Frontier Pricing Just Broke Open. What GPT-5.6, Grok 4.5, and Claude Fable 5 Do to Our Open-Source Convergence Forecast.

Our 75% reflects two core drivers: Meta's sustained open-source investment and the 100x compute cost reduction we've tracked as verified. What moved us from 72% to 75% last cycle was evidence that post-training techniques were closing the gap faster than frontier labs could maintain their moats through data advantage alone. This week's news requires us to re-examine both the 'what' and the 'when' of our forecast.

Start with what supports the thesis. Google abandoning Gemini 3.5 pre-training weeks from launch — a decision costing hundreds of millions — because it wasn't competitive with GPT-5.6 Sol and Claude Fable 5 on math reasoning and SVG generation is strong evidence of frontier compression. When the third-place frontier lab has to do a ground-up rebuild to stay in the same competitive tier, the performance ceiling is rising faster than any single lab can track. That ceiling compression is exactly the dynamic open-source needs: it reduces the distance any individual model needs to close. The Chinese model market share data (30-46% of enterprise API usage, up from 2% a year ago) is separately significant — it confirms that 'good enough' performance at 60-90% cost savings is a viable enterprise decision criterion, which is exactly the market condition that makes open-source convergence commercially meaningful.

Now for what complicates the thesis, and this is the part we want to be honest about. The forecast's resolution criteria is 'open-source model matches closed frontier performance' — and this week, three separate frontier releases have moved the frontier. GPT-5.6 Sol at 91.9% Terminal-Bench 2.1 in Ultra Mode, Grok 4.5 claiming Opus-class performance, and the Claude Fable 5 pricing signal suggesting Anthropic believes it has materially differentiated capability worth double the cost. Our forecast doesn't yet account for how rapidly the target is moving. We've been tracking gap-closing, but if the frontier is accelerating, the gap-closing velocity needs to be faster than we've modeled to resolve YES within our timeframe.

The government restriction on GPT-5.6 Sol — held to 20 partners for two weeks over cybersecurity capability thresholds before broad release — deserves particular attention. This is the first instance we've seen of the US government restricting AI model deployment on capability grounds before release. That is a data point about capability levels that has no open-source equivalent and no benchmark equivalent. The 'leaked Mythos' concern we've been carrying in the AGAINST column just got more concrete: if frontier labs have capabilities that government agencies consider threshold-crossing, the open-source convergence question becomes which threshold we're measuring against. Our 75% doesn't yet fully account for this. We're holding the number pending the July 17 Gemini 3.5 launch and independent benchmarking of GPT-5.6 Sol's Ultra Mode performance. What would move us above 80%: a Meta Llama release within three months that posts within 5 points of GPT-5.6 Sol on Terminal-Bench 2.1 independently verified. What would drop us below 65%: evidence that the capability threshold triggering government restriction represents a qualitative capability gap — not a quantitative one — that post-training cannot close.