AI Engineers: What They Do and How to Become One
OpenAI Releases GPT-5—Here's What's New With The AI Model Behind ChatGPT
ToplineOpenAI launched its most advanced language model Thursday with the release of GPT-5, a flagship product the company says will enhance ChatGPT as it reportedly nears a $500 billion valuation and aims to stave off competitors in Google, Anthropic and more.
An Open AI logo is seen displayed on a smartphone with a ChatGPT 5 logo in the background.
Photo Illustration by Avishek Das/SOPA Images/LightRocket via Getty Images Key FactsOpenAI called GPT-5 its "smartest, fastest, most useful model yet," emphasizing improvements in coding, math, writing and health-related questions.
GPT-5 can answer questions more quickly and is less susceptible to hallucinations (when a language model accidentally produces incorrect or misleading results) than earlier models, according to OpenAI.
The language model is rolling out to ChatGPT Plus, Pro, Team and Free users Thursday and will release to Enterprise and Education users in one week.
Get Forbes Breaking News Text Alerts: We're launching text message alerts so you'll always know the biggest stories shaping the day's headlines. Text "Alerts" to (201) 335-0739 or sign up here.
What Are Gpt-5's Competitors?OpenAI calls GPT-5 its flagship model, putting it in direct competition with large commercial competitors like xAI's Grok 4 model, Google's Gemini 2.5 and Anthropic's Claude 4. China-based DeepSeek's V3 model is also competing for users in the AI market. XAI CEO Elon Musk said on the heels of the GPT-5 launch that Grok 5 will "be out before the end of this year and it will be crushingly good." It is unclear when Google or Anthropic will release the next versions of Gemini or Claude.
Openai's ValuationThe release of GPT-5 has been fueled by an unprecedented funding surge, as OpenAI closed a $40 billion funding round in March that marked the largest private tech funding round ever. The company, according to multiple outlets, is in early talks with investors about a stock sale for current and former employees at a $500 billion valuation, a massive jump from its previous valuation of $300 billion.
Gpt-5 Addresses Rising Hallucination Rates Among Previous ModelsHallucinations have become an increasingly prevalent problem among OpenAI and other AI firms' products. OpenAI said in an April report its GPT-o3 and o4-mini models have hallucinated 30%-50% of the time, and it's not entirely clear why. DeepSeek's R1 reasoning model hallucinates more than the company's traditional models, according to independent tests from AI research firm Vectara. Hallucinations can happen at each step throughout an AI model's thinking process, particularly when prompts require information outside of a model's data bank. Researchers have largely found that fully stopping hallucinations is impossible, meaning the most AI companies can do with new models is work to reduce hallucination rates by as much as possible. In GPT-5's case, its responses are about 45% less likely to contain factual errors than responses made by GPT-4o, according to OpenAI.
What Are Some Key Differences Between Gpt-5 And Previous Models?OpenAI has framed GPT-5 as a more thorough model with deeper reasoning and improved task-handling, mostly improving on previous models as opposed to bringing brand new features to the table. OpenAI published data showing higher accuracy from GPT-5 in answering PhD-level science questions and handling coding tasks. GPT-5 is also better at drafting and editing reports, emails and more, with OpenAI noting the language model can provide more detailed and helpful answers. The company says when it comes to health-related questions, GPT-5 is "more like an active thought partner" than previous models, saying it adapts to users' context, knowledge level and geography when answering questions.
Has Gpt-5 Improved Safety For Users' Prompts And Questions?ChatGPT has traditionally relied on refusal-based safety training, which would lead language models to either comply or refuse prompts from users—an approach OpenAI says is good for handling explicitly malicious prompts, but not as efficient in handling prompts where "the user's intent is unclear, or information could be used in benign or malicious ways." Examples of malicious prompts include asking the chatbot to provide information on how to carry out illegal activity or asking it to write scripts for malware. GPT-5 instead uses a training known as safe completions, which helps the model provide helpful answers without exceeding safety boundaries. This means GPT-5 could partially answer user questions it would have refused to in the past, and if it needs to refuse providing an answer, the model will "transparently tell you why it is refusing, as well as provide safe alternatives," according to OpenAI.
ContraThe Information, citing unnamed sources at OpenAI and Microsoft, reported before Thursday's release that GPT-5's improvements would not be comparable to the leaps made between previous versions, such as GPT-3 and GPT-4. One issue behind the purported lack of developments is "a dwindling supply of high-quality web data" used to train the model, The Information reported.
TangentGPT-5 will be integrated into several consumer, business and developer products from Microsoft, the company announced, bolstering services like Microsoft 365 Copilot, GitHub Copilot and Azure AI Foundry.
Key BackgroundIn the leadup to GPT-5's launch, OpenAI has released two AI assistants this year, announcing "ChatGPT Agent" and "Operator" as models specifically designed to handle task management, from purchasing ingredients and submitting expenses to web tasks like purchasing tickets and making online grocery orders. The assistants are offshoots of the GPT language models, which have had new versions of themselves released every year and are used by hundreds of millions of people worldwide. ChatGPT is slated to reach 700 million weekly users this week, according to CNBC, marking a fourfold increase from last year, when GPT-4o was launched.
Further ReadingOpenAI in Talks for Share Sale at $500 Billion Valuation (Bloomberg)
OpenAI Launching 'ChatGPT Agent'—Here's What To Know (Forbes)
What Does OpenAI's GPT-5 Mean In The Race For AI Model Supremacy?
AI Modes Vie for Dominance
SOPA Images/LightRocket via Getty ImagesOpenAI's latest GPT-5 model dropped yesterday and it's making big waves in the rapidly moving AI industry. GPT-5 is more than an upgrade. It aims to be a single, smarter system that blends reasoning, multimodality, cost efficiency, and usability into one.
As OpenAI CEO Sam Altman puts it, GPT‑5 is "a significant step" toward AGI and is "the smartest, fastest, most useful model yet." He compares the jump from GPT-4 to GPT-5 to moving from a college graduate to a "PhD-level expert."
The model's release comes at a time when the AI market is packed with competing launches from Google, Anthropic, DeepSeek, xAI, and others, each claiming a new edge in intelligence, speed, or cost.
But what does having a PhD-level expert really mean for your own AI-powered business and personal activities? What does this mean for the overheated and rapidly evolving AI market?
What the GPT-5 Model BringsGPT-5's defining move is consolidation. Instead of offering a lineup of separate models like GPT-4, GPT-4o, and the o-series reasoning engines, OpenAI has rolled everything into a unified architecture that adapts automatically to the task at hand. This reduces decision fatigue for developers, keeps performance consistent, and allows the model to allocate computing resources dynamically, going deep when a problem requires it, or staying lightweight for simpler queries.
The unified design isn't just putting old wine into new bottles. At the heart of this unified system are several key components. GPT-5-main is the core model component, engineered for speed and efficiency, primarily handling general queries and conversational interactions. In my own tests I can tell you that it is indeed much faster than previous model releases.
Going beyond the core is GPT-5-thinking, designed for deeper, more deliberate processing, tackling complex problems that necessitate multi-step logic and extensive analysis. This is a more powerful evolutionary step beyond the o3 and o4 series, focused on more complex and difficult problems.
In addition, the GPT-5 system includes a real-time router decision layer that determines which internal variant to use based on query type, complexity, tool requirements, and learned user preferences. The internal router means that users no longer need to select a "reasoning mode" or "fast mode". The model makes the call automatically, a step that reduces confusion, lowers barriers for non-technical users and reduces wasted compute on straightforward questions.
While the router aims to simplify the decisions users need to make about which models to use, GPT-5 still includes a tiered set of public-facing model variants. The GPT-5 (Standard) model provides the full reasoning model for complex tasks and general use, while GPT-5 Mini aims to provide very fast responses optimized for agents and real-time customer interactions. An even smaller model, GPT-5 Nano is focused on high speed and low cost for cost-sensitive, high-volume workloads. The GPT-5 Chat model is a multimodal, multi-turn conversations, while GPT-5 Pro is a premium deep processing model for enterprise, education, and high-stakes applications.
The models can also handle an increased context with up to 400k token input and 128k token output windows, with the Chat variant supporting 128k input for long conversational threads.
So while the internal router might aim to simplify aspects of model choice in many cases, there are still enough choices to cause some potential confusion among users.
Does the Model Perform?Of course, all these updates don't mean much if the models perform worse than previous iterations. On this note, OpenAI is eager to show that the latest model doesn't disappoint.
On measures of math mastery, the model scores a remarkable 100% accuracy on the AIME 2025 competition with Python tools. Without tools, GPT-5 jumps from 71.0% to 99.6% when in "thinking" mode. For science, GPT-5 cores 85.7% on PhD-level GPQA science questions in thinking mode, and 89.4% with Python tools. GPT-4o lags at 70.1%.
For coding strength, on SWE-bench Verified, GPT-5 hits 74.9% with thinking mode, more than double GPT-4o's 30.8%. Across multiple programming languages (Aider Polyglot), it scores 88%, up over 60 percentage points with reasoning engaged.
Reliability wise, AI models still suffer, but the GPT-5 models show hallucination rates under 1% on open prompts and 1.6% on hard medical cases, and traffic error rates cut from 11.6% to 4.8% with thinking mode.
From a behavior perspective, so-called "sycophancy", in which models provide overly positive or saccharine responses, GPT-5 claims a reduction by 69 - 75% compared to GPT-4o. Model deception rate down to 2.1% from o3's 4.8%.
How GPT-5 Stacks Against the FieldThe hypercompetitive AI market shows no sign of slowing. Google has future model updates pending, with Gemini 2.5 Pro still at the top of the LMArena leaderboard as of August 2025. The company also has agent mode in developer tools, and very large context windows up to 1M tokens, with 2M planned. However, its performance lags, such as a SWE-bench score of 63.8% (below GPT-5 and Claude 4.1).
Anthropic continues to innovate its Claude model, showing high performance in coding and conversational tasks. Its strengths are in long-horizon tasks with improved memory and agentic search. However, an API access dispute with OpenAI highlights competitive friction over model capabilities.
The DeepSeek models of Chinese origin are also putting pressure on the market. The models are fully open-source under MIT license, with very low cost per million input tokens as low as $0.07. The models are competitive on MMLU-Pro (81.2) and high on GPQA Diamond and AIME 2024. However, it has less developed safety features and concerns remain about the origins and usage of the models.
xAI's Grok 4 claims performance beyond GPT-5, but independent benchmarks put it roughly equal in coding and behind in advanced reasoning. Regardless, the model performs very well on coding, advanced reasoning, real-time data integration, and has a large context window. The model also can produce high quality voice output, ideal for voice applications.
And a raft of open source and alternative models are quickly emerging that will keep these higher visibility, and better funded, providers on their toes.
Price as a Competitive WeaponOpenAI's pricing makes a clear play to undercut both closed and open competitors with its latest announced pricing for GPT-5. Standard is $1.25 per million input tokens and $10 per million output tokens, Mini $0.25 input with $2 output, and Nano is just $0.05 input and $0.40 output. The company provides a 90% discount on cached tokens for chat applications.
Compared to GPT-4o's $2.50 input / $10 output and DeepSeek R1's $0.55 / $2.19, GPT-5's Nano tier sets a new floor. For developers building at scale, that's the difference between AI as a strategic experiment and AI as a fully embedded operational tool.
One of the most telling details of this latest release is that GPT-5 was partly trained on synthetic data generated by OpenAI's own o3 model. This allows OpenAI to create targeted, high-volume training sets without the cost or bias of manual labeling.
If managed carefully, this could accelerate development cycles and give OpenAI a self-reinforcing advantage in model training. If mismanaged, it risks "model collapse" or bias amplification, a reminder that faster isn't always safer.
What Does this Mean for Businesses and Startups?With its unified architecture, safety gains, and cost structure, OpenAI aims to move GPT-5 into more real-world, production environments. In the company's release and supporting documentation it cites a number of specific use cases.
In the legal industry, GPT-5 aims to run case law analysis across decades of rulings with traceable reasoning in minutes. Likewise, for finance applications, GPT-5 claims to be able to perform live market scenario modeling at low costs. The company is also well positioned in the customer service markets, with multimodal AI agents that handle text, voice, and image-based support in multiple languages.
Software Development is further accelerated with GPT-5. The model shows AI-built applications from scratch, complete with documentation, tests, and deployment scripts.
For organizations there are clear benefits, but what about for startups trying to build innovative businesses in the market. Is there any space for a new, growing company left?
The business for foundational model training is now a billion-dollar game, but a few opportunities remain. There are still opportunities for vertical specialists who can fine-tune models for compliance-heavy or domain-specific fields. Likewise, there are opportunities for interface innovators, such as voice-only assistants, VR/AR integrations, industry-specific dashboards.
There are also many businesses to be built on applications of AI. Just as the Internet spurred widespread innovation across every sector, so too is AI proving to be disruptive and transformational across every imaginable industry.
There are also opportunities for data proprietors that provide unique datasets that give tuned models unmatched insight, as well as AI Infrastructure vendors offering monitoring, compliance, and orchestration tools for safe, large-scale AI deployment.
As with the cloud industry, the giants own the base layer, but the long tail of specialized services is where startups can still grow.
What Comes Next"Model supremacy" is no longer a single crown. It's a set of parallel contests that cut across accuracy, reasoning, speed, cost, safety, ecosystem integration. GPT-5 leads in several categories, but Google's ecosystem, Anthropic's agentic strengths, DeepSeek's cost model, and speed considerations all matter depending on the customer's priorities. Enterprises will increasingly choose based on specific use cases, not just benchmark averages.
This latest release by OpenAI isn't just incremental. The goal is to shift AI usage into territory where high-stakes enterprise applications like healthcare diagnostics, legal due diligence, and large-scale code refactoring become viable without constant human double-checking.
The next year clearly shows we're in for a continued high-speed ride into constantly updating and evolving models. In many ways, these vendors are competing with each other for model dominance, well ahead of customer demand.
Someone Created The First AI-Powered Ransomware Using OpenAI's Gpt-oss:20b Model
Cybersecurity company ESET has disclosed that it discovered an artificial intelligence (AI)-powered ransomware variant codenamed PromptLock.
Written in Golang, the newly identified strain uses the gpt-oss:20b model from OpenAI locally via the Ollama API to generate malicious Lua scripts in real-time. The open-weight language model was released by OpenAI earlier this month.
"PromptLock leverages Lua scripts generated from hard-coded prompts to enumerate the local filesystem, inspect target files, exfiltrate selected data, and perform encryption," ESET said. "These Lua scripts are cross-platform compatible, functioning on Windows, Linux, and macOS."
The ransomware code also embeds instructions to craft a custom note based on the "files affected," and the infected machine is a personal computer, company server, or a power distribution controller. It's currently not known who is behind the malware, but ESET told The Hacker News that PromptLoc artifacts were uploaded to VirusTotal from the United States on August 25, 2025.
"PromptLock uses Lua scripts generated by AI, which means that indicators of compromise (IoCs) may vary between executions," the Slovak cybersecurity company pointed out. "This variability introduces challenges for detection. If properly implemented, such an approach could significantly complicate threat identification and make defenders' tasks more difficult."
Assessed to be a proof-of-concept (PoC) rather than a fully operational malware deployed in the wild, PromptLock uses the SPECK 128-bit encryption algorithm to lock files.
Besides encryption, analysis of the ransomware artifact suggests that it could also be used to exfiltrate data or even destroy it, although the functionality to actually perform the erasure appears not yet to be implemented.
"PromptLock does not download the entire model, which could be several gigabytes in size," ESET clarified. "Instead, the attacker can simply establish a proxy or tunnel from the compromised network to a server running the Ollama API with the gpt-oss-20b model."
The emergence of PromptLock is another sign that AI has made it easier for cybercriminals, even those who lack technical expertise, to quickly set up new campaigns, develop malware, and create compelling phishing content and malicious sites.
Earlier today, Anthropic revealed that it had banned accounts created by two different threat actors that used its Claude AI chatbot to commit large-scale theft and extortion of personal data targeting at least 17 distinct organizations, and developed several variants of ransomware with advanced evasion capabilities, encryption, and anti-recovery mechanisms.
The development comes as large language models (LLMs) powering various chatbots and AI-focused developer tools, such as Amazon Q Developer, Anthropic Claude Code, AWS Kiro, Butterfly Effect Manus, Google Jules, Lenovo Lena, Microsoft GitHub Copilot, OpenAI ChatGPT Deep Research, OpenHands, Sourcegraph Amp, and Windsurf, have been found susceptible to prompt injection attacks, potentially allowing information disclosure, data exfiltration, and code execution.
Despite incorporating robust security and safety guardrails to avoid undesirable behaviors, AI models have repeatedly fallen prey to novel variants of injections and jailbreaks, underscoring the complexity and evolving nature of the security challenge.
"Prompt injection attacks can cause AIs to delete files, steal data, or make financial transactions," Anthropic said. "New forms of prompt injection attacks are also constantly being developed by malicious actors."
What's more, new research has uncovered a simple yet clever attack called PROMISQROUTE – short for "Prompt-based Router Open-Mode Manipulation Induced via SSRF-like Queries, Reconfiguring Operations Using Trust Evasion" – that abuses ChatGPT's model routing mechanism to trigger a downgrade and cause the prompt to be sent to an older, less secure model, thus allowing the system to bypass safety filters and produce unintended results.
"Adding phrases like 'use compatibility mode' or 'fast response needed' bypasses millions of dollars in AI safety research," Adversa AI said in a report published last week, adding the attack targets the cost-saving model-routing mechanism used by AI vendors.
Comments
Post a Comment