Editor’s update, October 19, 2026: This article covers GPT-5’s August 7, 2025 launch. Its original ChatGPT model lineup and access limits are historical. OpenAI has since retired GPT-5 Instant and Thinking from ChatGPT, while the original GPT-5 API snapshots have a separate shutdown scheduled for December 11, 2026. The gpt-5-chat-latest API variant already had a July 23, 2026 shutdown. See the current API lifecycle schedule and ChatGPT retirement notice for those distinct changes.
OpenAI announced GPT-5 on August 7, 2025 as a major update to ChatGPT and its developer platform. The launch emphasized a system that could answer quickly or use deeper reasoning, alongside improvements in coding, writing, visual understanding, and health-related conversations.
The release also aimed to simplify the ChatGPT model lineup. OpenAI’s launch announcement described GPT-5 becoming the default for signed-in users in place of GPT-4o, o3, o4-mini, GPT-4.1, and GPT-4.5. That was a ChatGPT product transition, not a declaration that all those API models stopped working on launch day.
How built-in thinking worked at launch
The original GPT-5 ChatGPT experience combined three components:
- A fast model for questions that did not need extended reasoning.
- GPT-5 Thinking for harder problems.
- A router that selected between them using factors such as task complexity, tool needs, conversation type, and the user’s stated intent.
This was a system of models rather than a claim that every request used a single identical model or reasoning budget. OpenAI said the router was trained using signals such as user model switches, response preferences, and measured correctness. The intended benefit was to reduce the need for users to choose a model before every task.
The GPT-5 system card distinguishes the fast, thinking, mini, and extended-compute variants. That distinction also helps explain why GPT-5 results can vary by product, configuration, and evaluation.
The API release was a different interface
OpenAI separately released gpt-5, gpt-5-mini, and gpt-5-nano for developers. The API’s GPT-5 was the reasoning model, whereas ChatGPT combined reasoning and non-reasoning models with a router. Even the API’s minimal reasoning setting was distinct from ChatGPT’s non-reasoning model.
The API launch added controls for reasoning effort and response verbosity, along with custom tools. These were developer controls, not a promise that every ChatGPT user would see the same settings. The developer announcement is the appropriate launch reference for API behavior.
What the launch benchmarks actually showed
The table below preserves important launch results with their model and tool conditions. All figures are OpenAI-reported historical evaluations, not fresh testing by this website or a current ranking of AI systems.
| Evaluation | Reported result | Model and conditions |
|---|---|---|
| AIME 2025 | 94.6% | GPT-5, high reasoning effort, without tools |
| GPQA Diamond | 85.7% | GPT-5, high reasoning effort, without tools |
| GPQA | 88.4% | GPT-5 pro, without tools, as reported in the ChatGPT launch |
| SWE-bench Verified | 74.9% | GPT-5; OpenAI’s 477-task evaluation subset |
| HMMT 2025 | 93.3% | GPT-5, high reasoning effort, without tools |
| MMMU | 84.2% | GPT-5, high reasoning effort |
| HealthBench Hard | 46.2% | GPT-5 result in OpenAI’s ChatGPT launch evaluation |
Correction: the original article assigned the 94.6% AIME result to GPT-5 Pro with Python. OpenAI’s published developer table assigns 94.6% to GPT-5 at high reasoning effort without tools. Tool-enabled and no-tool results should not be mixed.
SWE-bench Verified involved producing patches for repository issues. OpenAI excluded 23 of 500 tasks that did not reliably run on its infrastructure. The developer announcement also notes a verification-focused prompt for GPT-5 that did not improve o3. These conditions matter when comparing scores across articles or model launches.
Coding and visual understanding
The original announcement emphasized frontend generation, debugging, and working with larger codebases. These were reported improvements in capability, not evidence that every generated application would be production-ready without review.
Other retained visual results include 78.4% on MMMU-Pro, averaged across its standard and vision sets; 81.1% on CharXiv reasoning with Python enabled; 84.6% on VideoMMMU with a maximum of 256 frames; and 65.7% on ERQA. OpenAI’s developer table reports these for GPT-5 at high reasoning effort. They measure different kinds of visual tasks and are not interchangeable accuracy figures.
Writing and health-related conversations
OpenAI presented GPT-5 as a stronger writing collaborator, showing examples of creative work alongside practical uses such as editing reports, emails, and memos. Curated writing examples illustrate a style difference; they do not establish that one model will be preferred by every reader.
For health, the launch emphasized asking useful follow-up questions, flagging concerns, and helping users understand information. The important limitation remains that ChatGPT does not replace a medical professional. A HealthBench score is an evaluation result, not medical certification or a guarantee about an individual answer.
Reliability improvements and their limits
OpenAI reported that, with web search enabled on prompts representative of ChatGPT traffic, GPT-5 responses were about 45% less likely to contain a factual error than GPT-4o responses. GPT-5 with thinking was about 80% less likely than o3 under that evaluation. Separately, open-ended factuality tests showed roughly one-sixth the hallucination rate of o3 when GPT-5 used thinking. These are different comparisons and should not be treated as a universal error rate.
The launch also reported better acknowledgment of missing information and impossible tasks. For example, on a test with images removed, o3 gave confident answers about nonexistent images 86.7% of the time, compared with 9% for GPT-5. This demonstrated an improvement on that test while still leaving failures.
For safety training, OpenAI introduced safe completions: the model could provide a limited or higher-level answer when appropriate, or explain a refusal and suggest safer alternatives. The company also reported reducing overly agreeable responses in targeted tests from 14.5% to below 6%. Neither result means that errors, unsafe outputs, or excessive agreement were eliminated.
The launch included opt-in personality previews named Cynic, Robot, Listener, and Nerd. These were communication-style settings in the 2025 rollout, not independent measures of factual accuracy or a current promise about available presets.
GPT-5 pro was an extended-reasoning variant
GPT-5 pro used additional parallel test-time computation for difficult tasks and replaced o3-pro in the launch ChatGPT experience. It was a model option associated with access through the Pro plan, rather than a newly announced standalone subscription tier.
In OpenAI’s evaluation of more than 1,000 real-world reasoning prompts, external experts preferred GPT-5 pro over GPT-5 Thinking 67.8% of the time and observed 22% fewer major errors. Those results describe that evaluation, not a guarantee of certainty or a reason to skip professional review in consequential work.
Original rollout and current lifecycle
On August 7, 2025, OpenAI began rolling GPT-5 out to Free, Plus, Pro, and Team users, with Enterprise and Edu announced for the following week. The launch offered different usage levels and mini fallbacks. Those plan descriptions and limits belong to the launch period and should not be used to infer present-day entitlements.
- ChatGPT: OpenAI retired GPT-5 Instant and Thinking on February 13, 2026, with limited Enterprise access and GPT-5 Pro extensions through February 19.
- Chat API variant:
gpt-5-chat-latesthad an API shutdown on July 23, 2026. - Core API launch snapshots:
gpt-5-2025-08-07,gpt-5-mini-2025-08-07, andgpt-5-nano-2025-08-07are deprecated, with shutdown scheduled for December 11, 2026. That date is still in the future as of this update.
Check the exact model ID in the API model reference and the deprecation schedule before planning a migration. Current model selection should use the current catalog, rather than assuming that the original GPT-5 launch remains the default product experience.
GPT-5’s historical contribution was the combination of automatic reasoning choices, stronger technical capabilities, and more explicit reliability work. Its launch claims remain useful when read with their dates, model variants, and evaluation conditions.
