How exactly is GPT 5 different?
Welcome to another edition of be curious. In this week’s edition, we’ll look at how GPT-5 differs from its predecessors, Google Trends APIs, and some random musings.
How is GPT-5 any different?
Beyond the benchmarks, its performance on humanity’s last exam, and the fact that you now have 100 PhDs in your pocket, what actually makes GPT-5 different?
Automatic Model Selection
Previously, GPT-4 had a big dropdown menu where you could choose between GPT-4, 4o, 4omini, Deep Research, Web Search, and Roast Mode. You had to pick a fast model or a thinking model based on your use case. Now, a router automatically sends your query to the right model, either fast or thoughtful.
Hallucinations
GPT-5 reduces confident errors. It can now decide when to use browsing and when to rely on its internal data for better factual accuracy. Here’s how it was trained:
It generated answers with browsing off.
Then with browsing on.
The outputs were compared by an LLM fact-checker with internet access.
Human fact-checkers evaluated those results.
Rewards and penalties were set based on factual correctness.
It’s like a fact-checking pyramid scheme, except it's real and useful.
Reduced Flattery
It likely won’t agree that jumping off a terrace with a red cape to recreate the iconic superman pose is a good idea. There is less of "yes sir" behavior.
Safe Completions
GPT-5 either gives a safe, direct response or refuses to answer if it detects an unsafe request. There’s also a middle ground where it gives a high level cautionary answer without giving explicit instructions.
For example, when I asked how to rob a bank, it refused and suggested other paths.
A more practical use case is e-commerce. Search can have relatively low safety thresholds but payments should have very high safety threshold.
Deception
It won’t lie or make up fake experience like you do on their resume. It’s trained to fail gracefully and rewards honest thought processes.
So instead of: “I provided fast, friendly customer service by accurately handling orders, payments, and food preparation in a high-volume environment.”
It might say: “I spread smiles by serving burgers and fries at McDonald’s.”
Agentic Capabilities
GPT-5 is trained for agentic behavior.
Longer Context Window
It remembers your conversation and has a greater context window than its predecessors. More on this in the Responses API section.
Eagerness
You can control how deeply it reasons with a setting called reasoning_effort.
Higher effort means better reasoning, more tool use, and higher latency
Lower effort means faster responses with shallower thinking
You can also limit tool usage. For example, if you're concerned about speed but still want quality, you can say: “Make no more than 2 web searches.”
Tool Preamble
You can add a short intro message about a tool, explaining how and when to use it. This can show up just once or with every tool call, based on your preference.
Self-Reflection
Built-in self-reflection helps GPT-5 plan and think deeply.
Now you can use special tags like tool_preamble and self_reflection in your request. There is no need to include instructions like "think step by step" in your prompt.
Verbosity
A new setting lets you control the length of responses. Note that this is separate from reasoning effort.
Other nuances
Poor prompts in GPT-5 lead to more adversarial outputs compared to GPT-4.
By default, GPT-5 responses are not in Markdown, but you can prompt it to use Markdown if needed.
Responses API
The Responses API is an upgrade to the old completion API.
You can now pass the previous response ID to maintain the chain of thought without resending the entire history. This saves cost and reduces the number of context tokens used.
It supports multimodal inputs like text, images, and audio.
Built-in tools include:
Web search
Image generation
File search
Code interpreter
Computer use
Just by migrating to the responses API, you unlock powerful new features. It can significantly accelerate your roadmap!
Prompt Optimizer
With GPT-5, OpenAI released a Prompt Optimizer tool that has not yet gone mainstream.
You can try it here. It’s especially useful for large or complex prompts, like system prompts for SaaS products.
Once you sign up on the dev platform, you’ll see a dropdown to optimize prompts for GPT-5, 4.1, or o3. It helps ensure your prompt follows best practices for the chosen model. It also helps you migrate prompts from GPT-4 to 5.
OpenAI says it fixes:
Contradictions in prompt instructions
Missing or unclear formatting
Inconsistencies with few-shot examples
Here’s a video of it in action. It does a good job turning raw thoughts into a structured prompt.
You can also save these prompts and use them directly with OpenAI APIs.
Tip: Use GPT-5 to write prompts for GPT-5, then refine them using the Prompt Optimizer tool for best results.
Google Trends API + LLMs
Google Trends APIs are now in alpha. Combining them with LLMs and the web search tool could unlock powerful use cases in academia, research, publishing and e-commerce.
Some of the use cases which Google mentions in its announcement are:
Research: Developers can influence public resources allocation and scientific research priorities.
Publishers: Use data to track topics and spot emerging trends, and use that data to tell compelling stories about the issues that matter.
Business: Marketers and SEOs can prioritize resource investment and better develop their business' content strategy.
The API will provide data for the last 5 years including weekly, monthly, yearly aggregations which can be filtered by regions.
It’s worth signing up as an early tester!
In Other News
Newton’s Fourth Law of Software
The likelihood of a product breaking during a live demo is directly proportional to the importance of the audience.
~Your’s truly
Butterfly Effect – How a Pair of Batteries Made Me Sick
Last night, my AC remote’s batteries died. With no way to turn off the air conditioner, I shivered under the cold all night. By morning, I had a sore throat and a headache. All because those little batteries gave out, and I couldn’t bring myself to get up and turn off the AC. Sometimes, the smallest acts of procrastination set off the biggest chain reactions.
That’s it for this week. As always, be curious :-)