The Frontier Compression Event: Why We Moved Open-Source Parity to 75% — and What the 60-Day Release Cycle Actually Tells Us

First, the reasoning chain behind our move from 72% to 75%. The primary driver wasn't any single technical announcement — it was the Chinese open-weight model data. Q2 2026 OpenRouter traffic analysis showing Chinese providers at 45% of traffic, with Xiaomi's MiMo-V2-Pro at 21.1% versus OpenAI's 7.5%, is significant circumstantial evidence that open-weight models have crossed a deployment-preference threshold for a substantial portion of real users making real routing decisions. This isn't benchmark theater — it's revealed preference at scale. Users and developers are choosing open-weight Chinese models for actual workloads, which suggests functional performance parity in the dimensions those workloads require, even if not frontier-benchmark parity. We also weighted the Compute cost trajectory: 100x cost reduction over two years makes the economics of open-source training increasingly accessible. These factors together pushed us three points.

Now for what this week's releases actually tell us about the forecast — and this requires careful evidence classification. GPT-5.6 Sol achieving 91.9% on Terminal-Bench 2.1 for command-line coding, Grok 4.5 claiming to match Claude Opus 4.8 on a 1.5 trillion-parameter V9 foundation model, and Gemini 3.5 Pro introducing a 2 million-token context window after a complete architectural rebuild: these are proximate evidence that the closed frontier is advancing rapidly. That's relevant to our forecast only insofar as it tells us what open-source needs to match. The 60-day closed-model release cadence that OpenAI has established creates a moving target problem that is genuinely the most challenging element of our thesis. When frontier capability is advancing this fast, 'parity' becomes a time-stamped concept — open-source may achieve parity with GPT-5.4 exactly as GPT-5.6 ships.

This is where we need to be explicit about our forecast definition, because it matters enormously for resolution. Our forecast is 'open-source model matches closed frontier performance' — we define this as a point-in-time benchmark parity with the leading closed model, not sustained parity across a moving frontier. That means the forecast can resolve YES even if open-source immediately falls behind again when the next closed model ships. Under that definition, the 60-day compression cycle is actually mixed evidence: it advances the frontier faster (making parity harder to achieve) but it also compresses the window in which any closed model holds dominant position, increasing the probability that an open-source release coincides with a period between closed-model generations where the gap is narrow. Meta's Llama release cadence and the Chinese open-weight ecosystem's demonstrated velocity are what we're watching on the FOR side.

The counterargument we're genuinely wrestling with: the Gemini 3.5 Pro rebuild story is a signal that frontier architecture innovation is not slowing. Google scrapping the entire Gemini 2.5 Pro architecture rather than iterating suggests closed labs are making discontinuous advances, not just scaling runs. If architectural innovation is the primary driver of frontier performance — rather than compute and data, where open-source can compete — then benchmark convergence may be perpetually delayed by qualitative leaps that open-source training pipelines can't replicate quickly. Anthropic's leaked 'Mythos' remains in our AGAINST column for exactly this reason. We haven't downgraded from 75% because Meta's investment scale and the Chinese open-weight momentum are real, but if Gemini 3.5 Pro's Deep Think layer and GPT-5.6's subagent reasoning architecture prove to reflect a new architectural paradigm rather than iterative improvement, we'd revisit downward. Specific trigger: if by September 2026 no open-weight model achieves within 5 points of GPT-5.6 on Terminal-Bench 2.1, we'd move back toward 70%.