OpenAI’s weblog submit claims that GPT-5 beats its earlier fashions on a number of coding benchmarks, together with SWE-Bench Verified (scoring 74.9 p.c), SWE-Lancer (GPT-5-thinking scored 55 p.c), and Aider Polyglot (scored 88 p.c), which check the mannequin’s means to repair bugs, full freelance-style coding duties, and work throughout a number of programming languages.
Throughout the press briefing on Wednesday, OpenAI post-training lead Yann Dubois prompted GPT-5 to “create a gorgeous, extremely interactive net app for my companion, an English speaker, to study French.” He tasked the AI to incorporate options like every day progress, quite a lot of actions like flashcards and quizzes, and famous that he wished the app wrapped up in a “extremely partaking theme.” After a minute or so, the AI-generated app popped up. Whereas it was only one on-rails demo, the outcome was a modern web site that delivered precisely what Dubois requested for.
“It is an excellent coding collaborator, and likewise excels at agentic duties,” Michelle Pokrass, a post-training lead, says. “It executes lengthy chains and power calls successfully [which means it better understands when and how to use functions like web browsers or external APIs], follows detailed directions, and gives upfront explanations of its actions.”
OpenAI additionally says in its weblog submit that GPT-5 is “our greatest mannequin but for health-related questions.” In three OpenAI health-related LLM benchmarks—HealthBench, HealthBench Laborious, and HealthBench Consensus—the system card (a doc that describes the product’s technical capabilities and different analysis findings) states that GPT-5-thinking outperforms earlier fashions “by a considerable margin.” The pondering model of GPT-5 scored 25.5 p.c on HealthBench Laborious, up from o3’s 31.6 p.c rating. These scores are validated by two or extra physicians, based on the system card.
The mannequin additionally allegedly hallucinates much less, based on Pokrass, a typical problem for AI the place it gives false info. OpenAI’s security analysis lead Alex Beutel provides that they’ve “considerably decreased the charges of deception in GPT-5.”
“We’ve taken steps to scale back GPT-5-thinking’s propensity to deceive, cheat, or hack issues, although our mitigations are usually not good and extra analysis is required,” the system card says. “Specifically, we’ve educated the mannequin to fail gracefully when posed with duties that it can’t clear up.”
The corporate’s system card says that after testing GPT-5 fashions with out entry to net looking, researchers discovered its hallucination charge (which they outlined as “share of factual claims that include minor or main errors”) 26 p.c much less widespread than the GPT-4o mannequin. GPT-5-thinking has a 65 p.c lowered hallucination charge in comparison with o3.
For prompts that may very well be dual-use (probably dangerous or benign), Beutel says GPT-5 makes use of “protected completions,” which prompts the mannequin to “give as useful a solution as attainable, however throughout the constraints of remaining protected.” OpenAI did over 5,000 hours of purple teaming, based on Beutel, and testing with exterior organizations to verify the system was strong.
OpenAI says it now boasts practically 700 million weekly lively customers of ChatGPT, 5 million paying enterprise customers, and 4 million builders using the API.
“The vibes of this mannequin are actually good, and I believe that persons are actually going to really feel that,” head of ChatGPT Nick Turley says. “Particularly common individuals who have not been spending their time enthusiastic about fashions.”
