Chinese Labs Achieve Coding Parity in 12 Days. We've Moved Open-Source Frontier to 69%. Here's the Honest Case for Why We Didn't Go Higher.
TexTak moved 'open-source model matches closed frontier performance' from 67% to 69% this cycle. Today's Air Street Press report is the most direct evidence we've received since establishing this forecast: four Chinese AI labs released open-weights coding models within 12 days of each other — GLM-5.1, MiniMax M2.7, Moonshot's Kimi K2.6, and DeepSeek V4 — all achieving what the report characterizes as 'frontier capability on agentic engineering at less than one-third the cost of Claude Opus 4.7.' That's a remarkable coordination of capability release. But the NIST evaluation finding that DeepSeek V4 lags US frontier by roughly eight months on aggregate benchmarks is the sentence in that same report that keeps our move modest.
First, the definitional work our forecast requires. When we say 'open-source model matches closed frontier performance,' we are forecasting benchmark parity, not product parity, not developer preference parity, and not commercial impact parity. These are different things, and our forecast target needs to be read precisely: a general-purpose open-weights model achieving performance indistinguishable from the closed frontier on standardized capability benchmarks. Today's news gets us closer but doesn't cross that line. 'Frontier capability on agentic engineering tasks' in a specific domain is meaningful progress. An eight-month aggregate lag on NIST evaluation is not parity — it's a measurable gap that NIST, not a startup press release, is quantifying.
What today's evidence does prove strongly: the cost trajectory is now empirically established, not speculative. Less than one-third the cost of Claude Opus 4.7 at comparable coding task performance is a significant data point. It validates the compute cost compression thesis that anchors a key piece of our FOR case. And the coordination — four separate labs, four separate architectures, all landing within 12 days — suggests this isn't a lucky one-off. Chinese AI development has reached an organizational maturity where multiple independent teams are converging on the same capability frontier simultaneously. That's a structural signal, not a product signal.
The counterevidence we added to our AGAINST this cycle — Anthropic's reportedly leaked 'Mythos' model representing a step-change improvement — is the most important thing we're watching and the most honest gap in our model. If frontier labs have unreleased capabilities that are qualitatively ahead of what's been published, the benchmark comparison we're tracking may be measuring the wrong frontier. An eight-month lag against Claude Opus 4.7 is very different from an eight-month lag against a model not yet publicly released. Our 69% doesn't yet account for the possibility that the closed frontier is materially further ahead than public benchmarks suggest — because by definition, we can't measure what hasn't been published.
We moved to 69% rather than something higher for two reasons. First, the eight-month NIST aggregate gap is real measurement from an independent source — we weight independent benchmarking heavily precisely because self-reported performance claims are unreliable in this space. Second, our forecast has a specific resolution criterion: parity on standardized capability benchmarks. Coding task performance, while important, is one domain. The move from 67% to 69% reflects genuine progress on the cost-parity dimension and the acceleration of Chinese open-weights development, but not yet a claim that the aggregate gap has closed. What would move us above 80%: independent NIST or HELM evaluation showing an open-weights model within margin of error of GPT-5 or Claude's current production flagship on aggregate benchmarks. What would push us back toward 60%: confirmation that Anthropic's Mythos or an equivalent unreleased model is genuinely several generations ahead, making current parity comparisons against published models a misleading benchmark.