When Sam Altman said GPT-5 made him feel “useless relative to the AI,” he was describing a specific moment, not making a scientific claim that humans had suddenly become obsolete. In a July 2025 conversation with Theo Von, Altman said he had received an email question he did not fully understand, put it into GPT-5, and watched the model answer it in a way that he found remarkably difficult for himself to match. Contemporary reporting from the interview and Altman’s GPT-5 comment captured the moment in the context of GPT-5’s impending release.
The quote became memorable because it captures something more interesting than benchmark scores.
It raises a practical question:
What changes when an AI system can handle tasks that once required specialized human reasoning, and when people begin to rely on that capability as part of their everyday work?
GPT-5 launched on August 7, 2025 as OpenAI’s next major model family, bringing together reasoning and general-purpose capabilities in ChatGPT and adding improvements across coding, writing, vision, health, instruction following, and tool use. OpenAI described it as a unified system that could decide when to answer quickly and when to spend more effort reasoning through a difficult task. OpenAI’s original GPT-5 announcement provides the company’s technical overview and published evaluations.
There is also an important current context point. By August 2026, GPT-5 is no longer OpenAI’s newest frontier model. OpenAI now describes GPT-5 as a previous model and recommends its newer generation for current development, while GPT-5.5 has been deployed for more advanced real-world work.
So why does GPT-5 still matter?
Because GPT-5 marked a shift in how many people thought about AI capability. It was not simply another chatbot update. It pushed harder into reasoning, coding, multimodal understanding, research, and agentic work.
What Sam Altman actually meant by “useless relative to it”
The phrase sounds more extreme when removed from context.
Altman did not say that humans were useless in general. He described a particular experience in which GPT-5 handled an intellectual task that he personally struggled with.
His reported reaction was essentially:
I encountered a difficult question. I couldn’t understand it well enough to solve it. GPT-5 did. The gap felt uncomfortable.
That distinction matters.
The interesting part of the story is not that a language model answered an email. AI systems had already been doing that.
The interesting part is the subjective experience of encountering a system that appears to outperform you at a task you thought should be within your own capabilities.
That is psychologically different from using software as a calculator.
A calculator is obviously better at arithmetic.
A spreadsheet is obviously better at repetitive calculations.
But when a system begins to appear better at reasoning, interpretation, writing, coding, and research, people naturally start asking where their own advantage remains.
GPT-5 was designed around more than language generation
OpenAI’s original GPT-5 release described it as a unified system combining a fast model, a deeper reasoning model, and a router that decides which approach to use based on the conversation and task. This was intended to reduce the need for users to manually choose between separate reasoning and general models. OpenAI’s description of GPT-5’s unified system explains the architecture and the reasoning model behind that design.
That design matters because many real-world tasks are not purely “generation.”
Consider a request such as:
Analyze this codebase, identify the likely cause of the bug, propose a fix, explain the trade-offs, write the patch, and test it.
That’s a multi-stage reasoning task.
Or:
Review these documents, compare their claims, identify inconsistencies, research the missing information, and produce a concise recommendation.
Again, the task requires multiple cognitive operations.
GPT-5 was increasingly built around that style of work.
Coding was one of the clearest demonstrations
GPT-5’s coding capabilities were a major focus of the release.
OpenAI reported 74.9% on SWE-bench Verified and 88% on Aider Polyglot in its developer evaluation results. The company described GPT-5 as a stronger coding collaborator for debugging, editing code, answering questions about complex codebases, and handling agentic development tasks. OpenAI’s GPT-5 developer evaluation results document those figures and explain the model’s coding and tool-use capabilities. The SWE-bench Verified benchmark defines Verified as a human-validated subset designed for reliable evaluation of coding agents and language models.
That is where the “useless relative to it” idea becomes more concrete.
If AI can write a function, that’s useful.
If AI can inspect a large repository, understand the relationships between components, identify a failure, propose a solution, modify multiple files, and reason through testing, it begins to resemble a technical collaborator.
That does not eliminate software engineers.
It changes the division of labor.
The human increasingly becomes the person who:
● Defines the objective
● Establishes constraints
● Evaluates the proposed architecture
● Reviews changes
● Tests outcomes
● Owns the final decision
The machine takes on more of the implementation work.
GPT-5 also pushed further into multimodal AI
AI becomes more useful when it can reason across multiple forms of information.
GPT-5 supported text and vision, and OpenAI reported an 84.2% score on MMMU, a benchmark designed to test multimodal understanding. It also reported strong results in math, coding, health, and other areas. GPT-5’s multimodal and benchmark results show how the model was evaluated across more than straightforward text generation. The MMMU benchmark evaluates multimodal understanding and reasoning across college-level subjects using text and image information.
This matters because real work rarely arrives as pure text.
A business problem might involve:
Emails + spreadsheets + diagrams + screenshots + documents + images + code
A modern AI assistant needs to reason across those inputs rather than treating every format as a separate workflow.
That is one reason multimodal AI is becoming central to the broader AI competition.
AI coding and AI research are becoming connected
Coding is only one part of the change.
GPT-5 was also positioned as a model for research and complex analysis. OpenAI highlighted improvements in instruction following, tool use, factuality, and tasks requiring longer chains of reasoning.
This suggests a broader trajectory:
AI doesn’t just generate an answer. It increasingly helps complete the process that leads to the answer.
That distinction becomes important when comparing traditional chatbots with AI agents.
A traditional chatbot might tell you how to research a problem.
An agentic system may be able to:
1. Search for information
2. Compare sources
3. Write notes
4. Analyze data
5. Generate a report
6. Use software tools
7. Check its own work
8. Continue until a specified task is complete
GPT-5 moved further toward this model.
The trajectory continued with GPT-5.5. OpenAI’s 2026 release describes GPT-5.5 as designed for complex real-world work including online research, data analysis, documents and spreadsheets, software operation, and tool use, with a stronger emphasis on carrying tasks through to completion. OpenAI’s GPT-5.5 announcement illustrates how the capabilities first emphasized with GPT-5 are continuing to evolve.
The real breakthrough may be reduced friction
One reason GPT-5 felt different to many users was not simply raw intelligence.
It was friction reduction.
Older AI workflows often required users to understand:
● Which model to choose
● When to use reasoning
● Which tool to invoke
● How to structure prompts
● How to break complex tasks into steps
GPT-5’s unified-system design attempted to hide some of that complexity.
That sounds like a minor product decision, but it changes how people interact with AI.
Instead of learning the architecture of the AI system, users can increasingly describe the outcome they want.
The interface becomes:
Goal → AI planning → tools → execution → result
That is much closer to working with an assistant than operating a software component.
GPT-5 did not eliminate mistakes
This is where enthusiasm needs to be balanced.
A more capable AI is not a perfectly reliable AI.
OpenAI explicitly described GPT-5 as less prone to hallucination and better at communicating its limitations, but its system-card work still involved extensive safety testing because powerful models can fail in consequential ways. GPT-5’s system card and safety analysis provides the deeper technical context for its safety evaluation.
This distinction is essential:
Better reasoning does not equal guaranteed truth.
A model may produce an impressive answer and still be wrong.
It may misinterpret the user’s objective.
It may rely on an incorrect assumption.
It may fail on an unfamiliar edge case.
It may sound more confident than the evidence justifies.
The more powerful the system becomes, the more important it is to build workflows that allow verification.
Why “feeling useless” is actually a useful warning
Altman’s comment is interesting because it exposes a tension in AI development.
We want AI to become more capable.
But when it succeeds, humans may feel less capable by comparison.
That is not necessarily a reason to stop building better AI.
It is a reason to reconsider what human expertise means.
Suppose a researcher can ask AI to process thousands of papers in minutes.
The researcher’s comparative advantage is no longer reading every paper manually.
It may shift toward:
● Choosing the right question
● Designing the experiment
● Evaluating sources
● Recognizing weak assumptions
● Interpreting ambiguous findings
● Deciding which conclusions are worth acting on
Likewise, a programmer who has an AI coding partner no longer needs to spend as much time typing boilerplate.
The scarce skill becomes architectural judgment.
In other words, AI can make some forms of expertise less scarce while making judgment about AI output more valuable.
GPT-5 and the changing meaning of expertise
This is perhaps the most important lesson from the GPT-5 story.
Human expertise used to be closely tied to knowing information and producing work manually.
AI changes both.
If a model can remember more facts, write faster, and perform more calculations, then memorization and manual execution become less important differentiators.
The valuable skills increasingly become:
Problem formulation
Can you define the right problem?
Verification
Can you determine whether the AI is right?
Context
Can you understand what information matters?
Taste
Can you distinguish a technically correct output from a genuinely good one?
Responsibility
Can you decide what should actually be done?
These capabilities become more important precisely because the AI is capable.
What ChatGPT-5 pricing meant for developers
GPT-5 was also notable because it was exposed through an API with different model sizes and explicit cost controls.
OpenAI launched gpt-5, gpt-5-mini, and gpt-5-nano, with prices of $1.25 per million input tokens and $10 per million output tokens for GPT-5 at launch.
OpenAI’s current API documentation now labels GPT-5 as a previous reasoning model and recommends its newer models for current development, while listing the original model’s pricing and 400,000-token context window. OpenAI’s current GPT-5 model documentation provides the present status of the original GPT-5 model.
That matters for anyone evaluating GPT-5 today.
If you’re researching the model historically, GPT-5 is a milestone.
If you’re choosing an OpenAI model for a new production application in 2026, you should also evaluate the newer GPT-5.x generation.
Was GPT-5 really that transformative?
The honest answer is more nuanced than the headlines.
GPT-5 was a meaningful step forward.
OpenAI reported substantial gains across coding, math, health, multimodal understanding, and other benchmarks. Its unified architecture also attempted to make reasoning more automatic and reduce the burden of model selection.
But benchmarks do not equal general intelligence.
A model can improve significantly on standardized tests without becoming universally reliable.
It can solve more difficult coding tasks without understanding a software project’s business context perfectly.
It can answer complex questions without eliminating factual uncertainty.
It can reason better while still requiring human supervision.
So the right interpretation is not:
GPT-5 made humans obsolete.
It is:
GPT-5 pushed AI further into tasks that humans previously regarded as distinctly intellectual.
That distinction is much more important.
What GPT-5 means for ordinary users
The impact isn’t limited to developers or AI researchers.
A typical user can use advanced AI for:
● Writing
● Learning
● Research
● Data analysis
● Planning
● Coding
● Brainstorming
● Document analysis
● Image understanding
● Complex problem solving
OpenAI’s current GPT-5 product positioning still emphasizes writing, research, coding, analysis, and problem solving as core use cases, while newer GPT-5.x systems extend those capabilities further into longer, more autonomous workflows.The practical lesson is that AI literacy is becoming less about knowing how AI works internally and more about knowing how to delegate work effectively.
Good AI users increasingly behave like managers.
They specify outcomes.
They provide context.
They establish constraints.
They check important results.
They ask for revisions.
They decide what to trust.
The future is probably not human versus AI
The most useful conclusion from Altman’s “useless” comment is not that humans are losing a competition.
It is that the comparison itself may be misleading.
A human and an AI do not have the same strengths.
AI can:
● Process information extremely quickly
● Search large volumes of material
● Generate many alternatives
● Perform repetitive reasoning
● Work continuously
● Scale across many tasks
Humans can:
● Set goals
● Understand social context
● Establish values
● Take responsibility
● Interpret ambiguity
● Build trust
● Decide what is worth pursuing
The interesting future is what happens when those capabilities are combined.
That is already the direction implied by OpenAI’s newer systems, which increasingly focus on tool use, computer interaction, online research, coding agents, and long-running tasks. GPT-5.5 is a particularly clear example of that trajectory.
Final Takeaway
Sam Altman’s claim that GPT-5 made him feel “useless relative to the AI” became a powerful symbol because it described something many people had already begun to experience: AI was no longer impressive only because it could generate fluent text. It was becoming capable of handling increasingly difficult intellectual tasks.
GPT-5 helped accelerate that shift through stronger reasoning, coding, multimodal understanding, tool use, and a unified system that could decide when more deliberate reasoning was needed. OpenAI’s original GPT-5 research and release documentation provides the primary record of those capabilities and evaluations.
But GPT-5 did not make human expertise irrelevant.
It changed where expertise matters.
As AI becomes better at execution, humans have to become better at choosing the problem, evaluating the result, understanding the context, and deciding what deserves to happen next.
That may be the deeper meaning behind Altman’s uneasy reaction.
The future of AI is not simply a machine becoming better at what humans do.
It is a world in which the definition of valuable human work changes because machines can do more of it.
