GPT-6 is coming sooner than expected

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

Welcome back to Building LLM Compositions, your biweekly guide to everything new in the field of…composition AI? I’m not sure about this one.

In today’s issue…

• GPT-6 and the dawn of AI’s memory era

• OpenAI agents can now connect to your calendar and email

• Build your own stock portfolio agent

• Ranking the Chinese open model builders

• Why Amazon is betting on agents to win the AI race

…and more

🔍 SPOTLIGHT

GPT-5 was the end of an era. GPT-6 will mark the start of a new one.

Just weeks after the launch of OpenAI’s new flagship GPT-5 model drew mixed reactions, its CEO Sam Altman teased the arrival of GPT-6, saying that it would come more quickly than the 2.5 year gap between GPT-4 and GPT-5, and that it would lean more heavily on memory—recalling details about users in order to better personalize its responses.

Altman also discussed other interesting aspects of OpenAI’s design philosophy for GPT-6, such as its political leanings and the possibility of brain-computer interfaces, but it’s worth zooming in on the comment about memory, because it hints at the closing of the frontier that has defined the rise of AI over the past 5 hectic years—and the opening of another.

Since 2020, when OpenAI released GPT-3 and blew all competing large language models (LLMs) out of the water, the major AI labs have largely raced to improve along two axes: intelligence and price. Making LLMs bigger and training them on more data has lead to remarkable gains on the former, while improvements in their algorithms and the chips they run on collapsed the latter. GPT-5 is so much smarter than GPT-3 was when it first launched, it can’t even be evaluated on the same benchmarks. It’s also 40 times cheaper. Enter chatbots, agents, and the rest of the AI boom.

The moderate improvements it showed over the later GPT-4 versions, though, show that this paradigm is coming to an end. LLMs will continue to get smarter and cheaper, but the low-hanging fruit has been picked, and the lightning-fast progress we saw over the last 5 years is cooling into a slow, steady march. Price, too, has mostly leveled off. GPT-5 costs roughly the same as its predecessor GPT-4.1 (a little less for input and a little more for output), and OpenAI’s rival Anthropic actually raised its prices at the end of last year, something that would have been unthinkable in 2023.

But there are other frontiers available to explore. GPT-5 “remembers” details the exact same way GPT-3 did back in ancient times (early 2020): by stuffing them into its context window, which only holds 128,000 tokens—less than 100,000 words, give or take. This may sound like a lot, but it fills up very quickly when agents run turn after turn, trying to retain crucial details about what they have seen while still keeping room for more documents and data. It’s not a coincidence that I’ve featured so many memory systems for agents in the newsletter lately—everyone’s building them, because they aim to solve what’s become one of the biggest pain points in the field.

Altman’s comments indicate that this is the direction OpenAI is taking with GPT-6. If the company can revolutionize LLM memory the way it did LLM smarts, the results could be even more transformative.