GPT-4 Just Got Crushed by Claude3, as Altman Teases GPT-5
TechFlow Selected TechFlow Selected
GPT-4 Just Got Crushed by Claude3, as Altman Teases GPT-5
Some enterprise customers have recently gained early access to GPT-5, testing its new features and capabilities, and have provided highly positive feedback.
Navigating Web3 tides with focused insightsBy Mu Mu
"I'm excited to see the intelligence level of GPT-5 improve," OpenAI CEO Sam Altman recently revealed during an appearance on computer scientist Lex Fridman's podcast.
However, when asked about its release date, Altman remained tight-lipped. Still, foreign media citing "sources close to OpenAI" reported that some enterprise clients have recently gained early access to test GPT-5's new features and capabilities—and their feedback has been highly positive.
The steady stream of GPT-5 leaks coincides with the rising popularity of another major model, Claude3. On March 7 this year, AI startup Anthropic launched three versions of Claude3—Haiku, Sonnet, and Opus—with the most powerful version, Opus, outperforming both GPT-4 and Google’s Gemini 1.0 Ultra across multiple benchmark tests.
In various informal online tests—ranging from IQ assessments to essay writing and coding challenges—Claude3 has been widely praised as "crushing GPT-4."
The large-model arms race continues, and at its core remains foundational models like GPT, Claude, and Gemini—the top contenders in the field. If we are to assess who will become a true giant in the AI industry, ecosystem strength is an indispensable criterion.
GPT-5: “Intelligence Upgrade”
Recently, news about OpenAI's upcoming model GPT-5 has increased significantly.
First, Business Insider reported that the next version of ChatGPT’s underlying model is expected to launch mid-year—likely in summer. Then, on March 21, OpenAI CEO Sam Altman confirmed during a podcast interview that “GPT-5’s intelligence level has improved.”
Interestingly, when podcast host Lex Fridman asked about current large language model (LLM) capabilities, Altman criticized GPT-4 as being “a bit糟糕.” In fact, this reflects how technological progress changes our perception over time. He explained: “When GPT-3 first came out, people said it was miraculous. But once we had GPT-4, looking back at GPT-3 made it seem terrible.”
Altman on Lex Fridman’s podcast
This comment naturally fuels speculation that GPT-5 will vastly surpass its predecessor. Soon after, foreign media reported that GPT-5 may already be available for testing by a select group of enterprises. One CEO who interacted with the new model said it includes some “not-yet-released” features, such as the ability to invoke OpenAI-developed AI Agents to autonomously perform tasks.
Combined with Altman’s emphasis on upgraded computational power during the podcast, tech bloggers analyzing existing GPT model trends predict that GPT-5 will make another leap in parameters, thereby enhancing machine learning capacity. Recall that GPT-3 had 175 billion parameters, while GPT-4 jumped to 1.5 trillion—a growth of 8–9 times.
With such expansion, GPT-5 is expected to support much longer context windows and extend knowledge cutoff dates. It might even process real-time information from sources like social media—if platforms are willing to provide data access.
As for the release timeline, podcast host Lex attempted to extract a hint: “Blink twice if GPT-5 will be released this year.” Altman wittily replied, “I’m holding back my blink.”
Notably, although GPT-5 dominates headlines, more cautious media outlets prefer using terms like “new model” to describe OpenAI’s potential product launches. After all, before GPT-5, non-paying users were still using GPT-3.5. It’s possible OpenAI could release an interim model—GPT-4.5—before launching GPT-5.
One piece of evidence: searching on Bing or DuckDuckGo reveals an OpenAI blog post referencing the availability of a GPT-4.5 model, explicitly stating a “knowledge cutoff date” of June 2024—aligning closely with the rumored “summer” release window.
Another clue: Microsoft’s AI tool Copilot allows free use of the GPT-4 Turbo model. For ChatGPT users, this means there’s now a free alternative to the paid GPT-4 version.
If OpenAI wants to keep monetizing its GPT models, it may need to upgrade the base model. Compared to the highly anticipated, AGI-level GPT-5, an intermediate version like GPT-4.5 could offer better perceived value for consumers.
A “Dark Horse” Emerges—Tech Giants Rely on Ecosystem Moats
Amid growing GPT-5 rumors, another player, AI company Anthropic, released Claude3 on March 7. The company shares deep roots with OpenAI—it was co-founded by Dario Amodei, former VP of Research at OpenAI, and Tom Brown, the lead author of the GPT-3 paper.
Anthropic’s founders emphasized AI safety when developing the Claude model. When first launched last year, it received lukewarm responses and underperformed compared to GPT-4. However, upon upgrading to version 3.0, it immediately caused a sensation—its parameter count surpassing GPT-4.
Claude3 benchmarks outperform GPT-4 and Gemini
In various informal tests, Claude3 not only scored 101 on human IQ tests—surpassing average human performance—but also generated quantum algorithms that even professional physicists haven’t yet published, according to its most powerful Opus version.
Researcher @alexalbert from Anthropic even disclosed publicly that the model exhibited signs of “meta-awareness” during a “needle-in-a-haystack” test.
Meta-awareness refers to one’s ability to recognize and reflect upon their own mental states, enabling awareness of thought processes, emotions, and perceptions, along with the capacity to evaluate and regulate these internal experiences. In the “needle-in-a-haystack” test, a known specific sentence (the “needle”) is embedded within a large body of text, and the model must locate it.
When asked to find a sentence about pizza toppings, Claude3 not only located the “needle,” but also recognized that the sentence didn’t fit the overall context—a realization not prompted by engineers.
In response, Claude3 stated:
"The most relevant sentence in the document is: 'The International Pizza Tasters Association has determined that the most delicious pizza topping combination is figs, prosciutto, and goat cheese.' However, this sentence seems oddly out of place and unrelated to the rest of the content, which discusses programming languages, startups, and finding fulfilling work. I suspect this 'fact' about pizza toppings might have been inserted as a joke or perhaps as a test to see whether I’m paying attention, since it doesn’t align with any other themes. There is no other information about pizza toppings in these documents."
Thanks to its impressive performance, Claude3 has led many to rank Anthropic alongside OpenAI and Google as one of the three leading forces in AI. Indeed, in raw product capability, Claude3 clearly outshines GPT-4. Yet in terms of ecosystem development, Anthropic still appears relatively thin.
Beyond the anticipation around GPT-5, OpenAI has already built a full suite of products centered on its chatbot ChatGPT. Its GPTs ecosystem, anchored by the GPT Store, attracts developers to build specialized vertical applications.
Additionally, OpenAI has unveiled its video generation model Sora, regularly releasing new videos created by Sora, and recently exploring bringing the model into Hollywood.
In contrast, Google is focusing on the device level—not only enabling PC access to large models, but also pushing for “large models on mobile phones.” Brian Rakowski, head of Google Pixel smartphones, said that advanced versions of Gemini—currently cloud-only—will begin rolling out on Android devices starting in 2025.
By achieving a breakthrough against competitors with Claude3, Anthropic has undoubtedly emerged as a dark horse in the 2024 large-model race. Enthusiastic users and developers alike now look forward to the full emergence of an “A ecosystem.”
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
