GPT-5

GPT-5 is the fifth generation of OpenAI's GPT family of language models, announced on 7 August 2025. OpenAI presented it not as a single network but as a system: a fast general model, a slower reasoning model, and a router trained to decide which one should answer a given request. The design was meant to end the choice between speed and deliberation that had defined chatbot interfaces since the first reasoning models appeared 1.

The launch was one of the most heavily promoted in the company's history. GPT-5 appeared the same day in ChatGPT, in OpenAI's API, in Microsoft 365 Copilot and GitHub Copilot, and in Azure AI Foundry. OpenAI reported improvements in software engineering, mathematics, health questions and factual accuracy, and said the family hallucinated less and flattered users less than its predecessors. Those figures came from the developer. Independent evaluation of a system that changes frequently is difficult, and at launch it was scarce 2.

This entry covers what GPT-5 is, how the system is assembled, its release and version history, its published benchmark results, the access and pricing model, and the criticism that followed — including an error in the launch charts, complaints about the router, and a continuing argument about how its safety behaviour should be described.

GPT-5 is a family rather than one model. The API exposes gpt-5, gpt-5-mini and gpt-5-nano, which differ in size, latency and price, and ChatGPT exposes tiers that map onto them. Above the standard models sits GPT-5 Pro, which spends substantially more computation and is sold at a higher price for difficult reasoning work 3.

The distinguishing component is the router. In earlier products, users picked a model from a menu; in GPT-5, a trained component inspects the request and sends it either to the general model or to the reasoning model, which writes an internal chain of thought before answering. OpenAI says the router is trained on observed signals, including when users switch models, preference rates and measured correctness, and that it improves as more traffic is observed. The user sees one answer and, in the consumer product, no reliable indication of which component produced it 1 4.

Reasoning depth is also exposed as a parameter. Developers can set reasoning effort to minimal, low, medium or high, and a verbosity setting controls answer length; ChatGPT maps its thinking modes onto the same machinery. That makes GPT-5 less a fixed product than a set of configurations: two results labelled GPT-5 may come from different components at different effort levels 3.

The naming reflects a product decision as much as a research one: OpenAI judged that most people do not want to choose a model. The cost is transparency, since neither users nor outside evaluators can tell from the interface which path answered 5.

GPT-5 became available on 7 August 2025 to ChatGPT's free, Plus, Pro, Team, Enterprise and Edu users, with different limits on access to the reasoning and Pro variants, and simultaneously through the API and Microsoft's platforms. OpenAI described the free tier as receiving the standard model with limited reasoning requests, and the Pro tier as receiving extended access to the most expensive configuration 1.

The first weeks produced an unusual volume of complaint. The consumer product removed the model picker in favour of an automatic mode, and some users found that requests they had previously sent to a reasoning model were answered by the faster one. OpenAI restored access to an earlier model for paying subscribers, raised limits, and its chief executive acknowledged in a public discussion that the company had underestimated how attached some users were to the previous model's tone 5.

A point release, GPT-5.1, followed in November 2025. OpenAI described it as better at following instructions, warmer in tone and able to decide for itself how long to think, with separate Instant and Thinking variants and a coding-focused Codex version. The pattern — a large flagship launch, rapid point updates, and gradual retirement of older snapshots — has been consistent across OpenAI's model generations 4.

Versions are not permanent. OpenAI publishes deprecation dates for API snapshots and has retired models at short notice, which matters to organisations built on a specific version; the documentation recommends pinning to a dated snapshot rather than a moving alias 4.

OpenAI has not published GPT-5's parameter count, layer structure, tokeniser or training-data composition, and the system card does not supply them. What the company has described is the shape of the training process: pre-training on a mixture of web text, code and licensed material, followed by post-training that includes supervised fine-tuning, reinforcement learning from human feedback, and reinforcement learning on tasks with checkable answers 2.

The reasoning component is trained with reinforcement learning on problems where correctness can be verified, such as mathematics and code. That is the mechanism behind the long internal deliberation and the reasoning-effort control: the model has learned to use more computation when more is allowed. OpenAI has said the traces are not a full account of its computation and does not release them through the API 3.

Post-training also shapes tone. OpenAI states that it rewarded honest assessments and penalised flattery, in response to complaints that earlier models agreed with users too readily. The effect is visible in use but difficult to verify, because the preference data and reward models are not public 2.

Safety training follows an approach OpenAI calls safe-completions. Rather than refusing a sensitive request outright, the model is trained to reason about the policy and answer at a level of detail the company considers safe. OpenAI argues this is more useful without being more harmful; critics argue it turns a visible refusal into an invisible judgement that is harder to audit 2.

The strongest documented gains are in software engineering. GPT-5 handles multi-file edits, reads repository context, and produces runnable front-end code; OpenAI's launch demonstrations centred on fixing bugs and building small applications from a description. In agentic settings, where the model calls tools and iterates against test results, it is markedly more persistent than the GPT-4 generation, and that persistence is the main reason coding agents built on it can complete tasks that previously required a person 1.

On structured reasoning the model performs well on competition mathematics, scientific multiple-choice questions and data analysis, particularly at higher effort settings. It supports tool calls, structured outputs and long documents, and it accepts text and images as input. In health-related questions OpenAI reported a substantial improvement over its previous models, while cautioning that the system is not a medical device and should not be used for diagnosis 2.

The limits are equally documented. GPT-5 still fabricates citations, dates and quotations, and it can state a falsehood in the same register it uses for facts. Arithmetic and precise counting remain unreliable without a tool. Its knowledge stops at a stated cutoff, so anything later requires search or supplied documents. The advertised context window is large, but effective use of very long inputs degrades, and reasoning answers cost more and arrive more slowly 4.

Image generation and voice are handled by separate models, so the system is best understood as a coordinator rather than one network that does everything 1.

OpenAI reported a set of results at launch, all produced by the developer. On SWE-bench Verified, which asks a model to resolve real issues from open-source repositories, it reported 74.9 per cent at high reasoning effort, and on Aider's polyglot coding benchmark 88 per cent. It reported 88.4 per cent on GPQA Diamond, 94.6 per cent on AIME 2025 without tools, 84.2 per cent on MMMU, and 46.2 per cent on HealthBench Hard, a medical evaluation written with physicians. It also claimed roughly 45 per cent fewer factual errors than GPT-4o in standard mode and 80 per cent fewer in thinking mode, measured by its own graders 1.

These numbers should be read as vendor-reported. Harnesses differ in prompting, tool access, sampling and grading, and scores are comparable only under the same conditions. Several benchmarks are close to saturation, so the gap between leading models is smaller than the noise those choices introduce. SWE-bench Verified was itself constructed by OpenAI to remove ambiguous tasks, so the company controls part of the evaluation's design 6 7.

Contamination is the deeper problem. Public benchmark questions can appear in training data, and a model that has seen a problem may score well without generalising. The 2025 AI Index documents both saturation and the difficulty of proving a model has not seen a test; independent bodies have responded with private held-out evaluations, which are expensive and quickly stale 7.

The launch also produced a straightforward error: a bar chart in OpenAI's presentation mislabelled results, making the new model appear worse on a coding benchmark than it was. The company corrected the figure and attributed it to a design mistake, not to the numbers 5.

In ChatGPT, GPT-5 is available on the free tier with limits, and without limit on the paid tiers, with the Pro subscription costing substantially more and including access to GPT-5 Pro. Prices and tier names have changed repeatedly, and regional pricing varies with tax and payment rules, so the published plan page is the only reliable source at a given moment 8.

In the API, OpenAI prices the family per million tokens. At launch, gpt-5 cost 1.25 US dollars per million input tokens, 0.125 dollars per million cached input tokens and 10 dollars per million output tokens; gpt-5-mini and gpt-5-nano were priced at roughly a tenth and a fortieth of that. Reasoning tokens are billed as output, which means a hard problem can cost far more than its visible answer suggests. Batch processing is discounted, and cached context is cheaper on repeat calls 8.

Access is also mediated by platform. Microsoft offers the models through Azure AI Foundry and embeds them in Copilot products under its own terms; GitHub uses them in Copilot; other cloud providers host them through partnerships. Enterprise customers typically buy through those channels rather than directly, so the terms governing their data can differ from OpenAI's consumer terms 3.

Limits are part of the product: rate limits, caps on reasoning requests and regional restrictions apply, and OpenAI has withdrawn or delayed features in jurisdictions with specific requirements. For developers the binding constraint is often not price but the rate at which tokens can be consumed 4.

Among developers, the initial reception was strongly positive on coding. Agentic tools built on the model completed tasks that had previously needed close supervision, and third-party leaderboards generally placed it near the top of the field, though rarely by the margins OpenAI reported. Its willingness to keep working through failures was cited more often than raw scores 5.

Consumer reaction was more mixed. A vocal group of users found the model colder, more literal and more prone to refusing or hedging than the model it replaced, and objected to the removal of the picker that had let them choose. OpenAI adjusted tone and restored some access; the episode became a case study in how much users attach to a system's personality. The company's framing — that the model was better at saying it did not know — did not resolve the disagreement 5.

In enterprises, adoption was rapid because the model arrived inside products companies already used: Microsoft's Copilot, GitHub's coding assistant and Azure's catalogue all carried it within days, a distribution advantage no competitor matches. The counterweight is dependence, since an organisation that standardises on one vendor inherits its pricing, deprecation schedule and policy decisions 3.

Among safety researchers, the reception divided. Those who focus on measurable misuse risk welcomed the reported reductions in harmful output and the detailed system card. Those who focus on loss of control argued that a model which reasons about its own policy in hidden deliberation, deployed to hundreds of millions of people, needs external evaluation rather than self-assessment 2.

The benchmark record is contested in three ways. First, the headline numbers are vendor-reported, and no independent body has reproduced the full set under identical conditions. Second, the benchmarks themselves are weakening: SWE-bench Verified is approaching saturation, AIME problems circulate publicly, and GPQA's questions have been available long enough to raise contamination questions. Third, the router means that a reported score describes one configuration, not the product most people use 6 7.

Safety claims are disputed more sharply. OpenAI describes GPT-5 as its most robustly trained model and points to red-teaming, preparedness evaluations and safe-completions. External researchers and some former employees argue that self-reported evaluations are not evidence of safety at deployment scale, that hidden deliberation makes behaviour harder to interpret, and that the refusal boundary is a policy choice presented as a technical one. Neither position has been settled by experiment 2.

The capability narrative is contested too. OpenAI's language around the launch blurred the line between a strong product and a step towards general intelligence, while the measured record shows a model clearly better at coding and structured reasoning and still unreliable on facts and long-horizon planning. Claims of a qualitative difference are claims about degree, not kind 1.

What would settle these questions is unglamorous: independent, held-out evaluation of the deployed system, published methodology, and access for outside researchers. None of the major developers currently offers all three, and until one does, the honest description of GPT-5's benchmark record is that it is detailed, internally consistent and largely unverified from outside.