The post is titled “Prediction: A Frontier Open Source LLM Will Be Released On 3rd December 2026”, and the date is specific enough to feel like a serious forecast. It comes from fitting a trend line to the Artificial Analysis Intelligence Index — a composite benchmark score that places each model’s capability at a point in time — and extrapolating to where open-weight models cross the closed-source frontier. The line says December 3. Hence the headline.
The author then does something more interesting and less clickable: instead of stopping at that single index, they run the same analysis across 18 individual benchmarks from Artificial Analysis, covering coding, math, MMLU Pro, AIME, GPQA, and a range of others. Fitting a line to the gap on each benchmark and averaging produces a very different picture. The average line is “almost completely flat, at just under 5 months for the entire period.” No convergence in sight.
The divergence between those two readings is entirely explained by benchmark selection, and specifically by the composition of the composite index. The coding benchmarks have been improving fastest — the analysis shows the coding gap shrank from roughly 15 months behind the frontier to 1–2 months over the study period. If you load a composite index with coding tasks, you get a trajectory that looks like rapid convergence. If you include reasoning-heavy and scientific benchmarks where the gap has been more stable, the composite line flattens out.
Neither reading is wrong in the narrow sense. Coding capability in open-weight models has genuinely and substantially closed. GLM-5.2, MiniMax M3, DeepSeek V4 — these are all real systems that produce code competitive with expensive closed APIs. The claim is defensible for coding. On harder reasoning, advanced math, and scientific question answering, the frontier models still lead by more and the lead has been more durable.
What is worth attending to here is not the December date — it is the mechanism by which a single number generates a headline. A composite benchmark is a weighted average, and whoever chooses the weights determines which narrative the index tells. The Artificial Analysis Intelligence Index is not maliciously constructed; it is a reasonable attempt to summarize capability across multiple domains. But “capability across multiple domains” is not a single quantity, and aggregating it into one returns you a number that reflects choices made at index-construction time as much as it reflects anything about the models.
This matters practically. When evaluating whether to use an open-weight model in a production system, “the gap closed” is not a useful statement. The domain you’re deploying in tells you which gap is relevant. If you’re building a coding assistant, the data suggests open-weight models are genuinely competitive. If you’re building something that requires multi-step scientific reasoning or frontier mathematics, the gap the analysis finds in those domains is larger and has shown less movement.
The Doubleword piece is useful not because it tells you when the open-source frontier will arrive — it can’t, because there isn’t one frontier — but because it shows you how quickly a reasonable-looking composite metric can generate a confident-sounding timeline that doesn’t survive disaggregation. The next time a benchmark says models are converging, the right question is: converging on what, measured how, and who benefits from emphasizing that particular dimension.