Written by: Techub News Compilation

Introduction

Amidst the buzz following the release of GPT-5, Nathan Lambert, host of the Interconnects AI channel's podcast, invited fellow deep AI observer Sean for an unfiltered conversation. This excerpt taken from a longer live stream reflects the earliest and deepest impressions of users regarding GPT-5. They not only compared the performance of GPT-5 with competitors like Claude and Gemini but also delved into the implications of its unexpectedly low pricing strategy for the industry, as well as whether this signifies a pragmatic adjustment in OpenAI's grand narrative. For practitioners and enthusiasts aiming to understand GPT-5's real positioning and its impact on the AI competitive landscape, this conversation offers valuable perspectives beyond official propaganda.

Summary

• Performance Assessment: GPT-5 shows mixed results in blind testing against Claude 3.5 Sonnet, without overwhelming advantages, especially in specific tasks like coding, where products from Claude and Cursor are still preferred.
• Pricing Shock: The API pricing for GPT-5 (5 cents per million tokens) is far below expectations, even cheaper than GPT-4, and its “Nano” lightweight version also offers great value, being seen as a significant market strategy shift by OpenAI.
• Industry Narrative Shift: The release of GPT-5 resembles an iteration and optimization of the existing product line rather than a revolutionary leap, which may indicate a change in the industry from pursuing a grand narrative of "AGI breakthrough" to a more pragmatic path focused on efficiency and cost as an "exemplary software company."
• The Future of Agents: To achieve "Drop-in Replacement Workers" capable of handling complex, long-term tasks, fundamental progress is still needed in aspects like state management and long-term memory, rather than just relying on linear improvements in model capabilities.

True Performance: The Battle with Claude and the Coding Contest

At the start of the conversation, Nathan Lambert shared his preliminary testing results. Shortly after the early release of GPT-5 (even before Claude Opus 3.5 was launched), he conducted blind tests comparing GPT-5 with models codenamed “2.5 Pro,” “03,” “4.1,” and “Cloud 4 Sonnet.” His conclusion was that GPT-5 displayed inconsistent performance against Claude 3.5 Sonnet, sometimes outperforming and other times slightly lagging behind. He emphasized that this was still under default settings, and he anticipated that enabling the “high reasoning” mode would allow GPT-5 to surpass Sonnet.

However, performance assessment is far from simple benchmarking. Sean and Nathan Lambert quickly turned the topic to more specific application scenarios—coding. Both agreed that for complex codebases, Cursor (which usually integrates GPT models) decisively wins in product experience. Nathan Lambert admitted he had yet to discover any “magic prompt” in Claude Code (or similar products) that made them excel in the coding experience over Cursor. Sean concurred but added that for lightweight, exploratory tasks like quickly building a website or data visualization, he would lean towards using Claude Code or similar tools, as Cursor could similarly perform due to the underlying models being essentially the same.

This discussion highlights the complexity in evaluating current AI tools: model capabilities, product packaging, pricing strategy, and workflow integration all jointly determine user choices. GPT-5 might show improvements in pure capability metrics, but if the corresponding development tools (like integration through API into Cursor) provide a poor experience, users may still shift to other ecosystems.

Pricing "Nuclear Bomb": Market Strategy Shift Behind 5 Cents

If the performance discussion remained within expectations, then the pricing of GPT-5 brought true shock to both observers. Nathan Lambert revealed that the API pricing for GPT-5 is 5 cents per million tokens, which is "absurdly cheap." Even more surprising is that its lightweight version “Nano” does not perform much worse than GPT-5. Sean's reaction was, “That’s really interesting.”

This low pricing strategy raises deeper industry reflections. Sean asked, “Do you think there's still space in the market for pricier but slightly better models? Claude Opus, for instance, is significantly more expensive and, as you said, might be a bit better.” This question strikes at the core of industry competition. Previously, competitors like Gemini attempted to capture OpenAI's market share through lower pricing, while Claude built a reputation with powerful coding capabilities. Now, with GPT-5 simultaneously enhancing capabilities and drastically lowering prices, it undoubtedly resets the competitive baseline in the API market. Nathan Lambert emphasized that this is crucial for OpenAI, especially in the battle for “default choice.”

They cited a related tweet, indicating that if “price” becomes the differentiating factor, it will be a game entirely different from “claiming to build AGI and charging $20,000 a month.” This shift gave Sean a slight “bearish sign”: at the anticipated performance leap of GPT-5's release, OpenAI may have had to pivot its focus to pricing, suggesting that the narrative of driving growth purely through performance breakthroughs might be hitting a bottleneck.

From AGI Narrative to Exemplary Software Company: Has the Industry Turning Point Arrived?

The discussion on performance and pricing ultimately converged on a more macro theme: whether the narrative surrounding OpenAI and the entire AI industry is changing. Nathan Lambert shared his emotional journey: from initial disappointment (“Is this it?”), to gradually changing views with frequent use, to shock at seeing the pricing—“They just won.” This emotional curve from “AI is done” to “we are making a comeback” is quite representative.

Sean attempted to envision the next six months to a year and a half. He believes that the release of GPT-5 reinforced his worldview: AI progress will be a “long and slow march,” with prices and performance continuously pushed higher, but no new model will bring a twofold performance gain. This contrasts with earlier expectations of “each generation bringing a leap in progress.” He positions GPT-5 as a “default model update” of ChatGPT, akin to the previous transition from GPT-3.5 to GPT-4, aimed at “raising the baseline standard,” while also including a high-performing “high reasoning” model.

Nathan Lambert pointedly highlighted the narrative contradiction. He believes that the “AGI narrative” supporting massive industry funding is crucial. If the narrative shifts from “Sam Altman channels Prometheus to bring god-like AI” to “building a potentially trillion-dollar, ubiquitous, and highly useful software company,” then the development rhythm and capital logic of the entire industry could change. He noted that the public backlash against the former grand narrative may increase.

However, both saw another path: if AI tools could become a “must-have” like Microsoft Office for every legitimate company, even if it is “merely very useful software,” it could still support a massive business empire. Sean summarized, “OpenAI is building a potential multi-trillion-dollar software company that will become ubiquitous and highly useful, and then the world continues to function, and people's lives change a bit—these situations can coexist.” This may be the most pragmatic depiction of the current stage of AI commercialization.

The Future of Agents: Long-term Challenges and Technical Path Disputes

Besides evaluating GPT-5 itself, the conversation also touched on the next frontier of AI development: agents capable of genuinely replacing humans in complex tasks. Sean depicted an ideal scenario: an “accountant” agent capable of accessing your email and file system to help you handle tax returns. It needs to manage long time spans, track many states, maintain a to-do list, and require minimal human intervention throughout the process.

So, how far are we from achieving such an agent? Is it just a matter of “tightening screws for another nine months” (ongoing optimization), or do we need a fundamental breakthrough? Nathan Lambert's answer was, “Both.” He listed possible technical paths: updating model weights, trainable KB caches, LoRA fine-tuning, reinforcement learning, or graph retrieval-augmented generation (Graph RAG). These methods sound potentially effective but have yet to be proven feasible.

Sean tended to separate “model progress” from the “tools/framework layer.” He argues that models like Claude Code and Cursor demonstrate that if a model is proficient enough at maintaining context, it can self-record, track to-dos, accumulate context, and provide itself with sufficient “pointers” to acquire information, thereby maintaining state over longer periods without actual continuous learning. This suggests that progress in agents may rely more on innovative engineering frameworks rather than just linear growth in the model's inherent capabilities.

Finally, Nathan Lambert emphasized the genuine intrigue of incorporating the tool’s “inductive biases” directly into model training, rather than merely invoking them during inference. He noted that this is already reflected in DeepSeek's research, and OpenAI has also begun this work on its coding models. He advised the audience to personally compare GPT-4o (or similar models) with GPT-5 side by side in Cursor to feel the difference.

This candid conversation ultimately returned to a simple call to action: everyone should try all models. Users may come to trust certain models but will occasionally discover that their faith is misplaced—this has been the case since the dawn of the AI era. GPT-5 may not be the revolutionary leap many anticipated, but it provides solid improvements at a highly competitive price, and may be quietly rewriting the script of the AI industry from technical fervor to stable commercialization.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。