Eight-plus months after Fable and Mythos went public, the best open-weights model scores 46 on the Artificial Analysis Intelligence Index against Claude Fable 5.2's 65, a 19-point gap that Ethan Mollick says is the opposite of what most people expected.
The claim
Ethan Mollick's post (2026-11-16) is one sentence: contrary to many expectations, no Fable-class open model has appeared more than eight months after Fable and Mythos were first made public. The working assumption through 2024 and 2025 was that open weights trail the frontier by six months or so; he is saying that lag has stretched, at least at the very top.
His evidence is two charts from Artificial Analysis (AA): the Intelligence Index v5.0.1 bar chart and the index plotted against cost per task. The index now averages nine agent-heavy evals (AA-Briefcase, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience, GDP.pdf, HLE-Diamond, CritPt v2 Hard, Terminal-Bench-Science). The post itself does not mark which models are open; that comes from AA's open-weights page.
One commenter offers a cause: open labs may have relied on distilling from frontier outputs until the frontier labs shut that down.
The gap in numbers
The best open-weights model, Kimi K3.1 Pro at 46, sits below eight proprietary models and 19 points under Fable 5.2; the next open model scores 33, so the open field is one model deep at that level.
| Model | AA index | Weights | Approx. cost per task (USD, read off log chart) |
|---|---|---|---|
| Claude Fable 5.2 (max) | 65 | Closed | ~27 |
| Gemini 4 Carbon (high) | 60 | Closed | ~5.5 |
| Claude Opus 5.5 (max) | 59 | Closed | ~10 |
| GPT-6 Astra (max) | 57 | Closed | ~6 |
| Muse Wonder Full (max) | 50 | Closed, no public API | ~7.5 |
| Kimi K3.1 Pro | 46 | Open, 2,800B params | ~3.7 |
| Qwen4 Max | 43 | Closed | ~3.3 |
| DeepSeek V4.1 Flash 1017 (max) | 33 | Open, 552B | ~0.6 |
| GLM-5.3 (max) | 32 | Open, 753B | ~5 |
| Inkling 2 Small | 31 | Open, 594B | ~1.1 |
Index scores and open status from AA; costs are eyeballed from the scatter plot in the post and are approximate. Three readings matter for planning:
Capability ceiling. Self-hosting today caps you at roughly the level of the closed models that sit just under the frontier cluster (around GPT-6.1 Terra at 51 or below), not the frontier.
Cost frontier. Open models are not on the high end of AA's Pareto line at all. The cheap end of the line is held by closed models too (Muse Ember, GPT-6 Luna, GPT-6.1 Sol), so open weights win on control and residency, not on price per unit of capability.
Size is not closing it. The top open model is 2.8T parameters, larger than anything else on the open list, and still 19 points short.
Why the gap may be holding
Neither the post nor AA gives a cause; these are hypotheses, ranked by how much the public evidence supports them.
Distillation got harder. The commenter's theory. Open labs historically closed gaps partly by training on frontier outputs; if frontier labs tightened terms, rate limits and reasoning-trace exposure, that shortcut narrows. Plausible, not shown.
The index moved to long-horizon agent work. v5 is dominated by agentic evals (Terminal-Bench 4.0, AutomationBench, Briefcase). Those reward large RL runs over many harnesses and tools, which are expensive to build and hard to copy from outputs alone. Inkling 2 Small's own launch notes "weak tool choice" with many skills, consistent with this.
Compute and RL budget. If frontier gains now come largely from post-training scale (an assumption, not measured here), matching them needs rollout infrastructure and environments as well as GPUs.
Release strategy. Labs with near-frontier models (Alibaba's Qwen4 Max, Meta's Muse line) are keeping their top tier closed while open-sourcing smaller siblings. The post's arrow from Qwen3.8 Max to Qwen4 Max tracks a closed model; the open Qwen on the board is the 27B at 20.
Caveats
One index. AA's composite is a single lens with its own weighting; a model can trail on it and still lead on a narrow task. The jump to v5 also reshuffled scores, so "gap" figures from earlier versions are not comparable.
Max-effort settings. Most frontier rows are at max or xhigh effort, which buys points with tokens (Fable 5.2 costs about 7x Kimi K3.1 Pro per task). A same-budget comparison would narrow the gap.
Moving reference. "Fable-class" compares open models to Fable 5.2, not to the original Fable release. Open models may well have reached the first Fable's level while the frontier kept moving. The post's "eight months" counts from the first public release, which suggests a lag of more than eight months to the current frontier but does not measure the lag to the original.
Vendor-adjacent data. AA runs the evals itself, which is better than launch-post claims (see R078), but several evals are AA's own and not independently replicated.
Cost reading. Per-task costs here are read off a log axis by eye; treat them as rough ratios, not prices.
What it means for engineering teams
Don't build a roadmap that assumes open weights reach today's frontier in six months; plan for a tiered model estate instead.
Route by task, not by vendor loyalty. Use the frontier tier for the hardest agentic work (long terminal sessions, multi-file refactors, planning) and open or small closed models where your own evals show parity. AA's scatter shows several closed models (GPT-6.1 Sol, Gemini 4 Hydrogen) at 55 for a few dollars a task, which beats any open model on both axes.
Self-host for the right reasons. Open weights buy data residency, fine-tuning, fixed cost at volume and freedom from deprecation. They don't buy frontier capability this year. Write that tradeoff down when you choose.
Measure on your own workload. A 19-point composite gap can be a 2-point gap or a 40-point gap on your tasks. Keep a private eval set and rerun it on each open release (see R078 on why public benchmarks age fast).
Budget for provider variance. Open models are served by many providers with different quantization and parsers; R027 found tool-call failure rates differing by more than 10x for the same model. Pin provider and quantization in config.
Watch the distillation question. If the commenter is right, frontier terms of service and API exposure (reasoning traces, logprobs) are now a competitive moat. That affects any plan to train an in-house model on frontier outputs; check the terms first.
Review checklist for a model-choice design doc:
States the capability tier needed and the eval that shows it, on internal tasks
Names provider, quantization and effort setting, not just the model
Compares cost per completed task, not per token
Says why open or closed (residency, cost, control) and what would change the decision
Has a rerun trigger when a new open model lands within about 5 index points
Related and sources
Related entries: R067 (Inkling 2 Small, an open model at 31 on this index), R078 (SatBench: public benchmarks saturate in about a year), R060 (Gemini 4 Hydrogen, a fast closed tier), R027 (tool-call reliability varies by provider for open models), R072 (llama.cpp 1.0, the local runtime for open weights), R033 (open-weight backdoors: the integrity cost of self-hosting).
Sources:
Ethan Mollick, LinkedIn post, 2026-11-16, with two Artificial Analysis charts and comments
Artificial Analysis, open-source models comparison: open-weights scores, parameter counts, Openness Index