A new analysis from SemiAnalysis, surfaced via Techmeme on August 22, 2026, puts a hard number on something builders have felt anecdotally for two years: the performance gap between the best open-weight models and the best closed ones is now cutting in half with every new model generation, rather than just narrowing at a steady pace.

The analysis tracks the gap across four distinct eras of model development — pretraining scaling, post-training and alignment, reasoning and test-time compute, and now agentic AI. In each transition, the time it takes an open release to match the prior era's closed-model frontier has roughly halved compared with the era before it. Where early open models lagged the closed frontier by well over a year, recent open releases have closed comparable gaps within a single product cycle.

For teams shipping AI products, that's the part worth internalizing. This isn't a claim that any specific open model beats any specific closed one today — it's a claim about the slope of the curve, and slopes are what long-term model strategy should be built around, not this month's leaderboard screenshot.

Four eras, four narrowing gaps

According to Techmeme, SemiAnalysis frames the convergence as a pattern that repeats across eras rather than a one-time correction. The first era was defined by raw pretraining scale, where compute and data volume alone separated leaders from the rest. The second was post-training — RLHF, instruction-tuning, and preference optimization — where closed labs pulled ahead again by refining behavior rather than raw capability. The third was reasoning and test-time compute, the "think longer, answer better" wave. The fourth, playing out now, is agentic AI: models that plan, call tools, and execute multi-step tasks with minimal supervision. Each time a new axis of competition opened up, closed labs took the early lead — and each time, open-weight releases needed roughly half as long as the previous cycle to catch up.

Why "catching up" doesn't mean "caught up"

The halving trend describes a rate of convergence, not a finish line. Closed labs keep moving the frontier itself, so a shrinking gap can coexist with the absolute distance never reaching zero. Agentic benchmarks are also younger and messier than the static QA and coding leaderboards used to measure earlier eras, so — in our estimation — some of the apparent speed-up may partly reflect looser, less mature evaluation standards for agentic tasks rather than open models matching closed ones by every measure. The direction of the trend looks real; the precision of "half" should be read as an order-of-magnitude signal, not a physical constant.

What this means for model-selection strategy

If the gap keeps compressing on this schedule, the practical question for builders shifts from "which model is best right now" to "how fast will today's gap disappear, and what does that do to my architecture." A few implications worth acting on:

• Avoid hard-coding a single closed provider's API into core product logic; a model-agnostic routing layer costs little upfront and saves a rewrite later.
• Re-evaluate vendor choice every product cycle, not annually — if gaps are halving every era, a yearly review already lags the market by half a cycle.
• For high-volume, latency-sensitive inference, model the economics of self-hosting an open-weight model now, even if it isn't the cheaper option yet — the crossover point arrives faster each cycle.
• Track agentic-task benchmarks (tool use, multi-step planning, long-horizon execution) specifically, since that's the era where the gap is currently being contested, not the older static leaderboards.

AiiN's takeaway

The headline number — a gap halving every era — matters less as a benchmark stat and more as a planning input. Teams that treat model choice as a one-time architectural decision are optimizing for a snapshot that SemiAnalysis's own data says will be outdated well before the next roadmap review. The safer bet is building for model portability now, so that the next halving works in your product's favor instead of forcing a migration under pressure.