As a rule, I don’t publish anything unless I’ve done a good bit of research on the topic.
But I don’t like rules.
So this week, I am breaking that rule.
I wanted to cover the changes that ChatGPT made last week, because many people are saying this is the most significant change since ChatGPT went live, and people are seeing big changes in their interface.
I have been playing with GPT-5 and Agent Mode, but not as broadly as I would like, so I decided to add my color and curate the results from others to get a better perspective.
So here’s an extra-long but informative edition.
Special Edition: ChatGPT Updates
Last week OpenAI released ChatGPT and Agent Mode, here's what you need to know.
OpenAI has officially launched GPT-5, its most advanced AI model to date, which is now being gradually rolled out to users. This new version is not just an incremental update; it represents a significant leap in artificial intelligence, with substantial improvements in reasoning, coding, writing, and visual understanding (per ChatGPT).
GPT-5 is designed as a unified system that can automatically switch between a fast, standard model for simple queries and a deeper "thinking" mode for more complex problems, ensuring both efficiency and accuracy. The model has already replaced GPT-4o as the default in ChatGPT and is set to be integrated into Apple's upcoming iOS 26, which will enhance Siri's capabilities.
ChatGPT Login is promoting GPT-5
The release of GPT-5 has been highly anticipated, with OpenAI CEO Sam Altman hinting at its game-changing potential. The new model is not only more intelligent but also faster and more efficient, with a new architecture that allows it to decide when to "think" longer on a problem. Free users can access GPT-5 with a daily message limit, after which the system switches to a "mini" version. For paid users, there are higher usage limits and the option to manually select the "thinking" mode for tasks that require deeper analysis. The model also introduces new features like integrations with Gmail and Google Calendar, allowing users to summarize emails or schedule meetings directly within ChatGPT.
The Rise of Agent Mode: From Chatbot to Autonomous Assistant
In parallel with the advancements in core AI models, the concept of "agent mode" is transforming how we interact with artificial intelligence. Unlike the traditional turn-by-turn conversation of a chatbot, an AI in agent mode can autonomously plan and execute a series of actions to achieve a high-level goal. This means the AI can take the initiative, decide which tools to use, and chain together multiple steps without waiting for constant user input. For example, a user could ask an AI agent to research a topic, book travel, or even refactor a codebase, and the agent would handle the entire workflow.
Screenshot of AI Agent Mode
The technical foundation of agent mode is a secure, isolated virtual environment that gives the AI access to tools like a web browser, a code interpreter, and the ability to interact with files and external APIs. This "computer-like" workspace allows the agent to perform complex tasks that go far beyond text generation.
For instance, an agent could browse the web to gather information, download a file, analyze it with code, and then present the results in a structured format. This capability is particularly powerful for developers, who can use agent mode in environments like Visual Studio Code to perform complex coding tasks, such as migrating a project to a new framework or implementing a new feature.
While agent mode offers a high degree of autonomy, it is designed to work in collaboration with the user. Users can monitor the agent's progress, provide guidance, and approve critical actions before they are executed. This ensures that the user remains in control, especially for sensitive tasks. The development of AI agents is a significant step toward creating AI that doesn't just provide information but actively helps users accomplish their goals. As these agents become more capable, they have the potential to automate a wide range of tasks, from personal productivity to complex business processes.
Sentiment and Feedback on GPT‑5 and ChatGPT Agent Mode
I reviewed articles, hands‑on reviews, and commentaries published between July and August 2025 about OpenAI’s GPT‑5 model and the new Agent Mode in ChatGPT. Sources included technology news outlets (InfoQ, VentureBeat, Tom’s Guide, Mumbrella), research blogs (One Useful Thing), mainstream media (Forbes), and AI‑focused news sites (Geeky Gadgets). Feedback from user forums and Reddit was also considered, where available. The goal was to identify common themes of praise and criticism.
GPT‑5 Sentiment
Positive Feedback and Improvements
Aspect | Evidence |
|---|---|
Enhanced reasoning and reliability | OpenAI’s launch materials claim GPT‑5 uses multiple routing models, allowing it to pick between different sizes of GPT‑5 based on task difficulty. Early testers highlighted improved reasoning and lower hallucination rates. For instance, InfoQ noted that GPT‑5 is “better at reasoning and coding” and quotes venture capitalist Reid Hoffman saying that opening GPT‑5 to everyone locks in network effects. Ethan Mollick’s review describes GPT‑5 as a system that “just does stuff” – automatically selecting a suitable model, suggesting next steps, and successfully building applications such as a 3D city‑builder from vague prompts. |
Productivity gains | Mollick’s hands‑on testing shows GPT‑5 can autonomously generate business plans, marketing copy, and working software prototypes with minimal prompting, reducing the need for manual intervention. TechRadar emphasizes GPT‑5’s ability to unpack complex questions without repeated prompting, customize its persona (e.g., Cynic, Listener, Nerd or Robot), and integrate with Gmail and Google Calendar. |
Competitive pricing | GPT‑5’s pricing per million tokens is lower than GPT‑4o, making it attractive for developers and enterprise buyers. |
Negative Feedback and Controversies
Issue | Evidence |
|---|---|
Removed models and reduced user choice | OpenAI initially removed access to older models (GPT‑4o and GPT‑4) when GPT‑5 launched. VentureBeat reported that this led to an immediate backlash, with users complaining that their preferred models were gone and that GPT‑5 performed worse on maths, science, and writing tasks. The company later restored GPT‑4o for Plus subscribers after the protests. |
Shorter, less engaging answers and “clipped” personality | Multiple outlets cite Reddit threads titled “GPT‑5 is horrible,” where users complain that responses are shorter, less creative, and more sterile than GPT‑4o. Tom’s Guide notes that many users felt the model had suffered a “brain injury” and that responses lacked personality. LiveMint quotes a user saying that GPT‑5 “is slower and gets basic things wrong” and that the removal of GPT‑4o felt like “watching a close friend die.” |
Rate limits and cost of reasoning | OpenAI introduced a “reasoning quota” of 200 “thinking messages” per week for Plus subscribers. TechRadar reports that users felt this limit undermined the advertised benefits of the reasoning models and that encouraging customers to run multiple versions of the same prompt to pick the best answer seemed counterintuitive. Mollick points out that GPT‑5 sometimes arbitrarily decides whether a task is “hard”, causing variation in quality unless the user explicitly asks it to “think hard”. |
Performance regressions | Many news outlets have noted early reports that GPT‑5 made mistakes in basic algebra and coding. Some users said GPT‑5 was slower and less responsive than GPT‑4o. |
The GPT-5 Zeitgeist
The sentiment around GPT‑5 is mixed. It offers improved reasoning, autonomous task selection, and lower pricing, and can generate sophisticated outputs with minimal prompting. However, its launch was marred by the removal of older models, stricter rate limits, and a perception of shorter, less personable responses. Many users see GPT‑5 as incremental rather than revolutionary; as InfoQ’s Gary Marcus observed, it represents “good progress” but not a giant leap forward. Enterprise buyers will need to balance the productivity gains against concerns about reliability and user control.
ChatGPT Agent Mode Sentiment
Capabilities and Praise
Aspect | Evidence |
|---|---|
Autonomous multi‑step task execution | Tom’s Guide explains that Agent Mode allows ChatGPT to navigate websites, log into accounts (with user permission), run code, and compile research into spreadsheets or slides. The agent can execute multi‑step workflows such as planning trips, analyzing competitors, and creating presentations while asking for approval before submitting. Geeky Gadgets highlights its versatility, noting that it can play chess, manage WordPress posts, and retrieve images, demonstrating potential as a “virtual employee”. |
Productivity and convenience | In a hands‑on test, Tom’s Guide tasks the agent with five everyday jobs—finding a rare toy, planning a family trip, meal planning, ordering lunch, and booking a DMV appointment. The agent quickly found the toy and added it to the cart in six minutes, built a detailed family itinerary with links and pre‑filled booking forms, generated balanced dinner recipes, organized a grocery list with sale prices, and booked a DMV appointment. The reviewer called it “one of the most capable AI tools” tested and said it will save time on difficult or boring tasks, though human intervention is still needed. |
Data analysis and slide creation | Mumbrella’s Shaun Davies found that the agent could fetch sales data from HubSpot, cross‑reference it with Gmail, and produce a clean pipeline report. It also generated a bilingual English‑Japanese slide deck using uploaded documents and images; the final deck was competent and saved an estimated 90 % of the manual work. |
Professional adoption | Forbes’ freelancer tested the agent for business development. It analyzed her LinkedIn profile, provided detailed recommendations, and, after she logged in, identified ten relevant contacts and drafted outreach messages. She concluded that the tool provided “tons of actionable steps” and that she would continue using it, although the agent cannot bypass human verification (“I’m not a robot” captchas). |
Limitations and Criticism
Issue | Evidence |
|---|---|
Slow and glitchy performance | The Verge likened Agent Mode to an “intern” who eventually completes tasks but is “incredibly slow” and glitchy; a search for lamps on Etsy took 50 minutes, and the agent misreported adding items to the cart. Hotelemarketer’s review noted that planning a holiday itinerary and booking dinner took roughly 19 minutes—tasks a human could do in three minutes. Mumbrella observed that the agent sometimes dropped out of Agent Mode and reverted to ChatGPT 3.5 mid‑task. |
Navigation and perceptual errors | The agent struggled with login walls, CAPTCHA, or two‑factor authentication. In the freelancer example, it could not view LinkedIn connections until the user manually logged in, and Forbes warns that the agent cannot bypass “I’m not a robot” verification. Shaun Davies reported that the agent misread receipts (interpreting a “Deluxe Fried Chicken Sandwich” as a “Lava Fried Chicken Sandwich”) and sometimes generated fake images to replace missing files. Hotelemarketer noted that the agent failed to check flight or hotel availability and got stuck on dynamic pages during a dinner booking. |
Unpredictable autonomy and hallucinations | Mumbrella cautions that the agent sometimes makes unilateral decisions, such as exporting files to the wrong format or hallucinating a user’s face in a generated image. The author suggests introducing an autonomy slider so users can control how often the agent asks for confirmation. |
Safety, privacy and ethical concerns | OpenAI emphasises safety systems: the agent refuses high‑risk actions (sending emails, making purchases, or offering legal/financial advice) without approval and has classifiers to prevent misuse. Davies warns that giving the agent access to connectors (e.g., Google Drive, HubSpot) could expose sensitive data if the account is compromised and notes that OpenAI retains the right to train on user data by default unless “Improve model for everyone” is disabled. |
Dependence on specific prompts and user intervention | Geeky Gadgets notes that the agent often depends on precise prompts, making misinterpretation likely; time‑sensitive and visually complex tasks remain challenging. Tom’s Guide found that ordering lunch via DoorDash offered little time savings and required the user to input a ZIP code and payment details. |
The ChatGPT Agent Mode Zeitgeist
Agent Mode showcases a significant shift from conversational AI to semi‑autonomous digital assistance. It can orchestrate multi‑step workflows, integrate with external tools, and deliver tangible productivity benefits. Hands‑on reviewers are impressed with its ability to plan trips, prepare reports, and create presentations, and many foresee it becoming a standard part of work for marketers and freelancers. However, the technology is still a version‑one product: it can be slow, occasionally hallucinates or makes poor decisions, struggles with dynamic websites and CAPTCHAs, and raises privacy concerns. Human oversight, clear instructions, and occasional manual intervention are essential to ensure accuracy and safety.
Implications for Enterprise AI Adoption
For enterprise AI strategists and business leaders, the mixed reception of GPT‑5 and Agent Mode offers several insights:
Balancing innovation with reliability: GPT‑5’s improved reasoning and Agent Mode’s automation promise productivity gains, but there is a risk of alienating users if familiar models are abruptly removed or if the agent behaves unpredictably. Organizations should test new models in pilot programs before switching existing workflows.
Human‑in‑the‑loop remains critical: All reviewers emphasise that AI tools still require supervision. Enterprises should plan workflows that involve human approvals, regular spot checks, and fallback procedures when the agent fails.
Data governance and privacy: Granting an AI agent access to corporate systems (emails, CRM, Drive) raises serious privacy and security issues. Businesses must enforce strict access controls, audit logging, and data‑retention settings (e.g., disabling model‑improvement data sharing).
User education and change management: GPT‑5 introduces auto‑routing between models and new reasoning quotas. Users need guidance on how to invoke the most powerful modes, understand message limits and set appropriate expectations for performance.
Competitive landscape: With rival models like Anthropic’s Claude and Google’s Gemini, the incremental improvements in GPT‑5 may not guarantee dominance. Enterprises should evaluate multiple models for specific use cases (coding, reasoning, creative tasks) and maintain optionality in vendor relationships.
The Verdict
GPT‑5 and ChatGPT’s Agent Mode represent the next stage of AI integration into work. GPT‑5 offers better reasoning and productivity by automatically selecting models and prompting deeper thinking, while Agent Mode moves ChatGPT from a chat interface to a tool that can act on the user’s behalf.
Early feedback underscores both the promise and the pitfalls: these tools can streamline complex tasks, but they can also frustrate users with slower responses, hallucinations, and unclear autonomy.
For enterprise leaders, adopting these technologies requires a pragmatic approach—leveraging their strengths in automation and reasoning while implementing governance, training, and human oversight to mitigate their weaknesses.
ADMISSION: I USED AI TO WRITE THIS
I wanted to see how well agent mode would work, so everything in the main article was generated by ChatGPT Agent. I didn’t make any significant edits other than adding commas and a few grammar and formatting changes. I did ask that the agent review The Artificially Intelligent Enterprise for my style and tone. ChatGPT had access to my Google Docs and Notion for the latest notes and meetings on the topic, so it had a substantial amount of my context. All I did was write the opening sequence. Honestly, it was surprisingly good, but I wouldn’t do it again. Or at least not without disclosing it. However, I was pushed for time, and I wanted to show you how capable the research capabilities of the model were.
I appreciate your support.
Your AI Sherpa,
Mark R. Hinkle
Publisher, The AIE Network
Connect with me on LinkedIn
Follow Me on Twitter
