GPT-5 arrives

The announcement and the narrow claim

On 7 August 2025, OpenAI introduced GPT-5 and made it the default model in ChatGPT for every tier, including free accounts, on a product that the company said reached roughly 700 million weekly active users. The official announcement framed the release less as a single larger model than as a reorganization of the ChatGPT experience: a system that could answer quickly for routine requests and spend more time reasoning for hard ones. Sam Altman told reporters that using it was like having, in his words, a "team of Ph.D. level experts in your pocket," a framing NBC News recorded on launch day.

The narrow, documented claim is modest and checkable: on 7 August 2025 OpenAI shipped a new flagship model family and a unified routing layer, replaced the previous defaults, and published benchmark figures and a system card. The broad historical judgment, that GPT-5 marks a decisive step toward general intelligence, is a separate order of claim, and the useful discipline is to keep the two apart. A launch is an event. A capability is a measured quantity. A turning point is a retrospective verdict. This report treats them as three different things.

What OpenAI actually shipped

GPT-5 was not one artifact but a small family plus a control system. Inside ChatGPT, OpenAI described a fast path (marketed as GPT-5 Instant), a deeper reasoning path (GPT-5 Thinking), and a "real-time router" that decides, per request, which path should answer. Above them sits GPT-5 Pro, an extended-reasoning configuration reserved for the 200 dollar per month Pro subscription. In the API, TechCrunch reported that OpenAI exposed three sizes, gpt-5, gpt-5-mini and gpt-5-nano, differing in how much computation is allocated to reasoning, with the flagship priced at 1.25 dollars per million input tokens and 10 dollars per million output tokens, and a context window on the order of 400,000 tokens.

The API also added two controls that make the product philosophy concrete: a reasoning_effort setting (from minimal through high) that governs how much internal deliberation the model performs before answering, and a verbosity setting that governs how long the answer runs. GPT-5 replaced OpenAI o3 and GPT-4o as the ChatGPT defaults, and consolidated a lineup that had grown to include o4-mini, GPT-4.1 and GPT-4.5. The important structural point, noted in this event's grounding, is that product behavior and a single model identifier were no longer the same thing: the string "GPT-5" in ChatGPT names a routed system, while "gpt-5" in the API names a specific model called with its own effort and verbosity settings.

The GPT-5 system card discloses the naming underneath that marketing layer. The fast, high-throughput model is documented as gpt-5-main, with a lighter gpt-5-main-mini variant, while the deeper reasoning path is gpt-5-thinking, alongside its own gpt-5-thinking-mini. That internal nomenclature is worth noting because it shows that the "GPT-5 Instant" and "GPT-5 Thinking" labels ChatGPT users see are a marketing gloss over an engineering distinction that predates the launch: main models optimized for latency and cost, thinking models optimized for depth, joined by a router that decides which one answers a given message. The naming also clarifies a detail easy to miss, that "GPT-5 Thinking" is itself a family with a smaller sibling, so the reasoning tier has its own internal cost and speed tradeoff before a request ever reaches GPT-5 Pro.

The lineage that made a merge possible

GPT-5's design only makes sense against the two research lines it fused. The first is the pretrained autoregressive transformer scaled up: Brown and colleagues' GPT-3 paper showed in 2020 that a 175-billion-parameter language model could perform many tasks from a prompt alone, and the GPT-4 technical report in 2023 pushed that capability and added multimodal input while withholding architectural detail. That line optimized what a model knows after pretraining.

The second line optimized what a model does at inference. In September 2024 OpenAI introduced o1 with the post "Learning to reason with LLMs", a model trained through reinforcement learning to produce a long internal chain of thought before answering, extending an idea popularized by Wei and colleagues' chain-of-thought work. Its successor, o3, pushed reasoning further on coding, mathematics and science. These reasoning models established a second scaling axis: spending more computation at test time, generating more thinking tokens, could substitute for some of the gains previously bought only by a larger pretraining run. GPT-5's premise was that a user should not have to choose between the two lines. The router chooses for them.

The constraint it was built to remove

The problem GPT-5 addressed was not primarily a modelling deficiency; it was a product one. By mid-2025, ChatGPT presented a menu of models (GPT-4o, o3, o4-mini, GPT-4.1) whose names encoded engineering history rather than user intent. Most people could not tell which model suited which task, and even sophisticated users maintained ad hoc rules of thumb. A general audience asked to reason like a systems engineer just to send a message will, on the whole, simply pick whatever is selected by default and never revisit the choice, which defeats the purpose of offering options at all. Enterprise customers evaluating the API faced a parallel version of the same problem, choosing among o-series and GPT-series options that varied in latency, cost and reasoning depth without one obvious default for an ordinary request.

The router is OpenAI's structural answer: make model selection a decision the system makes from the request, not a decision the user makes from a dropdown. Fortune argued that this routing pattern, whatever its rough launch, may be where consumer AI is heading, because it hides an increasingly complex model portfolio behind one entry point.

That reframing carries a cost worth naming. A dropdown is legible: the user knows exactly what they invoked. A router is convenient but opaque: two identical prompts can be handled by different underlying models, and the user cannot always tell which. The design trades user control for user simplicity, and the launch would immediately test whether that trade was acceptable.

How the routing works, and why a model ID is not the product

Mechanically, the router is a lightweight decision layer that inspects a request and estimates whether a fast response or extended reasoning will serve it better, then dispatches accordingly. OpenAI described it as learning from signals including which responses users preferred and when they explicitly asked the model to think harder. When it works, a factual lookup returns quickly and a multi-step coding or math task is escalated to the reasoning path without the user doing anything.

This is why the API surface and the chat surface diverge. In the API a developer can pin gpt-5 and set reasoning_effort to high for a hard agentic task, or to minimal for cheap classification, making behavior explicit and reproducible. In ChatGPT the same underlying capabilities are mediated by the router, so behavior depends on a classification the user does not see. The practical consequence is that a benchmark run against the API with a fixed effort level and a casual chat in the app are not the same system, and comparisons between them can mislead. The grounding for this event states the point plainly: the product's behavior and a single model identifier are not identical.

The evidence OpenAI presented

OpenAI supported the launch with a battery of task-specific results, and these should be read as the provider's own reported figures pending independent replication. On SWE-bench Verified, a benchmark of real GitHub issues that a model must resolve by editing a codebase, OpenAI reported that GPT-5 scored 74.9 percent, with strong results on the Aider Polyglot coding test. On competition mathematics it reported 94.6 percent on AIME 2025 without tools, and on the graduate-level GPQA benchmark of expert science questions it reported results near the high-eighties percent range. It also cited multimodal understanding around 84 percent on MMMU and new numbers on HealthBench for medical queries.

Two cautions apply. First, as the template for this timeline insists, these are scores on defined tasks under stated conditions, not percentages of human intelligence, and a benchmark can be strong while real-world reliability varies. Second, several of these benchmarks were designed by the research community precisely because earlier ones saturated: GPQA was built so that PhD-level experts reach only about 65 percent and skilled non-experts about 34 percent, which is what makes a high machine score notable and also what makes it worth verifying. The honest reading is that GPT-5 posted state-of-the-art or near-state-of-the-art numbers on hard public benchmarks according to OpenAI, and that task-specific evidence, not the launch itself, is what settles whether a claim holds.

A launch that undercut its own story

The most instructive part of this event is how the rollout diverged from the pitch. Within a day, users reported that GPT-5 felt worse than the models it replaced. The cause was partly infrastructural. Altman conceded on Reddit that "yesterday, we had a sev and the autoswitcher was out of commission for a chunk of the day, and the result was GPT-5 seemed way dumber." In other words, the router, the feature meant to make the product smarter, failed, and requests that should have been escalated to the reasoning path were not, so the system underperformed exactly where it was supposed to shine.

The presentation added a self-inflicted wound. A benchmark chart shown during the livestream displayed bars whose heights did not match their values, an error Altman later called a "mega chart screwup" and the internet dubbed a "chart crime," even though the published blog post carried correct figures. Then came the product backlash: removing GPT-4o as an option angered users who valued its conversational tone, some of whom described the new default as colder and more clipped. OpenAI responded by restoring 4o for paying subscribers and promising to roughly double rate limits for Plus users while the rollout finished. Altman summarized the episode bluntly, admitting the company "totally screwed up some things on the rollout," a candor Fortune documented. The lesson for a history of AI is that a capability claim and a deployment are separable, and that a strong model can ship inside a weak launch.

Safety reframed: safe completions and the hallucination question

Alongside the model, OpenAI published a GPT-5 system card describing its safety approach and evaluations. The headline conceptual shift is what OpenAI calls "safe completions": instead of a binary choice between fully answering and outright refusing, the model is trained to give the most helpful response it safely can, which matters most for dual-use prompts whose intent is ambiguous. On factual reliability, OpenAI reported that on challenging conversations, hallucinations from the reasoning configuration were substantially lower than from o3, citing large multiples of improvement on its own internal measures, and that the thinking model made several times fewer factual errors across browsing settings.

These are meaningful directional claims, and they are also exactly the claims that require outside scrutiny. As this event's grounding states, the launch does not establish that every earlier weakness, including hallucination, has disappeared; a lower rate on a benchmark is not zero, and reduction in one setting does not guarantee reduction in another. Reduced sycophancy and better instruction-following, both reported by OpenAI, are welcome if they hold up under independent evaluation. The correct posture is neither dismissal nor acceptance on faith: treat the system card as the provider's evidence and weigh it against task-specific tests conducted by others.

The field was converging, not diverging

GPT-5 did not appear in isolation, and reading it against its competitors clarifies what was and was not novel. By 2025 the leading labs were converging on the same structural idea: a single product that blends fast answers with optional deep reasoning. Anthropic and Google had each moved toward models that expose a reasoning or thinking mode: Google's Gemini 2.5 family, introduced in early 2025, built an adjustable thinking budget directly into its flagship model rather than shipping reasoning as a separate product line, and Anthropic's Claude 3.7 Sonnet, released that February, was marketed as a hybrid reasoning model able to show extended, step-by-step thinking on demand from a single checkpoint rather than requiring a separate reasoning-only model. Both moves anticipated the same insight GPT-5 acted on: that forcing a user to pick between a fast model and a smart one is a worse product than making that choice automatic or adjustable. So by mid-2025 the interesting question was never whether reasoning-plus-generation would combine, but who would package it most cleanly and route between modes most reliably. GPT-5's distinctive bet was to hide the choice behind an automatic router for consumers while exposing explicit effort and verbosity controls to developers, a split its two closest rivals had not fully committed to at that point.

Seen this way, GPT-5 is better understood as consolidation than as invention. The transformer architecture underneath is still the one described in the 2017 paper that this timeline treats as the field's architectural turning point. The reasoning behavior descends from the o-series. The scaling intuition that data and compute must be balanced, not merely maximized, traces to work like the Chinchilla study. GPT-5's contribution is a product-level synthesis of these threads plus the routing layer, which is a genuine piece of engineering but not a new scientific foundation. That is not a criticism; consolidation is how research capability becomes usable at scale.

How to place GPT-5 in AI history

The defensible placement is specific. GPT-5 is a significant product event and an incremental capability event, not an architectural revolution. Its clearest importance is that it delivered advanced reasoning to hundreds of millions of non-technical users by default, folding a fragmented model lineup into one entry point, and that it did so while OpenAI publicly stumbled on the router, the charts and the deprecation of a beloved predecessor. The CNBC coverage of the launch and OpenAI's own materials agree on the facts of scale and availability; they diverge, appropriately, on interpretation, which is left to independent evaluation over time.

Three questions help locate the event without overclaiming. Which earlier idea did it inherit? The pretrained transformer and the reinforcement-learned reasoning model, neither introduced here. Which constraint did it actually remove? A product constraint, the burden on users to pick a model, addressed by routing. Which limitation still required future work? Reliability under distribution shift, hallucination that is reduced but not eliminated, and the opacity that routing introduces. A timeline that answers those three questions carries more information than one that records a version number and a superlative. GPT-5 matters because it made frontier reasoning ambient and ordinary for a mass audience, and because its rollout is a case study in the gap between what a model can do and what a launch demonstrates. The router that made GPT-5 legible as a single product is the same router whose outage made the launch look like a regression, and that pairing, capability bundled with an operational single point of failure, is arguably the more durable lesson than any individual benchmark score.

References

• OpenAI (2025). Introducing GPT-5.
• OpenAI (2025). GPT-5 System Card.
• OpenAI (2024). Learning to reason with LLMs.
• Zeff, M. (2025). OpenAI's GPT-5 is here.
• Bort, J. (2025). Sam Altman addresses 'bumpy' GPT-5 rollout, bringing 4o back, and the 'chart crime'.
• Fortune (2025). GPT-5's model router ignited a user backlash against OpenAI, but it might be the future of AI.
• Fortune (2025). Sam Altman admits OpenAI totally screwed up its GPT-5 launch and says the company will spend trillions of dollars on data centers.
• NBC News (2025). OpenAI releases GPT-5, calling it a team of Ph.D. level experts in your pocket.
• CNBC (2025). OpenAI launches GPT-5 model for all ChatGPT users.
• Vaswani et al. (2017). Attention Is All You Need.
• Brown et al. (2020). Language Models are Few-Shot Learners.
• OpenAI (2023). GPT-4 Technical Report.
• Wei et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.
• Jimenez et al. (2023). SWE-bench: Can Language Models Resolve Real-World GitHub Issues?.
• Rein et al. (2023). GPQA: A Graduate-Level Google-Proof Q&A Benchmark.
• Hoffmann et al. (2022). Training Compute-Optimal Large Language Models.