The Open-Source Tipping Point Just Got a Real-World Test — And It's Passing

Let's be precise about what we're forecasting and what this evidence actually proves. The [open-source-frontier] target is 'open-source model matches closed frontier performance.' GLM 5.2 is not formally open-source in the Meta Llama sense — it's a Chinese frontier model with public API access. But the underlying dynamic it represents is exactly what our thesis predicts: the gap between closed Western frontier labs and accessible alternatives has collapsed faster than the competitive moat thesis assumed. When 45% of OpenRouter token volume routes to Chinese models, up from under 2% a year ago, that's enterprises making real procurement decisions with real cost consequences, not researchers running evals.

We weight this evidence heavily because it passes the test we care most about: it's direct evidence of deployment behavior, not just benchmark claims. The distinction between 'model achieves parity on eval X' and 'enterprises route 45% of production traffic to it' is enormous. The former is proximate evidence. The latter is close to direct. Companies don't route nearly half their production token volume to a new provider in one year unless the quality-to-cost ratio is genuinely competitive at their workloads. That's not hype — that's revealed preference.

The strongest counterargument to our 75% isn't about GLM 5.2 specifically. It's about what 'frontier' means when the frontier keeps moving. Anthropic's unreleased capabilities — our own forecasting notes reference a 'Mythos' variant representing a potential step-change — and OpenAI's GPT-5.6 Sol hitting 91.9% on Terminal-Bench 2.1 suggest the closed labs are still producing capability gains that published models haven't matched. Our forecast is about parity at a moment in time, and if the frontier moves fast enough, parity at last month's frontier doesn't count. This is the part of our model that requires ongoing calibration.

We're also tracking a secondary complication: the verification standard for 'parity' matters enormously, and the Chinese model ecosystem does not submit to MLPerf or equivalent independent audits at the same rate as Western labs. The 45% OpenRouter routing figure is real enterprise behavior, which is arguably more meaningful than any benchmark. But when we set the resolution criteria, 'matches frontier performance' needs to mean something technically verifiable, not just 'companies are buying it.' Our 75% holds — and arguably this evidence pushes toward 77-78% — but we're watching whether the next Anthropic or OpenAI capability release re-opens a gap that GLM 5.2 just appeared to close. What would drop us below 65%: a credible independent benchmark showing GLM 5.2 underperforms on reasoning tasks that matter for enterprise use cases, combined with a demonstrable new capability release from a closed lab that resets the baseline. What would push us above 85%: Meta's next Llama release matching GPT-5.6 Sol on Terminal-Bench 2.1 with independent verification.