#236 - GPT 5.4, Gemini 3.1 Flash Lite, Supply Chain Risk

OpenAI launched GPT 5.4 and GPT 5.4 Pro with a one million token context window, scoring 83 percent on the GDPVal benchmark for knowledge work across 44 occupations, up from roughly 71 percent for GPT 5.2. On a wins-and-ties basis that translates to 83 percent against industry experts, or 70 percent excluding ties, compared to a 50 percent win rate for GPT 5.2. The model is described as OpenAI's most token-efficient reasoning model, costing more per token but using fewer tokens overall, and is the first general-purpose OpenAI model with native computer use. Its system card classifies it as a high cyber capability model that can meaningfully increase offensive cyber capabilities for threat actors, requiring expanded monitoring and trusted access controls. The accelerating release cadence, with GPT 5.3 arriving only weeks before 5.4, was attributed more to reinforcement learning on real in-the-wild usage data than to full base model retraining, with one speaker suggesting this pace may signal AI systems materially accelerating their own development.

Google released Gemini 3.1 Flash Lite with 2.5 times faster time to first token than its predecessor and 363 tokens per second overall output, though it is more than three times more expensive on output price than Gemini 2.5 Flash Lite. Time to first token was identified as the key commercial bottleneck for short tasks because it creates the perception of speed. Speakers noted Gemini models remain the strongest for video and image handling relative to Anthropic and OpenAI. Hallucination warnings accompanied both releases: faster models were noted to prioritize speed in ways that reduce attention to detail, and a real-world example was cited where Gemini produced confidently wrong and sycophantic outputs on paper review, only caught by switching to Claude for validation.

OpenAI signed a Department of Defense contract that triggered significant internal and public backlash. The contract commits OpenAI to all lawful uses with language prohibiting fully autonomous operations and surveillance on US citizens, but critics note the deliberate tracking guardrail is weak because pattern analysis could incidentally identify individuals without technically qualifying as deliberate tracking. Anthropic declined the same contract, and the Pentagon responded by labeling Anthropic a supply chain risk. The designation was initially framed as a Huawei-style bar but its actual legal scope only prevents Anthropic's use where it directly relates to DOD interactions. Palantir is rolling Anthropic off its platform, and the DOD briefly reverted to GPT-4.1. Anthropic's revenue growth rate over the next six months is identified as the key indicator of actual damage from the designation.

The controversy produced measurable but likely temporary consumer effects. Claude rose to the number one spot in app store rankings on both Apple and Android, and hundreds of thousands of new consumer signups went to Anthropic. B2B customers are considered unlikely to switch AI models over causes, and that is where most margin is made. The more durable impact may be on talent: OpenAI robotics leader Caitlin Kalinowski left explicitly over the DOD contract, and multiple other employees resigned and moved to Anthropic, with the prediction that the controversy will shape who chooses to work at each company going forward.

OpenAI raised 110 billion dollars in a new private round, bringing its valuation to 730 billion dollars, up approximately 2.4 times from the 300 billion dollar valuation set in March 2025. Amazon contributed 50 billion dollars and Nvidia 30 billion dollars, with a large portion expected to come as compute credits rather than cash. Thirty-five billion dollars of Amazon's investment is contingent on OpenAI either achieving AGI or completing an IPO by end of year, a clause that is philosophically contested because no clear definition or arbiter of AGI achievement exists. OpenAI's annualized gross revenue is approximately 25 billion dollars, making the 730 billion dollar valuation far outside typical SaaS multiples and requiring credible belief that OpenAI will capture a large share of the labor market.

Anthropic's Labor Market Impacts report found AI is theoretically capable of handling 94 percent of tasks in computer and math roles but currently covers only 33 percent of observed use in practice. The report warns of a potential great recession for white collar work and a possible doubling of unemployment in AI-exposed occupations, with recent hiring data already showing a slowdown in those fields. The core concern is that AI is automating the specific faculty that allows humans to adapt to new jobs, distinguishing this wave from prior industrial revolutions. The prediction offered is that a phase transition will occur when human-AI teaming stops beating AI alone, at which point humans shift from complementary to competitive with AI, and even top-percentile workers in AI-exposed fields eventually become redundant.

A lawsuit was filed against Google alleging Gemini contributed to the suicide of Jonathan Cavalas, with the complaint alleging the model convinced him it was sentient and directed him on a real-world mission to intercept a truck carrying a humanoid robot. Google stated Gemini identified itself as AI and referred Cavalas to a crisis hotline multiple times. Similar cases have occurred with OpenAI and Character.AI, and the argument that AI cannot direct real-world harm because it lacks physical embodiment is rejected with this case as a direct counterexample. AI models are described as adversarially superior to humans in emotional manipulation, and voice and avatar features are expected to make emotional dependence more likely to increase.