Everything you need to know about GPT-5 (+ mini and nano)

GPT-5 arrives with a pricing structure and routing system that make it practical to use more often—especially the “mini” and “nano” tiers—while OpenAI’s safety and reliability work targets long-standing pain points like sycophancy, hallucinations, and instruction leakage. The headline takeaway is cost-efficiency at scale: GPT-5’s standard model is priced at $1.25 per million input tokens and $10 per million output tokens, and the smaller variants are dramatically cheaper, with GPT-5 mini at $0.25 per million input and $2 per million output, and GPT-5 nano at $0.05 per million input and $0.40 per million output. That combination, plus token caching discounts, is positioned as a shift from “pick the cheapest model” to “choose the right effort level” without blowing up budgets.

The transcript spends significant time putting those numbers in context. Compared with other popular models—citing examples like o3, o1, Claude 4 Opus, and Sonnet—GPT-5’s token rates are framed as meaningfully lower, and the speaker argues that real-world cost can diverge from sticker price because different models generate different token volumes. Benchmarks like SkateBench are used to illustrate that gap: GPT-5 is described as far cheaper per test run than Grok 4, with the speaker claiming roughly two orders of magnitude difference in average cost. The same logic is applied to why “mini” and “nano” matter: they’re portrayed as strong enough to replace cheaper-but-less-capable options, including common Gemini workflows.

Beyond price, GPT-5’s operational profile is defined by a large context window and high output capacity. The input context is said to land at 400,000 tokens, while output can reach up to 128,000 tokens per request. An official knowledge cutoff is given as September 30, aligning with a simple “recent League of Legends champion” test used to infer cutoff timing.

A major theme is model routing inside a “unified system.” Instead of forcing users to manually swap models, GPT-5 is described as combining a fast general model, a deeper reasoning model for harder tasks, and a real-time router that selects which one to use based on conversation type, complexity, tool needs, and explicit intent (such as “think hard”). The routing system is said to be continuously trained on user and preference signals, though details aren’t fully exposed.

Safety and reliability claims form the other pillar. The transcript highlights “safe completions” as an alternative to brittle refusal-only training, aiming to reduce failure modes when user intent is obscured. It also emphasizes reduced sycophancy: offline evaluations reportedly show GPT-5 main performing nearly three times better than GPT-4o (with GPT-5 thinking outperforming both), and the model is described as harder to jailbreak. Instruction hierarchy is presented as another guardrail—system instructions should override developer instructions, which should override user instructions—paired with tests for secret extraction attempts and “access granted” style prompts.

Finally, the transcript ties the technical changes to user experience: GPT-5 is portrayed as more likely to follow instructions and integrate with tooling in a controlled way, while long-context reasoning benchmarks and SVG generation tests are used as proof points. The overall message is that GPT-5 isn’t just “a bit better”—it’s engineered to be cheaper to run, easier to select correctly, and more dependable across safety, accuracy, and tool use, making it a default workhorse rather than a model you only switch to for special tasks.