OpenAI brings GPT-5 to free ChatGPT with an automatic selector

On August 7, 2025, OpenAI introduced GPT-5 and a router that chooses between fast answers and additional reasoning. The original source supports the documentary core of the event; the announcement's tables are vendor evaluations and do not replace a test on the reader's task.

GPT-5 arrives nearly three years after ChatGPT popularized generative AI, and with an unusual distribution scale: OpenAI says the service now has more than 700 million weekly users. That makes the launch significant beyond benchmark comparisons: any improvement or mistake in the model will quickly flow into millions of everyday conversations, tasks and decisions. Document supporting the figure.

A model that decides how much to think

The company describes GPT-5 as a unified system. It combines the speed of the GPT family with the reasoning capabilities introduced in the o-series models, which break down complex problems before answering.

An internal router analyzes each request and decides whether to provide an immediate response or activate a more intensive reasoning mode. The goal is to reduce a friction that had become commonplace in ChatGPT: manually choosing which model to use and figuring out when it is worth spending more computing time.

In practice, a simple question should get a quick answer, while a coding, analysis or planning request may be routed through a longer process. Paid subscribers retain more control and higher usage limits. The Plus plan, at $20 a month, expands access; the Pro plan, at $200 a month, includes unlimited use of GPT-5 and GPT-5 Pro, a version that uses more resources to tackle difficult problems. Document supporting the figure.

OpenAI is also adding four configurable personalities to ChatGPT — Cynic, Robot, Listener and Nerd. They do not change the system’s knowledge, only the style in which it responds. It is a meaningful product decision: the company wants the assistant to adapt to user preferences without turning every conversation into a collection of instructions about tone.

Coding, science and a promise of fewer hallucinations

OpenAI says GPT-5 is particularly competitive at coding. On SWE-bench Verified, a test based on real issues from GitHub repositories, it scored 74.9% on its first attempt. That is slightly above Claude Opus 4.1’s 74.5% and clearly ahead of Gemini 2.5 Pro’s 59.6%, according to results compared by the company. Document supporting the figure.

That is a significant result for people who use AI to build complete applications from a description, a practice known as vibe coding. But it does not mean the model can replace a programmer’s review: SWE-bench measures bounded problems, and a rate near 75% still leaves one in four unresolved on the first attempt. Document supporting the figure.

GPT-5 Pro scored 89.4% on GPQA Diamond, a PhD-level science question exam, and 42% on Humanity’s Last Exam when using tools. On the latter test, it fell short of the 44.4% reported for Grok 4 Heavy. The new model therefore does not top every frontier leaderboard; OpenAI is presenting an advantage in specific areas, not an insurmountable lead over Anthropic, Google or xAI. Document supporting the figure.

The most practical improvement may be reliability. OpenAI says GPT-5 with reasoning produced incorrect or fabricated answers in 4.8% of cases in its evaluation of responses to ChatGPT prompts, compared with 22% for o3 and 20.6% for GPT-4o. Hallucinations are plausible but false answers, one of the problems that most limits the professional use of these systems. Document supporting the figure.

In healthcare, the company puts GPT-5’s hallucination rate with reasoning at 1.6% on HealthBench Hard Hallucinations, compared with 12.9% for GPT-4o and 15.8% for o3. That would be a meaningful improvement if confirmed in real-world use, but it does not turn the chatbot into a healthcare professional. Interpreting symptoms, clinical results or treatments still requires medical judgment and data that a conversational model cannot verify on its own. Document supporting the figure.

The developer offering is changing too

GPT-5 is available through OpenAI’s API in three sizes: gpt-5, gpt-5-mini and gpt-5-nano. The variants let developers balance cost, speed and reasoning capacity depending on the task. The company has also added a verbosity control so developers can specify whether they want more concise or more extensive responses.

The flagship model costs $1.25 per million input tokens and $10 per million output tokens. A token is a unit of text that models use to process and generate language; it does not exactly correspond to a word. The difference between input and output matters for products that generate long reports, code or detailed answers: producing text remains considerably more expensive than reading it. Document supporting the figure.

The release comes just days after OpenAI published gpt-oss, a family of open-weight reasoning models that companies and developers can download and run themselves. With GPT-5, the company is maintaining a complementary strategy: offering open models for certain uses while reserving its most advanced system for ChatGPT and its commercial infrastructure.

Team, Edu and Enterprise accounts will receive GPT-5 as their default model next week. The next test will not be another benchmark chart, but whether the automatic selector consistently makes the right call, whether the model reduces errors outside internal tests and whether its improvements justify the cost for those building products on the API.

Turning the headline into a check

The first step is to freeze the system's identity. OpenAI introduced GPT-5 and a router that chooses between fast answers and additional reasoning. A commercial name may cover different revisions, automatic routes and tools. A test record should preserve date, access mode, configuration, permissions and the full output. Without that snapshot, an improvement or failure observed today cannot rigorously be attributed to the version another person will use tomorrow.

Next, turn how to evaluate the whole system, including the router, rather than only the model name into cases with acceptance criteria. Build a local sample containing easy, ambiguous, long and deliberately impossible tasks. Record the input, the information the model may consult and what outcome would count as sufficient. A vendor-selected demonstration shows possibility; a test set preserved by the user measures reliability.

Autonomy needs a permission ladder. Reading and proposing are not the same as editing, sending or buying. A safer setup begins with read-only access, requires a preview and reserves execution for explicit approval. It also keeps a log and a rollback path. Judge the model by the errors the surrounding system contains, not by the confidence of its plan.

What the record must preserve

Cost and quality must be measured together. A cheap answer that must be reviewed from scratch can cost more than a slower but verifiable one. Measurement includes waiting, retries, consumption, human oversight and the consequences of failure. the announcement's tables are vendor evaluations and do not replace a test on the reader's task. That boundary turns the announcement into a testable hypothesis rather than a promise to be believed.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.