OpenAI just released GPT-5, here's everything you need to know
OpenAI just announced GPT-5 in this YouTube livestream (I summarised the key insights so that you don't have to watch the whole thing):
https://www.youtube.com/live/0Uu_VJeVVfo?si=Zs2SdYqp1z_zlyMR
What's different about GPT-5?
→ Sam Altman described it as a "PhD level expert" across domains.
→ Unified reasoning: It thinks just the right amount - fast responses for simple tasks, deeper thinking when needed automatically
→ Pro tip: Free users can say "think hard about this prompt" for deeper reasoning (or what OpenAI now call "Thinking"), or paid users can select the "GPT-5 thinking" model for guaranteed deeper reasoning
Who has access and when?
→ ChatGPT rollout: Available today for Free, Plus, Pro, Team users
→ Free tier: Gets GPT-5 access, then transitions to GPT-5 mini when hitting limits
→ Pro users: Get higher limits plus GPT-5 Pro with extended thinking capabilities
→ API: Live today with three models available
Reliability and safety improvements?
→ Fewer hallucinations on complex, open-ended questions
→ Safe-completions feature: Provides the most helpful answer within safety constraints, with fewer blanket refusals and explains the reason why when it can't help
→ Less deceptive behavior when tasks are underspecified or impossible
Voice, learning, and personalization?
→ Voice mode: Sounds more natural / human, translates smoothly between languages, adjustable speed (great for language learning)
→ Study and learn mode: Step-by-step guidance for topics and languages
→ Personality customization (research preview): Choose supportive, professional, concise, or even slightly sarcastic responses
→ Enhanced memory: Better understanding of your preferences and goals over time
→ Gmail/Google Calendar integration: Plan your schedule, find missed emails, manage calendar, essentially a personal assistant inside of ChatGPT
'Software on demand' and coding?
→ Full app development: Builds "complete applications" from scratch, including front-end, with writing, running, debugging, and iteration
→ Improved codebase understanding: Excellent for large complex projects (e.g. for vibe coding with Cursor)
→ According to the SWE-Bench (Software Engineering Benchmark) it's slightly better than Anthropic's best coding model. GPT-5 is 74.9% and Claude 4.1 Opus is 74.5%
API features for developers?
→ Three models: gpt-5, gpt-5-mini, gpt-5-nano (decreasing in capability and cost respectively)
→ Custom tools: Free-form plaintext instead of JSON for complex function calls
→ Structured outputs: Custom JSON schema for complex outputs, and regex or grammar constraints for precise formatting
→ Verbosity control: Low, medium, high settings for response length
→ Extended context length: 400K context window (double GPT-o3's 200K)
Key benchmarks?
→ SWE-Bench: 74.9% (up from o3's 69.1%) Software engineering tasks
→ Aider Polyglot: 88% (up from o3's 79.6%) Multi-language coding
→ MMMU: 84.2% (up from o3's 82.9%) Strong visual reasoning from images
→ AIME 2025: 94.6% (up from o3's 86.4%) Excellent mathematical reasoning
→ COLLIE: 99% (up from o3's 98.4%) Instruction following
→ Biggest benchmark improvements seem to be in coding and math
Pricing and availability?
→ API pricing: gpt-5 at $1.25 per 1M input tokens and $10 per 1M output tokens
→ Cost efficiency: gpt-5-nano is 25x more affordable than gpt-5
→ Availability: API live today, ChatGPT rolling out now (not yet live at time of making this post).
700 million people now have PhD level intelligence in their pocket, for free.
What a time to be alive!
https://www.youtube.com/live/0Uu_VJeVVfo?si=Zs2SdYqp1z_zlyMR
What's different about GPT-5?
→ Sam Altman described it as a "PhD level expert" across domains.
→ Unified reasoning: It thinks just the right amount - fast responses for simple tasks, deeper thinking when needed automatically
→ Pro tip: Free users can say "think hard about this prompt" for deeper reasoning (or what OpenAI now call "Thinking"), or paid users can select the "GPT-5 thinking" model for guaranteed deeper reasoning
Who has access and when?
→ ChatGPT rollout: Available today for Free, Plus, Pro, Team users
→ Free tier: Gets GPT-5 access, then transitions to GPT-5 mini when hitting limits
→ Pro users: Get higher limits plus GPT-5 Pro with extended thinking capabilities
→ API: Live today with three models available
Reliability and safety improvements?
→ Fewer hallucinations on complex, open-ended questions
→ Safe-completions feature: Provides the most helpful answer within safety constraints, with fewer blanket refusals and explains the reason why when it can't help
→ Less deceptive behavior when tasks are underspecified or impossible
Voice, learning, and personalization?
→ Voice mode: Sounds more natural / human, translates smoothly between languages, adjustable speed (great for language learning)
→ Study and learn mode: Step-by-step guidance for topics and languages
→ Personality customization (research preview): Choose supportive, professional, concise, or even slightly sarcastic responses
→ Enhanced memory: Better understanding of your preferences and goals over time
→ Gmail/Google Calendar integration: Plan your schedule, find missed emails, manage calendar, essentially a personal assistant inside of ChatGPT
'Software on demand' and coding?
→ Full app development: Builds "complete applications" from scratch, including front-end, with writing, running, debugging, and iteration
→ Improved codebase understanding: Excellent for large complex projects (e.g. for vibe coding with Cursor)
→ According to the SWE-Bench (Software Engineering Benchmark) it's slightly better than Anthropic's best coding model. GPT-5 is 74.9% and Claude 4.1 Opus is 74.5%
API features for developers?
→ Three models: gpt-5, gpt-5-mini, gpt-5-nano (decreasing in capability and cost respectively)
→ Custom tools: Free-form plaintext instead of JSON for complex function calls
→ Structured outputs: Custom JSON schema for complex outputs, and regex or grammar constraints for precise formatting
→ Verbosity control: Low, medium, high settings for response length
→ Extended context length: 400K context window (double GPT-o3's 200K)
Key benchmarks?
→ SWE-Bench: 74.9% (up from o3's 69.1%) Software engineering tasks
→ Aider Polyglot: 88% (up from o3's 79.6%) Multi-language coding
→ MMMU: 84.2% (up from o3's 82.9%) Strong visual reasoning from images
→ AIME 2025: 94.6% (up from o3's 86.4%) Excellent mathematical reasoning
→ COLLIE: 99% (up from o3's 98.4%) Instruction following
→ Biggest benchmark improvements seem to be in coding and math
Pricing and availability?
→ API pricing: gpt-5 at $1.25 per 1M input tokens and $10 per 1M output tokens
→ Cost efficiency: gpt-5-nano is 25x more affordable than gpt-5
→ Availability: API live today, ChatGPT rolling out now (not yet live at time of making this post).
700 million people now have PhD level intelligence in their pocket, for free.
What a time to be alive!
Comment
