GPT-5 for Operators: What Actually Changed

Three things changed on August 7, and one thing broke.

OpenAI released GPT-5. It became the default for everyone, including free users. And a router designed to pick the right kind of thinking for each question was, according to coverage that week, largely broken on launch day.² That combination is the most useful thing about this release for a small business owner. The model matters. The dependency matters more.

What is GPT-5, in plain terms?

It is not just a bigger model. TechCrunch described it as a unified model with a real-time router that decides whether to answer quickly or take more time to think, so that users do not have to choose.¹

Availability was broader than past releases. GPT-5 became the default for all free ChatGPT users, which TechCrunch called the first time free users got a reasoning model. Plus subscribers got higher usage limits, Pro subscribers got unlimited GPT-5 plus GPT-5 Pro, and Team, Edu and Enterprise organizations got it as their default the following week.¹

For developers, there are three sizes (gpt-5, gpt-5-mini and gpt-5-nano), control over response verbosity and pricing of $1.25 per million input tokens and $10 per million output tokens for the base model.¹

What do the benchmarks actually tell you?

Something, but less than the headlines suggest.

TechCrunch's rundown: GPT-5 scored 74.9% on SWE-bench Verified, a test of real coding tasks, narrowly ahead of Claude Opus 4.1 at 74.5%. GPT-5 Pro scored 89.4% on GPQA Diamond, a science question set. On a simulated-tasks benchmark called Tau-bench, results were mixed: it scored 63.5% on the airline tasks, below o3's 64.8%, and 81.1% on retail, below Claude Opus 4.1's 82.4%.¹ The article's own summary was that GPT-5 is on par with other frontier models in several areas and slightly behind in some, and that benchmarks tell only part of the story.¹

I read that the way I read any launch: genuinely better in some places, roughly equal in others, and not a reason to rebuild your business on a Tuesday afternoon.

Matt Wolfe, who posts a weekly AI news video, made the same observation on launch day in a video posted on August 7, 2025: "Even OpenAI is starting to admit that the benchmarks mean less and less to most people these days."⁴ He also said the launch left him underwhelmed outside of coding, which is close to my own read.

Plays here, from YouTube

Does GPT-5 hallucinate less?

OpenAI says so, and the numbers are striking. On a health-focused hallucination test, GPT-5 with thinking hallucinated 1.6% of the time, versus 12.9% for GPT-4o and 15.8% for o3. On ChatGPT prompts, it gave incorrect information 4.8% of the time versus 22% for o3 and 20.6% for GPT-4o.¹

These are the company's claims about its own product. They are encouraging, and they are not permission to stop checking. A model that is wrong less often is still wrong sometimes, and the errors tend to be delivered with the same confidence as the right answers. For anything with a number, a legal claim, a medical statement or a price, verify.

What changes for writing, analysis and agents?

Writing

OpenAI says GPT-5 is stronger in subjective domains like creative writing, with its VP of ChatGPT saying the model has "better taste" on creative tasks.¹ It also added four selectable personalities: Cynic, Robot, Listener and Nerd.¹

For an operator, "taste" is only useful if it matches yours. A better default voice is not your voice. Generic voice is the problem I described in The AI Copy Tell. Give any model three examples of your best writing and a list of phrases you never use, then compare outputs. The model that wins on a benchmark may lose on your newsletter.

Analysis

The router and the "thinking" option are the interesting change here, the same dial idea I covered in Claude 3.7 Sonnet. For analysis, such as summarizing a long report, comparing vendors or working through a pricing model, you want the slower, reasoning path. After launch, the picker added Auto, Fast and Thinking, and Altman said GPT-5 Thinking allowed 3,000 messages per week.² That is generous for a single operator, and it is a limit worth knowing about.

Agents

OpenAI had just launched ChatGPT agent in July (see ChatGPT Agent: Delegating Multi-Step Tasks Safely), and multi-step work is where these models are heading. But GPT-5's results on agent-style benchmarks were mixed, a reminder that "agentic" is not one skill.¹ My advice has not changed: keep a human checkpoint before anything that sends, spends or deletes, and test on read-only tasks first.

What did the rocky first week teach?

That the tool you rely on can change underneath you.

At launch, OpenAI deprecated its older models and moved everyone to GPT-5, and the router appeared largely broken on day one. Users complained, and OpenAI restored GPT-4o, GPT-4.1 and o3 for paid users within days.² On August 12, Altman announced the Auto, Fast and Thinking options, and said a coming personality update would feel warmer than GPT-5's current one but, in his words, "not as annoying (to most users) as GPT-4o."²

If you have built a process around "ask ChatGPT to do X," the August lesson is that behavior can shift overnight. People who had tuned prompts to GPT-4o's style suddenly got different output.

How to protect a workflow from model changes

So should you switch?

Probably not wholesale, and probably not this week.

Use GPT-5 for the work where you have already tested it and it won. Keep your tested alternatives. Treat the new default as one more tool, not a replacement for judgment. Every tool I adopt has to earn its place on a task list, not a launch page.

The long arc is clear: these systems are getting more capable, cheaper (as in DeepSeek R1 and the Falling Price of Intelligence) and more available. The advantage will go to operators who keep a calm testing habit, not to the ones who chase every release. That testing habit is built into the AI-assisted systems I build for clients.

Keep building.Ekai