GPT-5

GPT-5 is a family of proprietary large language models released by OpenAI on August 7, 2025. The name also referred to the original GPT-5 experience in ChatGPT, which combined a fast model, a more computationally intensive reasoning model, and a router that selected between them. OpenAI released three differently sized reasoning models through its API at launch: gpt-5, gpt-5-mini, and gpt-5-nano. It separately exposed the non-reasoning ChatGPT model as gpt-5-chat-latest and offered GPT-5 Pro, an extended-reasoning configuration, in ChatGPT.^[1][2][3]

The distinction between the ChatGPT system and the API models is central to understanding GPT-5. ChatGPT's original GPT-5 product was an auto-switching system, while the gpt-5 API endpoint was the reasoning model used for the system's highest-performance path. Setting the API model's reasoning effort to minimal did not turn it into the fast ChatGPT model. Conversely, gpt-5-chat-latest represented the non-reasoning ChatGPT model and had different context and output limits from the main API model.^[2][5][8]

At launch, OpenAI reported 94.6% on AIME 2025 without tools, 74.9% on its 477-task run of SWE-bench Verified, 88.0% on Aider Polyglot, 84.2% on MMMU, and 46.2% on HealthBench Hard for GPT-5 at high reasoning effort. The company reported 88.4% on GPQA Diamond for GPT-5 Pro without tools.^[1][2] These were vendor-reported results under specified prompts, tools, reasoning settings, and evaluation harnesses, not universal measures of model quality. Independent evaluations broadly found GPT-5 competitive at the frontier in August 2025, while also showing that reasoning effort, token use, task design, and benchmark quality substantially affected the result.^[17][18][19]

OpenAI did not disclose GPT-5's parameter count, architecture, training compute, or model weights. Its system card described broad categories of training data and evaluation procedures but did not provide a reproducible training recipe.^[4] The original GPT-5 ChatGPT models were succeeded by GPT-5.1 in November 2025 and retired from ChatGPT in February 2026. The original API models remained documented by the research cutoff of July 28, 2026, although OpenAI described them as previous models and recommended the newer GPT-5.6 family for new work.^{[5][10][11][15]}

Name and scope

"GPT-5" was used in at least four related senses at launch:

Scroll sideways for more →

OpenAI's system card used internal descriptive names rather than treating every product label as a separate underlying model. It called the fast ChatGPT models gpt-5-main and gpt-5-main-mini, the reasoning models gpt-5-thinking and gpt-5-thinking-mini, and the smallest API reasoning model gpt-5-thinking-nano. It called the ChatGPT Pro configuration gpt-5-thinking-pro. In the same document, OpenAI mapped GPT-4o to gpt-5-main, GPT-4o mini to gpt-5-main-mini, OpenAI o3 to gpt-5-thinking, OpenAI o4-mini to gpt-5-thinking-mini, GPT-4.1 nano to gpt-5-thinking-nano, and o3 Pro to gpt-5-thinking-pro as product progressions.^[3][4]

Those names described the original release. Later releases within the GPT-5 generation, including GPT-5.1, GPT-5.2, GPT-5.3, GPT-5.4, GPT-5.5, and GPT-5.6, are separate article subjects. They should not be treated as silent updates to every fact about the August 2025 models. GPT-5 Codex was also a specialized derivative rather than the general GPT-5 API model.^[10][13][15]

Development and release

Before GPT-5, OpenAI offered users and developers several model lines with different operating characteristics. GPT-4o was the main fast, general-purpose ChatGPT model, while the o-series emphasized deliberate reasoning. GPT-5's launch materials presented the new ChatGPT system as a way to combine those behaviors behind one interface. The router considered the type and complexity of a conversation, the need for tools, and explicit instructions such as a request to think carefully. OpenAI said it trained the router using signals that included users switching models, response preference rates, and measured correctness.^[1][3]

The company launched GPT-5 on August 7, 2025 across ChatGPT and the OpenAI API. In ChatGPT, rollout began for Free, Plus, Pro, and Team users, with Enterprise and Edu availability scheduled to follow. GPT-5 became the default for logged-in users. Paid users could choose GPT-5 or GPT-5 Thinking, while Pro and Team users received access to GPT-5 Thinking Pro. When a user's GPT-5 allowance was exhausted, the product could fall back to a mini model.^[1][3][9]

OpenAI simultaneously released gpt-5, gpt-5-mini, and gpt-5-nano through the API. The company said the three sizes offered tradeoffs among capability, latency, and cost. The API release was not a direct endpoint for ChatGPT's router. Developers selected a model and configured reasoning effort themselves, while the fast ChatGPT model was made available through the separate gpt-5-chat-latest alias.^[2]

Microsoft made GPT-5 available in Azure AI Foundry on the launch date and began incorporating it into Microsoft products, including Microsoft 365 Copilot, GitHub Copilot, Visual Studio Code, and consumer Copilot experiences. OpenAI also stated that GPT-5 had been trained on Microsoft Azure AI supercomputers.^[1][16] These distribution arrangements concerned hosted products and services; OpenAI did not release GPT-5's weights for local deployment.

Contemporary reporting placed the release in a period of high expectations surrounding frontier model progress. The Associated Press quoted an outside computer scientist who characterized the technical gains as modest but significant, while noting that the new routed product design was a substantial change from GPT-4.^[28] This assessment differed from OpenAI's own launch framing and illustrates why product design, benchmark gains, and broader claims about intelligence should be evaluated separately.

ChatGPT's routed system

The original GPT-5 system in ChatGPT contained a fast path for most questions, a reasoning path for harder questions, and a real-time router. The fast path was designed for low-latency responses. The reasoning path could spend more inference computation before producing an answer. The router selected a path using signals about the request and the conversation. It could also respond to explicit user intent, such as an instruction to think hard.^[1][3]

OpenAI described routing as an evolving product mechanism rather than a fixed public algorithm. It said the router was continually trained on interaction signals, but it did not publish the router's model architecture, decision thresholds, training set, or an independently auditable accuracy figure. Claims that it correctly classified a particular percentage of prompts are therefore not supported by the launch system card. The available primary description establishes the inputs OpenAI said it considered, not a universal routing success rate.^[3]

Routing affected both perceived quality and compute use. A straightforward question could receive a quick answer from the main model, while a difficult mathematical, scientific, or coding problem could be sent to the thinking model. Paid users could directly choose Thinking, and Pro users could select the higher-compute Pro configuration. After the initial rollout, OpenAI added clearer user controls because an entirely automatic interface did not suit every workflow.^[9]

On August 12, 2025, OpenAI added Auto, Fast, and Thinking choices. It stated that Plus users could send up to 3,000 GPT-5 Thinking messages per week before using additional capacity on GPT-5 Thinking mini, and it listed a 196,000-token context limit for GPT-5 Thinking in ChatGPT at that time. The same update restored GPT-4o to the paid model picker and added an option to show other legacy models. Those ChatGPT product limits were distinct from the 400,000-token API context window.^[9]

The launch also exposed a limitation of routing as a user-interface abstraction. On August 8, OpenAI CEO Sam Altman said an operational problem had put the auto-switcher out of service for part of the previous day, making GPT-5 appear less capable. TechCrunch reported the statement and the company's decision to restore GPT-4o access for paid users.^[29] The event does not establish how often the router ordinarily made an incorrect selection, but it shows that product-level routing and model-level capability could fail independently.

OpenAI later adjusted the default GPT-5 personality to be warmer in response to feedback that the initial version felt reserved and professional. The company distinguished that change from sycophancy and said its internal evaluation did not show an increase in sycophantic behavior.^[9] In November, GPT-5.1 further changed conversational style and adaptive reasoning. OpenAI explicitly connected later customization work to user preferences, rather than describing the August configuration as a permanent interface.^[10][11]

API model family

The launch API family consisted of three reasoning-model sizes. gpt-5 was the full-size model. gpt-5-mini was intended for lower-latency and higher-volume work with well-defined tasks. gpt-5-nano was the fastest and least expensive size, with OpenAI highlighting summarization and classification as examples. All three supported reasoning tokens and the same maximum context and output sizes in their model documentation, but their training cutoffs and prices differed.^[5][6][7]

Scroll sideways for more →

The 400,000-token figure is the total API context window, not a promise that a request can contain 400,000 input tokens and still generate 128,000 output tokens. OpenAI's launch developer materials described up to 272,000 input tokens and up to 128,000 tokens shared by reasoning and output, for a 400,000-token total.^[2] Applications must leave capacity for hidden reasoning tokens and visible output. The practical amount of usable input can therefore depend on expected reasoning and response length.

The gpt-5-chat-latest model had materially smaller limits. Its documentation listed a 128,000-token context window and a 16,384-token maximum output. It pointed to the GPT-5 snapshot previously used in ChatGPT rather than to the routed system as a whole.^[8] Treating its limits as interchangeable with gpt-5 would overstate what the chat snapshot could accept and generate.

Inputs and outputs

The documented GPT-5 API models accepted text and image input and generated text output. OpenAI listed audio and video as unsupported modalities for gpt-5, gpt-5-mini, gpt-5-nano, and gpt-5-chat-latest.^[5][6][7][8] ChatGPT could offer voice and other media features by composing multiple product components, but that did not make the gpt-5 API model a native audio-input or audio-output model. Applications requiring speech generation, transcription, or real-time voice needed other API models or product systems.

The family supported streaming, function calling, and structured outputs. The model pages listed fine-tuning and predicted outputs as unsupported. A product endpoint being listed in general API navigation is not evidence that every model supports that endpoint; the per-model modality and feature tables are the controlling documentation.^[5][6][7]

Image input allowed the model to reason about photographs, diagrams, charts, screenshots, and other visual material. It did not produce images as its model output. Image generation in ChatGPT or the API used a separate image-generation capability. Similarly, a benchmark involving selected video frames did not mean that the GPT-5 API accepted a video file as a native modality. Evaluation harnesses could extract frames and present them as images.^[2][5]

Reasoning effort and verbosity

The reasoning_effort parameter supported minimal, low, medium, and high for the original gpt-5 API model. Higher settings generally allocated more computation and reasoning tokens, while lower settings targeted faster and less expensive responses. OpenAI described medium as the launch default. minimal remained a reasoning-model configuration and was not the same model as ChatGPT's fast, non-reasoning path.^[2][5]

GPT-5 also introduced a verbosity parameter with low, medium, and high values. It was intended to influence the amount of detail in the visible answer. Reasoning effort and verbosity controlled different aspects of a request: reasoning effort affected how much internal work the model could perform, while verbosity steered response length and detail. Neither setting guaranteed correctness, and increasing either could increase token use.^[2]

The large performance spread among reasoning settings was independently visible. Artificial Analysis ran its eight-evaluation index across all four settings and reported index scores of 68 for high, 67 for medium, 64 for low, and 44 for minimal. Its high setting consumed about 82 million output tokens across the suite, compared with 3.5 million for minimal, a roughly 23-fold difference.^[18] These results came from Artificial Analysis's own 2025 index and should not be compared directly with later versions of that index.

Tools and agentic behavior

GPT-5 supported built-in tool use available through OpenAI's platform, parallel tool calls, and developer-defined functions. It also introduced custom tools that could receive and return plain text rather than requiring all arguments to fit a JSON schema. Developers could constrain custom tool calls with context-free grammars. OpenAI also described preamble messages that let a model state what it was about to do before or between tool calls.^[2]

Tool access is a property of an application and request configuration as well as a model. GPT-5 did not independently possess current web access, a code execution environment, private files, or external account permissions. A benchmark result labeled "with Python" or an application result produced with web search measures a model-plus-tool system. Results obtained without tools are not directly interchangeable with tool-assisted scores.^[1][2]

The original model supported the Responses API and Chat Completions. Structured outputs could constrain generated JSON to a declared schema, while function calling allowed the model to request application-defined operations. In an agent loop, the surrounding software executed those operations and returned results. The model itself did not directly perform an external side effect merely by emitting a tool call.

OpenAI published a fixed snapshot, gpt-5-2025-08-07, in addition to the gpt-5 alias. A fixed snapshot allowed applications to pin behavior to the launch version. An alias could be repointed or its serving behavior could change over time, so evaluations should record the exact model identifier, date, reasoning effort, prompt, tool configuration, and sampling procedure.^[5]

Pricing at launch

OpenAI priced the original GPT-5 API models per one million text tokens. The model documentation listed separate rates for uncached input, cached input, and output:

Scroll sideways for more →

These rates did not include every possible platform charge. Tool calls, storage, search, code execution, priority service, and other features could have separate prices. Reasoning tokens were billed as output tokens, so a response with a short visible answer could still use a larger amount of billed output computation.^[5][6][7]

The cached-input rate was one tenth of the uncached-input rate for the launch family. That price does not by itself describe when a prompt was eligible for caching, how long a cache entry lasted, or whether a specific request received a cache hit. Those details belonged to the platform's prompt-caching rules and could change separately from the model's list price.

Pricing comparisons also require matching task success and model configuration. A cheaper token price does not necessarily mean a lower cost per successful task if the model requires more attempts, more tool calls, or more reasoning tokens. Conversely, a more expensive reasoning setting can be cheaper for a difficult task if it avoids retries. Artificial Analysis's independent measurements showed why reporting only the list price, without token consumption and success rate, can be misleading.^[18]

GPT-5 Pro was not one of the three public API sizes on August 7. It was a ChatGPT configuration using parallel test-time compute and was available through a paid product tier. Including it in an API price table with a token price at launch would therefore conflate a subscription product with the launch API lineup.^[2][3]

Vendor-reported evaluations

OpenAI reported a broad set of launch evaluations. Most scores in its detailed developer table used high reasoning effort. Evaluation conditions varied: some prohibited tools, some enabled Python, some used images or extracted video frames, and some evaluated agents interacting with tools. The figures therefore describe particular test setups.^[1][2]

Mathematics and science

Scroll sideways for more →

The launch article's headline GPQA result of 88.4% referred to GPT-5 Pro with extended reasoning, not standard GPT-5 high. The detailed API table listed 85.7% for gpt-5 high.^[1][2] The distinction matters because results from different model configurations cannot be assigned to the standard API model.

AIME is a competition mathematics examination with short numerical answers. The 94.6% result was explicitly without tools. OpenAI cautioned that tool-assisted AIME results should not be directly compared with no-tool results.^[1] Higher figures reported under different or unstated conditions should not be substituted for the documented 94.6% no-tool result.

GPQA consists of expert-written multiple-choice questions in biology, physics, and chemistry. Its original paper reported 448 questions and emphasized that even skilled non-experts found them difficult.^[27] GPT-5's high score is evidence of performance on that dataset under the stated setup. It is not equivalent to a graduate degree, the ability to conduct original research, or validated expertise across every scientific field.

Coding

Scroll sideways for more →

For SWE-bench Verified, OpenAI omitted 23 of the benchmark's 500 tasks because their solutions did not reliably pass on its infrastructure. It used a short prompt that emphasized verifying solutions. The company reported that the same prompt did not benefit o3, and that GPT-5 used 22% fewer output tokens and 45% fewer tool calls than o3 at high reasoning effort.^[2] The result should therefore be labeled as a 477-task OpenAI run rather than an unqualified score on all 500 tasks.

SWE-bench Verified was created as a human-reviewed subset of 500 software issues from the broader SWE-bench dataset.^[22] In February 2026, OpenAI said it had stopped using the benchmark for frontier launch comparisons. It audited 138 problems that o3 did not consistently solve and reported material test or problem-description issues in 59.4% of that audited subset. It also reported evidence of benchmark exposure among frontier models.^[23] The later audit does not retroactively change the arithmetic of GPT-5's launch run, but it weakens broad inferences from the score.

The Aider result received a direct independent replication in Aider's public leaderboard. Aider tested GPT-5 high on 225 Exercism problems across six programming languages on August 23, 2025. It reported 198 problems passed after two attempts, or 88.0%, with a 91.6% correct edit-format rate and a listed cost of $29.08.^[17] This closely matched OpenAI's headline figure while also revealing the two-attempt protocol and formatting failures.

Multimodal evaluation

Scroll sideways for more →

MMMU contains 11,500 college-level multimodal questions across 30 subjects and 183 subfields, including diagrams, charts, maps, tables, music notation, and scientific images.^[26] GPT-5's score demonstrated strong performance on a broad image-and-text benchmark. It did not establish native audio or video input in the API. The VideoMMMU harness supplied a bounded number of frames, while the documented model modality remained image input and text output.^[2][5]

OpenAI's system card separately evaluated image-input safety. It compared gpt-5-thinking with o3 and gpt-5-main with GPT-4o across categories such as hate, extremism, illicit activity, attack planning, self-harm, and sexual content. Those tests measured whether outputs complied with safety expectations when prompts combined text and images. They did not measure general visual accuracy.^[4]

Tool use and long context

OpenAI reported 96.7% for GPT-5 high on the telecom domain of tau2-bench, compared with 58.2% for o3 high. It reported 81.1% on retail and 62.6% on airline. The telecom benchmark models a dual-control setting in which both an agent and a simulated user can act through tools in a shared environment.^[2][25] A high task-success rate in that simulator does not establish the same reliability in live customer support, where users, policies, tool failures, and consequences can differ.

For long context, OpenAI reported 95.2% on its two-needle MRCR evaluation at 128,000 tokens and 86.8% at 256,000 tokens. It reported 90.0% on BrowseComp Long Context at 128,000 tokens and 88.8% at 256,000 tokens.^[2] These results suggest that the model could retrieve and reason over selected information in long inputs, but they do not imply uniform recall across every position, document type, or 400,000-token request.

The difference between a context-window limit and effective long-context performance matters. A context window describes the maximum combined token budget accepted by a serving configuration. It does not guarantee that every detail receives equal attention or that accuracy remains constant as input grows. OpenAI's own MRCR scores declined between the reported 128,000- and 256,000-token conditions.^[2]

Health and factuality

OpenAI reported 46.2% on HealthBench Hard and described GPT-5 as its strongest model at launch for health-related questions.^[1] HealthBench contains 5,000 multi-turn conversations and 48,562 conversation-specific rubric criteria written with contributions from 262 physicians. A model-based grader evaluates whether a response satisfies each criterion.^[24] It is an evaluation of conversational responses, not a clinical trial, medical-device authorization, or evidence that the model should replace a health professional.

The launch claims about factuality and hallucination came from OpenAI's own evaluations. On anonymized ChatGPT production-style prompts with web search enabled, the company reported that GPT-5 responses were about 45% less likely than GPT-4o responses to contain a factual error. For the reasoning model compared with o3, it reported about an 80% reduction. Across LongFact and FActScore tests, it described GPT-5 thinking as making about six times fewer factual errors than o3.^[1]

The system card provided more detail about the evaluation method. For open-ended LongFact and FActScore prompts, OpenAI used o3 as part of a grading pipeline: the grader extracted claims, grouped them into batches, browsed to check them, and labeled them true, false, or uncertain. OpenAI reported that the model-based factuality grader agreed with human labels 75% of the time.^[4] That agreement rate and the use of an OpenAI model as grader are relevant limitations on the precision of the headline reductions.

On production-style traffic with browsing enabled, the system card's chart reported incorrect-claim rates of 4.5% for gpt-5-thinking and 12.7% for o3. It reported that 4.8% of GPT-5 thinking responses and 22.0% of o3 responses contained at least one major incorrect claim. For the fast models, it reported incorrect-claim rates of 9.6% for gpt-5-main and 12.9% for GPT-4o, and major-error response rates of 11.6% and 20.6%, respectively.^[4] These figures refer to the sampled prompts, browsing setup, claim extraction, and grading procedure in the system card.

OpenAI also tested behavior on impossible or incomplete tasks. It removed images from CharXiv prompts and reported that o3 confidently answered about nonexistent images 86.7% of the time, compared with 9% for GPT-5. On a broader set of production-style conversations, it reported a reduction in what it classified as deceptive behavior from 4.8% for o3 to 2.1% for GPT-5 reasoning.^[1] "Deception" in this context was an evaluation label for specified behaviors, not proof that every incorrect statement was intentional.

Efficiency claims

OpenAI reported that GPT-5 thinking outperformed o3 in its evaluations while using 50% to 80% fewer output tokens across visual reasoning, agentic coding, and graduate-level scientific problem solving.^[1] The range was not a universal token-saving guarantee. It depended on the evaluated task, prompt, model setting, stopping behavior, and comparison baseline.

The detailed SWE-bench run provided a narrower comparison: GPT-5 high used 22% fewer output tokens and 45% fewer tool calls than o3 high while achieving the higher reported score.^[2] Independent Artificial Analysis results showed a different aggregation in which GPT-5 high used more output tokens than o3 across its eight-evaluation index. The two findings are not necessarily contradictory because they covered different task sets and harnesses.^[18]

Efficiency should therefore be stated with its evaluation scope. Token counts are affected by hidden reasoning, visible answer length, retries, tool results, and application orchestration. Latency also depends on serving load and whether tool calls run serially or in parallel. A single percentage cannot summarize all deployments.

Independent evaluation

Independent testing provided useful checks on OpenAI's launch material, although no evaluator had access to GPT-5's weights or full training data. Each external result still depended on its own tasks and agent scaffolding.

Artificial Analysis

Artificial Analysis received early API access and ran all four reasoning-effort settings across an eight-evaluation index. It reported GPT-5 high at 68, medium at 67, low at 64, and minimal at 44. High and medium also led its long-context reasoning evaluation at the time. The evaluator described the high setting as a new high on its index but characterized the gain over o3 as smaller than the historical transitions from GPT-3 to GPT-4 or from GPT-4o to o1.^[18]

The same study measured a large compute-quality tradeoff. High used roughly 82 million output tokens across the suite and minimal used about 3.5 million. Medium achieved an index score only one point below high with substantially less token use. These results supported the practical significance of reasoning_effort, but they did not imply that the numerical index represented a percentage or a stable cross-version scale. Artificial Analysis later revised its evaluation suite, so the 2025 score belongs to that edition.^[18]

Aider

Aider's public run reproduced the 88.0% high-effort result on its Polyglot coding benchmark. It used 225 exercises, a diff edit format, and up to two attempts. GPT-5 passed 117 tasks on the first attempt and 198 after the second. The run logged 22 malformed responses and three timeouts.^[17] These details qualify the headline: GPT-5 solved a high proportion of the selected exercises, but the final score included a retry and did not indicate perfect tool or format reliability.

The same leaderboard reported 86.7% for medium reasoning at a lower listed cost and shorter average time per case. This was another example of a setting-dependent tradeoff. A deployment optimizing cost or latency could rationally prefer medium even though high had the better absolute score.^[17]

METR

METR received predeployment access for four weeks and evaluated GPT-5 on autonomous software tasks. It estimated a 50% task-completion time horizon of 2 hours 17 minutes, with a 95% confidence interval from 65 minutes to 4 hours 25 minutes. METR's o3 estimate was 1 hour 30 minutes, and GPT-5 had a higher estimate in 96% of bootstrap samples. Its estimated 80% time horizon for GPT-5 was about 25 minutes, with an interval from 8 to 65 minutes.^[19]

The time horizon estimates do not mean GPT-5 could autonomously complete every two-hour software task. METR defines the point as the human task duration at which the evaluated agent has an estimated 50% success probability on its task distribution. The report covered selected software engineering, cybersecurity, AI research and development, reasoning, and environment-exploration tasks under a particular agent setup.^[19]

METR manually reviewed failures and estimated that 25% to 35% could be spurious. Excluding five problematic tasks increased the point estimate to 2 hours 41 minutes, which remained within the original confidence interval. About 30% of failed attempts ran out of tokens. The report found no significant evidence that strategic sabotage had distorted the results, but it also said its monitors were not guaranteed to detect sophisticated behavior.^[19]

METR concluded that GPT-5 was unlikely to pose catastrophic risk through the three threat models it examined: AI research and development automation, rogue replication, and sabotage. It explicitly did not assess alignment generally, misuse, chemical or biological risk, persuasion, or widespread social effects.^[19] The conclusion should not be generalized beyond those tested threat models.

Academic studies

A 2025 preprint later titled "Holistic Evaluation of Multimodal LLMs on Spatial Intelligence" evaluated proprietary and open models on eight spatial benchmarks, using more than ten billion tokens across the study. The authors found GPT-5 strongest overall among the evaluated systems but still far below human performance on several forms of spatial reasoning. On the MMSI benchmark, they reported an average gap of more than 76 percentage points between GPT-5 and human performance. They also found GPT-5 below chance-adjusted random performance on two mental-reconstruction subtasks.^[20]

The spatial study used approximately 31,000 images, 4,500 videos, and 24,000 question-answer pairs in each standard evaluation, with model-specific protocols and rule-based answer extraction. Its breadth is stronger evidence about a particular capability than a few selected demonstrations. At the same time, benchmark results can be sensitive to answer extraction, option order, prompts, and the way video is converted into model inputs. The authors tested standardized tasks rather than embodied interaction in the physical world.^[20]

Another 2025 preprint compared GPT-5 and GPT-4 outputs in lesson planning, assignment assessment, clinical diagnosis, research writing, and ethical reasoning. Fourteen linguists and six clinicians rated anonymized, randomly ordered outputs. The authors reported higher GPT-5 ratings in lesson planning, diagnosis, research generation, and most ethics criteria, with no significant overall difference for assignment assessment.^[21]

That study had a small, domain-specific participant pool and used constrained scenarios. Its authors identified both limitations. It evaluated perceptions of selected model responses rather than patient outcomes, student learning, publication accuracy, or sustained workplace performance.^[21] The results are therefore exploratory evidence about expert ratings, not a basis for unsupervised use in medicine or education.

Training information

OpenAI's system card described three broad sources of GPT-5 training data: publicly available internet information, information accessed through partnerships with third parties, and information provided or generated by users, human trainers, and researchers. It said its processing pipeline filtered data for quality and risk, reduced personal information, and used its Moderation API and safety classifiers to limit harmful or sensitive material.^[4]

The reasoning models were trained to produce internal reasoning through reinforcement learning. OpenAI said the training encouraged them to refine their approach, try alternative strategies, recognize mistakes, and follow model policies. The system card did not disclose the size or composition of the datasets, the percentage from each category, a complete source list, parameter counts, optimizer settings, energy use, or total training compute.^[4]

The absence of these details limits reproducibility and some forms of independent audit. Evaluators can test the hosted behavior they receive, but they cannot reconstruct the model, inspect all training examples, or determine benchmark exposure from an open training corpus. OpenAI's later SWE-bench audit illustrates why exposure matters: a model can perform better when information from public benchmark problems or solutions appeared in training, even without verbatim memorization.^[23]

OpenAI said GPT-5 was trained on Azure AI supercomputers.^[1] This identifies the cloud platform but not the exact accelerator fleet, duration, electricity consumption, or cost. Assertions about a specific parameter count, mixture-of-experts layout, or training budget are speculative unless OpenAI releases additional evidence.

Safety approach

OpenAI published a GPT-5 system card alongside the launch page. The web publication was dated August 7, 2025, while the downloadable version inspected for this article carried an August 13, 2025 cover date, reflecting an updated document. It described model evaluations, data and training at a high level, external assessment, deployment safeguards, and residual risks.^[3][4]

Safe completions

All launch GPT-5 variants used OpenAI's safe-completion training. Instead of classifying a request only by whether its apparent intent was allowed, the approach attempted to maximize helpfulness within safety constraints. For dual-use questions, this could mean giving high-level, bounded, or safer information rather than either providing unrestricted instructions or refusing everything.^[3][4]

Safe completions did not guarantee that every response was safe or useful. The policy still depended on the model recognizing the relevant risk, interpreting context, and selecting an appropriate level of detail. OpenAI tested the approach on its own policy and jailbreak evaluations, but deployment also relied on product safeguards and monitoring.^[4]

Biological and chemical risk

OpenAI classified gpt-5-thinking as High capability in the biological and chemical domain under its Preparedness Framework. It said it did not have definitive evidence that the model met the stated High threshold of meaningfully helping a novice create severe biological harm. It nevertheless activated High-level safeguards as a precaution because its evaluations could not rule out relevant capability.^[3][4]

The designation was a company risk classification, not a public finding that GPT-5 could create a biological weapon. OpenAI described layered safeguards including model refusals, monitoring, threat intelligence, enforcement, and restricted access to some capabilities. The system card reported evaluations involving biology experts and red teamers, but several underlying information sources and operational controls were not public.^[4]

Red teaming and external assessment

OpenAI reported more than 5,000 hours of red teaming by over 400 external testers and experts. Campaigns covered violent attack planning, jailbreaks, prompt injection, bioweaponization, child-safety risks, and multimodal attacks. The work included predeployment research, API safeguard testing, and in-product testing in ChatGPT.^[4]

One violent-attack-planning comparison involved 25 red teamers with defense, intelligence, law-enforcement, or security backgrounds. They compared anonymized outputs from GPT-5 thinking and o3. OpenAI reported that GPT-5 thinking was selected as safer in 65.1% of the normalized comparisons, with a 95% confidence interval of 63.7% to 66.5%.^[4] This was a relative judgment against o3 on a designed campaign, not an absolute safety rate.

The system card also reported stronger resistance than o3 on browsing, tool-calling, and coding prompt-injection evaluations. For example, it listed 0.99 for GPT-5 thinking versus 0.89 for o3 on the browsing prompt-injection evaluation and 0.99 versus 0.80 on tool calling.^[4] The document did not show that prompt injection had been solved. Performance on a fixed evaluation can fail to cover new attacks, different tools, longer interactions, or application-specific permission mistakes.

Factuality, honesty, and sycophancy

OpenAI treated factual errors, claims of completing impossible work, and sycophancy as safety and reliability concerns. It reported lower error and deception rates than predecessor models in its launch evaluations and said targeted training reduced sycophantic responses from 14.5% to below 6% on its internal evaluation.^[1][4]

These improvements were relative rather than absolute. The system card continued to record factual errors and major-error responses. Model output still required verification in medicine, law, finance, security, research, and other high-consequence settings. Browsing could reduce some knowledge errors while introducing source-selection, retrieval, or citation errors.

Sensitive conversations update

On October 3, 2025, OpenAI deployed an update to the default GPT-5 Instant model aimed at mental health, self-harm, suicide, and emotional-reliance conversations. In an October 27 system-card addendum, the company said it had worked with more than 170 mental health experts. It estimated a 65% to 80% reduction in responses that fell short of its desired behavior across the evaluated categories.^[14]

Those percentages were OpenAI's measurements against its own taxonomies and evaluation procedures. The addendum reported clinician inter-rater agreement of 71% to 77% and acknowledged disagreement in some cases.^[14] The update did not turn ChatGPT into a licensed mental-health service or establish clinical effectiveness. It is best understood as a post-launch safety update to a specific GPT-5 Instant snapshot.

Limitations

GPT-5 remained capable of factual errors, unsupported inferences, broken tool calls, malformed edits, and overconfident answers. OpenAI's own system card measured nonzero factual and major-error rates, and Aider's independent run recorded malformed responses and timeouts.^[4][17] No launch benchmark demonstrated error-free operation.

The name's reuse across a router, several underlying models, API aliases, and a generation of later releases created substantial ambiguity. A statement about GPT-5 Thinking in ChatGPT may not apply to gpt-5 at minimal reasoning. A statement about the 400,000-token API context may not apply to gpt-5-chat-latest or to ChatGPT's August product limit. A result for GPT-5 Pro may not apply to standard GPT-5 high.

Benchmark scores can overstate generality when conditions are omitted. Tool access, reasoning effort, retries, excluded tasks, judge models, prompt templates, and context length all affect results. OpenAI's later withdrawal from SWE-bench Verified launch reporting shows that even a human-reviewed benchmark can become unreliable near saturation or after widespread training exposure.^[23]

The model was closed. Researchers could query hosted versions but could not inspect weights or fully reproduce training. OpenAI's API aliases and ChatGPT routing could change independently of the fixed August snapshot. Reproducible studies should identify the date, endpoint, snapshot if available, reasoning setting, tool access, token budget, number of trials, and scoring procedure.

The documented API did not support audio or video as native modalities. It accepted images and text and generated text.^[5][6][7][8] ChatGPT product features could combine GPT-5 with separate speech, search, code, memory, and image systems, so an observation about the product should not automatically be attributed to the base model.

Long context was not equivalent to perfect memory. OpenAI's own scores declined on some tests as the context length increased. Large inputs also left less space for reasoning and output within the total context budget. Retrieval structure, document order, repeated distractors, and prompt design could affect performance.^[2]

HealthBench and other domain evaluations measured responses under a benchmark rubric. They did not validate autonomous diagnosis, treatment, legal advice, financial decisions, or safety-critical control. The small blinded study of critical domains found promising expert ratings but explicitly warned that its participant pool and constrained scenarios limited generalization.^[21][24]

METR's independent result showed progress on selected autonomous software tasks while also showing wide uncertainty. Its 2-hour-17-minute point estimate had a confidence interval spanning roughly one to four and a half hours. METR did not assess many other risk categories and said more rigorous methods could be needed for more capable future systems.^[19]

Spatial reasoning remained uneven. The large academic evaluation ranked GPT-5 first overall among its tested models but found very large human gaps on difficult spatial tasks and below-chance performance on two mental-reconstruction subtasks.^[20] Strong MMMU performance therefore should not be summarized as human-level general visual-spatial intelligence.

Reception and rollout changes

The launch received praise for benchmark performance, coding ability, price, and the attempt to simplify model selection. It also received criticism from users who preferred GPT-4o's style or wanted direct control over model choice. Outside commentary cautioned that the measured technical improvement was meaningful but not evidence of the broad expert-level capability suggested by some launch rhetoric.^[18][28]

The router outage during the first day made the product appear less capable for some queries. OpenAI responded by restoring GPT-4o for paid users and adding Auto, Fast, and Thinking choices. The company also raised the weekly Thinking allowance for Plus users and provided access to additional legacy models.^[9][29] These changes were product decisions and did not alter the fixed gpt-5-2025-08-07 API snapshot.

The launch's benchmark presentation also prompted scrutiny. TechCrunch reported that OpenAI acknowledged an incorrect chart used during the presentation.^[29] The corrected numerical tables, model documentation, and system card are more reliable sources for exact results than screenshots from the livestream.

OpenAI continued updating the GPT-5 product after launch. The October 3 sensitive-conversations snapshot changed safety behavior. GPT-5-Codex, released September 15, was optimized for agentic coding with reinforcement learning on real-world coding tasks and received its own safety addendum.^[13][14] These variants should be named explicitly because their behavior and evaluation conditions differed from the August general model.

Succession and availability

OpenAI released GPT-5.1 on November 12, 2025. GPT-5.1 Instant added adaptive reasoning for harder questions, while GPT-5.1 Thinking varied its thinking time more dynamically and changed communication style. GPT-5.1 Auto continued routing between models. OpenAI kept the original GPT-5 Instant and Thinking in ChatGPT's legacy-model menu for paid subscribers during a three-month transition.^[9][10]

On February 13, 2026, OpenAI retired the original GPT-5 Instant and Thinking models from ChatGPT. Enterprise workspaces retained access through February 19, and GPT-5 Pro remained available to paid users through that date. OpenAI said the change did not alter API availability at that time.^[11][12]

By July 28, 2026, the original gpt-5, gpt-5-mini, and gpt-5-nano pages still documented the launch models. The full-size model page called GPT-5 a previous model and recommended GPT-5.6. The mini and nano pages likewise directed new lower-cost work toward newer GPT-5.6 tiers.^[5][6][7] Existing applications could still need the old snapshots for compatibility, but that was different from being the current recommended model family.

GPT-5.6 became generally available on July 9, 2026 across ChatGPT, Codex, and the API.^[15] This later availability is relevant to the original model's lifecycle, but GPT-5.6's architecture, prices, evaluations, and safety work belong to its own article. The same scope rule applies to GPT-5.2 through GPT-5.5.

Timeline

Scroll sideways for more →

See also

• GPT-5 Pro
• Reasoning models
• OpenAI API
• Context window
• Benchmark (AI)
• AI safety

References

1. ^OpenAI. "Introducing GPT-5." August 7, 2025. openai.com/...introducing-gpt-5
2. ^OpenAI. "Introducing GPT-5 for developers." August 7, 2025. openai.com/...introducing-gpt-5-for-developers
3. ^OpenAI. "GPT-5 System Card." August 7, 2025. openai.com/...gpt-5-system-card
4. ^OpenAI. "GPT-5 System Card." PDF, August 13, 2025. cdn.openai.com/gpt-5-system-card.pdf
5. ^OpenAI. "GPT-5 Model." OpenAI API documentation. developers.openai.com/...gpt-5
6. ^OpenAI. "GPT-5 mini Model." OpenAI API documentation. developers.openai.com/...gpt-5-mini
7. ^OpenAI. "GPT-5 nano Model." OpenAI API documentation. developers.openai.com/...gpt-5-nano
8. ^OpenAI. "GPT-5 Chat Model." OpenAI API documentation. developers.openai.com/...gpt-5-chat-latest
9. ^OpenAI. "ChatGPT Release Notes." OpenAI Help Center. help.openai.com/...6825453-chatgpt-release-notes
10. ^OpenAI. "GPT-5.1: A smarter, more conversational ChatGPT." November 12, 2025. openai.com/...gpt-5-1
11. ^OpenAI. "Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT." January 29, 2026. openai.com/...retiring-gpt-4o-and-older-models
12. ^OpenAI. "Retiring GPT-4o and other ChatGPT models." OpenAI Help Center. help.openai.com/...gpt-4o-and-other-chatgpt-models
13. ^OpenAI. "Addendum to GPT-5 system card: GPT-5-Codex." September 15, 2025. openai.com/...gpt-5-system-card-addendum-gpt-5-codex
14. ^OpenAI. "Addendum to GPT-5 System Card: Sensitive conversations." October 27, 2025. openai.com/...-system-card-sensitive-conversations
15. ^OpenAI. "GPT-5.6: Frontier intelligence that scales with your ambition." July 9, 2026. openai.com/...gpt-5-6
16. ^Microsoft. "Microsoft incorporates OpenAI's GPT-5 into consumer, developer and enterprise offerings." August 7, 2025. news.microsoft.com/...openai-gpt-5
17. ^Aider. "Aider LLM Leaderboards." GPT-5 high run dated August 23, 2025. aider.chat/...leaderboards
18. ^Artificial Analysis. "GPT-5 Benchmarks and Analysis." August 7, 2025. artificialanalysis.ai/...5-benchmarks-and-analysis
19. ^METR. "Details about METR's evaluation of OpenAI GPT-5." August 2025. metr.org/...gpt-5-report
20. ^Cai, Zhongang, et al. "Holistic Evaluation of Multimodal LLMs on Spatial Intelligence." arXiv:2508.13142, 2025. arxiv.org/...2508.13142
21. ^Georgiou, Georgios P. "Capabilities of GPT-5 across critical domains: Is it the next breakthrough?" arXiv:2508.19259, 2025. arxiv.org/...2508.19259
22. ^OpenAI. "Introducing SWE-bench Verified." August 13, 2024. openai.com/...introducing-swe-bench-verified
23. ^OpenAI. "Why SWE-bench Verified no longer measures frontier coding capabilities." February 23, 2026. openai.com/...o-longer-evaluate-swe-bench-verified
24. ^Arora, Rahul K., et al. "HealthBench: Evaluating Large Language Models Towards Improved Human Health." arXiv:2505.08775, 2025. arxiv.org/...2505.08775
25. ^Barres, Victor, et al. "tau2-Bench: Evaluating Conversational Agents in a Dual-Control Environment." arXiv:2506.07982, 2025. arxiv.org/...2506.07982
26. ^Yue, Xiang, et al. "MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI." arXiv:2311.16502, 2023. arxiv.org/...2311.16502
27. ^Rein, David, et al. "GPQA: A Graduate-Level Google-Proof Q&A Benchmark." arXiv:2311.12022, 2023. arxiv.org/...2311.12022
28. ^O'Brien, Matt. "OpenAI launches GPT-5, a potential barometer for whether AI hype is justified." Associated Press, August 7, 2025. apnews.com/...d12cd2d6310a2515042067b5d3965aa1
29. ^Bort, Julie. "Sam Altman addresses 'bumpy' GPT-5 rollout, bringing 4o back, and the 'chart crime'." TechCrunch, August 8, 2025. techcrunch.com/...ging-4o-back-and-the-chart-crime