parallel test-time compute, and that the safety evaluations of gpt-5-thinking are treated as "strong proxies" for GPT-5 Pro, while certain risks are also evaluated separately for the Pro configuration because the additional compute "could materially impact the relevant risks."[^25]
OpenAI also reported that scores on its internal sycophancy benchmark fell from 0.145 for GPT-4o to 0.040 for gpt-5-thinking, with an A/B test in production showing a roughly 75% reduction in sycophantic responses for paid users.[^25] OpenAI engaged more than 400 external testers in the GPT-5 safety evaluation, contributing over 5,000 hours of red-teaming and review.[^25]
The broader GPT-5 release on August 7, 2025, received mixed reviews. Critics on social media reported factual errors, weak math results on certain prompts, and basic spelling mistakes, while developers and enterprise users tended to praise the model's instruction-following and well-structured code output.[^26] Researcher Ethan Mollick, who had early access, characterized GPT-5 as "very good" and described its behavior as "[just doing] stuff, sometimes amazing things, sometimes puzzling things, entirely on its own," while emphasizing that the "jagged frontier" of AI capability remained — with GPT-5 excelling at some tasks and surprisingly weak on others.[^27]
Independent developer and AI commentator Simon Willison described the underlying GPT-5 model as "competent" and "occasionally impressive" but not a major leap over GPT-4-class models.[^28] When GPT-5 Pro was added to the API in October 2025, Willison documented running a single image-generation test (an SVG of a pelican) that took 6 minutes and 8 seconds and cost approximately $1.10, illustrating both the cost profile and latency characteristic of the model.[^29]
At DevDay 2025, the API release of GPT-5 Pro was widely framed as one of the headline announcements, alongside Sora 2 in the API, the Apps SDK, and AgentKit.[^4][^13] ChatGPT had reached 800 million weekly active users by the time of DevDay 2025, doubling from approximately 400 million in February 2025, which OpenAI cited in the keynote as part of the rationale for opening high-capability models to developers.[^14][^4]
Trade publications described GPT-5 Pro at DevDay as "OpenAI's smartest and most expensive model," with multiple outlets noting that the input price was about ten times that of standard GPT-5 and that the model was positioned for "advanced enterprise and research use cases."[^18][^4] Coverage by InfoQ characterized the release as "the first time OpenAI has offered developers access to its very top-end reasoning model in production."[^4]
OpenAI named the financial-services AI startup Hebbia as a research partner for GPT-5 in financial workflows. Hebbia stated that GPT-5 enabled it to build three-statement financial models from unstructured documents, including SEC filings, Virtual Data Rooms, and provider databases such as S&P Capital IQ, PitchBook, and FactSet.[^30] Biotech company Amgen separately reported piloting GPT-5 in scientific and business workflows and found that the model "met the high bar" for scientific accuracy in early evaluations.[^31]
Reviewers have identified several limitations of GPT-5 Pro. The most consistent observation is latency: developer reports describe single requests taking several minutes to complete because the model is configured to use the highest reasoning effort, and the API documentation itself recommends background processing for many use cases.[^6][^29] A separate critical write-up by Geeky Gadgets characterized GPT-5 Pro as producing "robotic" outputs that "lack conversational tone," and argued that the model is best suited to highly structured analytical tasks rather than open-ended creative or conversational work.[^32]
Cost is another widely cited limitation: at $15 per million input tokens and $120 per million output tokens, GPT-5 Pro is substantially more expensive than the standard GPT-5 ($1.25 / $10 per million in the API at GPT-5's launch) and is generally only economical for tasks where additional accuracy materially changes downstream outcomes.[^16][^18]
GPT-5 Pro is the highest-capability and most expensive tier within the GPT-5 family. The standard GPT-5 model in the API was launched at $1.25 per million input tokens and $10 per million output tokens, with a 400,000-token context window and a smaller maximum output of 128,000 tokens.[^17] GPT-5 Thinking, the dedicated reasoning configuration of GPT-5, occupies a tier between standard GPT-5 and GPT-5 Pro in both capability and cost.[^15][^11]
Compared to OpenAI o3-pro — the prior generation's parallel-compute reasoning model — GPT-5 Pro shares the conceptual approach of scaling test-time compute, but is described by OpenAI as part of a unified GPT-5 reasoning system rather than as a member of a separate o-series.[^2][^8] OpenAI's evaluations published with the GPT-5 launch reported that GPT-5 with thinking achieves performance comparable to or better than OpenAI o3 using "50–80% less output tokens" on a range of capabilities, and Artificial Analysis lists GPT-5 (high) ahead of o3-pro on several frontier benchmarks.[^8][^23]
Within ChatGPT specifically, GPT-5 Pro is selectable for the most difficult tasks and long-running workflows, while users of lower tiers see only the standard GPT-5 with thinking modes determined by the GPT-5 router.[^11] OpenAI subsequently introduced higher-numbered models in the GPT-5 family, including GPT-5.1, GPT-5.2, GPT-5.4, and GPT-5.5, each accompanied by their own Pro variants at higher price points.[^33]