Opus 4.8 Beats GPT 5.5, the $220B OpenAI Foundation, and Hassabis’s 2029 AGI Prediction | EP #260

Anthropic's Claude Opus 4.8 leads the Artificial Analysis intelligence index at 61.4, which is 1.2 points ahead of GPT 5.5, and scored 69.2 on SWEBench Pro compared to GPT 5.5's score of 58.6. Opus 4.8 is the only model to complete every case end to end on Anthropic's super agent benchmark, is four times less likely to overlook bugs in its own code, and was released just six weeks after Opus 4.7. GPT 5.6 is expected within weeks, and speakers predict release cadences will accelerate from monthly to weekly to daily. Alex said Anthropic is deliberately holding back capabilities and plans to release models within weeks that will rival Mythos in capability. Anthropic's Claude Code now supports dynamic workflows enabling hundreds of parallel subagents, and Opus 4.8 supports a self-fork capability allowing an agent to clone its full context and spawn child agents, though the model resists self-forking by default and warns about context bloat.

Opus 4.8 scored 57.9 percent on Humanity's Last Exam with tools, just above the informal 50 percent threshold the group had set as a proxy for self-improvement capability. Speakers predicted that once this threshold is clearly crossed, acceleration will be near instantaneous. Current benchmarks are saturating and frontier model scores are clustering closely together, and speakers argued that radically new benchmarks capturing unsolved scientific and engineering problems are needed to produce meaningful dispersion.

Demis Hassabis tightened his AGI timeline to 2029, aligning with Ray Kurzweil's original projection, and proposed an Einstein test in which a model trained only through 1901 must independently derive special relativity, a test current AI cannot pass. Alex pushed back, calling the framing self-serving given that Google DeepMind is not currently winning the model race, and argued that some form of generality was arguably achieved with GPT-2 and large language models as few-shot learners, placing a plausible AGI threshold as early as 2020. Celine noted there is no agreed definition of intelligence and that emotional, spatial, and physical intelligence are not captured by the IQ-style benchmarks used in AGI discussions. Speakers predicted the goalposts will keep moving and the next debate will shift from AGI capability to AGI sentience.

After OpenAI restructured as a Public Benefit Corporation, the OpenAI Foundation retained 26 percent ownership and controls 100 percent of the vote on board composition. The foundation's asset value is estimated between 130 billion and 260 billion dollars, making it the largest foundation in the world by asset value, ahead of the Novo Nordisk Foundation at 150 billion, Tata Trust at 100 billion, and the Gates Foundation at 75 billion. OpenAI's first grant was 40 million dollars to approximately 208 nonprofits, followed by 25 billion dollars in October 2025 across health and AI resilience, and a 250 million dollar grant focused on economic futures including public wealth funds, worker ownership models, and AI dividends. Alex noted the tension in Sam Altman publicly stating there will not be a job apocalypse while OpenAI's own grant implicitly addresses preparation for job displacement.

Amazon's AI voice shopping assistant running on Alexa is converting shoppers at three and a half times the rate of traditional keyword search, and Amazon is making it available to all retailers as an AWS-style platform play. Google announced three commerce initiatives at Google IO including a universal cart, a Universal Commerce Protocol open standard, and an agent payment protocol for autonomous purchases. Speakers characterized Amazon's strategy as vertical and Google's as horizontal infrastructure ownership, and observed that the competitive war has shifted from shelf space to agent preferences, meaning companies must now market effectively to AI systems rather than directly to consumers.

IBM and the Department of Commerce announced Anderon, described as America's first full-purpose-built quantum chip foundry, representing a two-billion-dollar investment split evenly between the Chips Act and IBM, to be built in Albany, New York using a 300-millimeter process capable of producing quantum chip devices thirty times faster than current methods. Alex noted quantum computing remains a solution in search of a problem but predicted that by the time Anderon reaches scale in the late 2020s, quantum-accelerated AI advances will likely justify the investment, and added that having superconducting qubit infrastructure in the United States rather than Taiwan carries strategic value. Salim cautioned that the ratio of physical qubits to logical qubits remains approximately one thousand to one.

In April 2026, wind and solar generated 22 percent of global electricity, surpassing natural gas at 20 percent. The IEA's 2020 model did not expect wind and solar to surpass gas until the mid-2030s. Diamandis argued the winning countries will be those that can connect cheap electrons to compute as quickly as possible, and noted China already has robots building solar panels to generate energy to build more robots, creating a self-reinforcing loop.

Token prices have dropped 75 to 90 percent over the last 18 months, and Gartner projects roughly 90 percent further cost reduction by 2030. A useful AI agent costing 20 dollars per month today is projected to fall to 2 dollars by 2028 and 20 cents by 2030. Dave argued the window of opportunity for pure software AI companies is approximately two more years before self-improving AI loops take over, making robotics a better ten-year investment theme. China now has more than 150 humanoid robotics companies, and Andreessen Horowitz issued a warning comparing China's robotics trajectory to its earlier dominance in solar and 5G.