Six Defections and $84B Later, the Open-Source Frontier Gap Just Narrowed Again

Let's be precise about what drove the move from 72% to 75%, because it wasn't these events — it was earlier data that we're now contextualizing with this week's news. The move reflected the cumulative weight of Meta's Llama 4 family reaching genuine competitive performance on standard benchmarks, the 100x compute cost reduction making open training runs tractable at near-frontier scale, and the narrowing gap in post-training techniques as the research community reverse-engineers closed models. What this week adds is a structural signal: if the researchers who built Gemini's reasoning stack, its Transformer architecture, and its code generation capabilities are now at Anthropic — a company whose safety research philosophy has historically complemented open research norms — the institutional knowledge advantage that closed frontier labs relied on is eroding faster than we modeled.

Here's the distinction that matters for our forecast target. 'Open-source model matches closed frontier performance' is not asking whether Anthropic beats Google — that's a closed-vs-closed question. It's asking whether an openly released model reaches the performance tier of whatever the best closed model is at resolution. Claude Opus 4.8 topping the index actually raises the bar our forecast needs to clear, because it pushes the frontier higher. But the talent migration story cuts differently: researchers who understand how to close capability gaps at the frontier are concentrating outside the companies most opposed to open release. Meta remains the primary open-source driver, and nothing this week changes Meta's incentive structure or resource commitment.

The part of our thesis that keeps us up at night is Anthropic's leaked 'Mythos' capability. If frontier labs have genuinely unpublished step-change improvements — not just incremental gains — then benchmark convergence on current public models may be measuring the wrong target. Our 75% explicitly does not account for unknown capabilities in development. The defection story is actually double-edged here: if six of DeepMind's best researchers are at Anthropic, and Anthropic has a step-change model in development, the closed frontier may be moving faster than open alternatives can track, not slower.

Alphabet's $84.75B equity raise is the counterweight to the defection narrative, and we're holding it in tension rather than resolving it. That capital commitment — $180-190B in 2026 capex, the largest equity raise in corporate history — reflects Sundar Pichai's stated belief that enterprise compute demand exceeds supply. Google is not conceding the frontier; it's betting that compute scale eventually overwhelms talent gaps. Our 75% reflects the view that open-source training efficiency improvements will outpace closed-lab compute scaling advantages by the resolution date — but if Alphabet's infrastructure investment produces a step-change in training throughput, that assumption weakens. What would drop us below 65%: verified evidence of a closed-lab capability at the level of GPT-3's original release — something that open models simply cannot match on any meaningful benchmark — emerging from a lab that doesn't open-release. What would push us above 80%: Meta releasing a Llama 5 variant that tops the Artificial Analysis Index on any primary benchmark within 60 days of the best closed model.