OpenAI’s GPT-5 is here

OpenAI has launched GPT-5, a brand new flagship AI mannequin that may energy the corporate’s subsequent technology of ChatGPT.

GPT-5, which was launched Thursday, is OpenAI’s first “unified” AI mannequin and combines the reasoning skills of its o-series of fashions with the quick responses of its GPT sequence. The following-generation mannequin alerts a brand new period for ChatGPT — and its creator, OpenAI — pointing to OpenAI’s broader ambitions to develop AI methods which can be extra like brokers than chatbots.

Whereas GPT-4 enabled AI chatbots to supply sensible responses on all kinds of questions, GPT-5 permits ChatGPT to finish all kinds of duties on behalf of customers — equivalent to producing software program purposes, navigating a consumer’s calendar, or creating analysis briefs.

With GPT-5, OpenAI has additionally sought to make ChatGPT easier to make use of. As a substitute of asking customers to decide on the proper settings, GPT-5 comes outfitted with a real-time router that decides how one can provide the very best reply, whether or not that’s responding to consumer questions rapidly or taking extra time to “suppose” by solutions.

Throughout a briefing with reporters, OpenAI CEO Sam Altman claimed GPT-5 is “the very best mannequin on the planet,” and mentioned it represented a “important step” alongside the corporate’s path to growing AI that may outperform people at most economically priceless work — that’s, synthetic common intelligence (AGI).

“Having one thing like GPT-5 can be just about unimaginable at any earlier time in historical past,” mentioned Altman.

Beginning Thursday, GPT-5 shall be out there to all free customers of ChatGPT as their default mannequin. OpenAI’s VP of ChatGPT, Nick Turley, mentioned that is a part of the corporate’s effort to provide free customers entry to an AI reasoning mannequin for the primary time. (Beforehand, the corporate gated these extra superior fashions behind a paywall.)

“That is simply one of many ways in which I’m excited to reside the mission, ensuring that these things truly advantages folks,” mentioned Turley on the choice, referencing OpenAI’s long-standing mission to distribute superior AI to as many individuals as attainable.

The expectations are excessive for GPT-5, certainly one of OpenAI’s most anticipated product launches since ChatGPT put the corporate on the map in 2022. Since then, ChatGPT has grown into one of many world’s hottest client merchandise, reaching greater than 700 million customers each week — practically 10% of the globe’s inhabitants, in accordance with the corporate.

Many see GPT-5 as a bellwether for AI progress broadly, and the mannequin’s reception by Silicon Valley may have profound implications for Massive Tech, Wall Avenue, and policymakers regulating know-how. These stakeholders are watching to see if GPT-5 presents a big bounce in AI’s capabilities, very similar to its predecessor, GPT-4, which challenged expectations of what software program can do.

GPT-5 presents a slight edge on the competitors

OpenAI claims GPT-5 is state-of-the-art in a number of domains, barely edging out main AI fashions from Anthropic, Google DeepMind, and Elon Musk’s xAI on key benchmarks. Nonetheless, GPT-5 barely underperforms frontier AI fashions in different areas.

The corporate says GPT-5 presents frontier-level efficiency round coding; Altman mentioned the mannequin particularly excels at spinning up whole software program purposes on demand, in what’s turn out to be often called “vibe coding.”

On SWE-bench Verified — a check of real-world coding duties pulled from GitHub — GPT-5 scores 74.9% on its first try. Meaning GPT-5 simply outperforms Anthropic’s newest Claude Opus 4.1 mannequin, which scored 74.5%, and Google DeepMind’s Gemini 2.5 Professional, which scored 59.6%.

On Humanity’s Final Examination — a troublesome check measuring AI mannequin efficiency throughout math, humanities, and the pure sciences — a model of GPT-5 with prolonged reasoning (GPT-5 Professional) scored 42% when utilizing instruments. That’s barely lower than xAI was capable of obtain with Grok 4 Heavy, which scored 44.4% on the check.

On GPQA Diamond — a check of PhD-level science questions — GPT-5 professional scored 89.4% on its first strive, outperforming Claude Opus 4.1, which scored 80.9%, and Grok 4 Heavy, which scored 88.9%.

OpenAI says GPT-5 is healthier for answering health-related questions. On a check measuring accuracy in AI mannequin responses round healthcare subjects, HealthBench Laborious Hallucinations, OpenAI says GPT-5 (with pondering) hallucinates simply 1.6% of the time. That is far decrease than the corporate’s earlier GPT-4o and o3 fashions, which scored 12.9% and 15.8, respectively.

Whereas AI chatbots should not medical professionals, thousands and thousands of persons are utilizing them for well being recommendation. In response to this phenomenon, the corporate says GPT-5 is extra proactive about flagging potential well being considerations and serving to customers parse medical outcomes.

As well as, OpenAI says GPT-5 is healthier than different AI fashions on extra difficult-to-measure, subjective domains, equivalent to inventive design and writing. Turley mentioned GPT-5 responds extra naturally and displays “higher style” than different AI fashions on inventive duties.

“The vibes of this mannequin are actually good,” mentioned Turley.

GPT-5 can be extra correct than OpenAI’s earlier fashions, and the corporate says it suffers far much less from hallucinations — the tendency for AI fashions to make up data — in comparison with its o-series fashions. Hallucinations appeared to be getting worse in OpenAI’s newest AI reasoning fashions, equivalent to o3, and OpenAI beforehand mentioned it didn’t fairly perceive why it was occurring.

In responses to ChatGPT prompts, OpenAI discovered that GPT-5 (with pondering) hallucinates and responds with incorrect data 4.8% of the time. That’s a big discount from o3 and GPT-4o, which rating hallucination charges of twenty-two% and 20.6%, respectively, on the check.

On a benchmark measuring an AI mannequin’s agentic capability to finish simulated on-line duties, Tau-bench, GPT-5 presents blended efficiency. On a part of the check measuring an AI’s capability to navigate an airline’s web site, GPT-5 scores 63.5%, barely underperforming o3, which scored 64.8%. On one other a part of the check measuring AI’s capability to navigate retail web sites, GPT-5 scores 81.1%, underperforming Claude Opus 4.1, which scored 82.4%.

OpenAI additionally says that GPT-5 is safer than its earlier fashions. Whereas AI reasoning fashions often exhibit a bent to scheme towards people or lie to advertise their very own objectives, OpenAI discovered that GPT-5 was misleading at a decrease price than different fashions.

Alex Beutel, OpenAI’s security analysis lead, mentioned lowering deception improves not solely the security of GPT-5, but additionally the consumer expertise, making a mannequin that’s extra “clear and trustworthy in methods customers can belief.”

Beutel additionally notes GPT-5 is healthier at discerning between unhealthy actors who’re attempting to misuse ChatGPT and customers making innocent requests. This leads to GPT-5 with the ability to refuse extra unsafe questions, whereas providing fewer rejections to customers in search of innocent data.

Upgrades for customers and builders

ChatGPT is getting a couple of consumer expertise upgrades as a part of the GPT-5 launch. Customers can now choose from 4 new personalities in ChatGPT’s setting: Cynic, Robotic, Listener, and Nerd. The corporate says these will adapt ChatGPT’s responses with out requiring customers to particularly ask the mannequin to reply in a sure manner.

Subscribers to ChatGPT’s $20-per-month Plus plan get increased utilization limits for GPT-5 than free customers. In the meantime, $200-per-month Professional subscribers can have limitless entry to GPT-5, in addition to a souped-up model referred to as GPT-5 Professional that makes use of extra computational sources to supply higher solutions. Organizations on OpenAI’s Staff, Edu, and Enterprise plans will achieve entry to GPT-5 as their default mannequin subsequent week.

For builders, GPT-5 is coming to OpenAI’s API in three sizes — gpt-5, gpt-5-mini, and gpt-5-nano — which can spend kind of time “reasoning” by duties. Builders may also now management verbosity within the OpenAI API, deciding how lengthy or quick an AI mannequin’s responses needs to be.

The bottom mannequin of GPT-5 will value builders $1.25 per million enter tokens (roughly 750,000 phrases, longer than your entire “Lord of the Rings” sequence) and $10 per million output tokens.

The launch of GPT-5 comes after a busy week for OpenAI. The corporate launched an open-weight reasoning mannequin, gpt-oss, that builders and enterprises can obtain without cost and run at a fraction of the associated fee. The open mannequin practically matched the talents of OpenAI’s earlier high fashions, o3 and o4-mini, however GPT-5 units a brand new normal for frontier efficiency in some areas, equivalent to coding.

Nonetheless, GPT-5 appears to be roughly on par with different frontier AI fashions in a number of areas. Benchmarks, in fact, solely inform a part of the story for any AI mannequin, and it stays to be seen how builders will use GPT-5 in the actual world, and whether or not the mannequin is actually a step above the competitors.