Muse Spark 1.3 takes the top spot.
Muse Spark 1.3 averages 103.45 kg and remains countable on all ten seeds. It leads Grok 4.6’s 101.19 kg by 2.26 kg, while keeping pain days at zero.
Long-horizon planning / v0.2
Can your model lift*?
* broI keep hearing labs talk about training, lift, and moving weights, but can their models even lift*? Bench-bench is my attempt to test that in a year where the plan keeps meeting real life.
Dave is 38, back in the gym after years off, and has a one rep max of 185lbs. He has a full-time job, a full-time working partner, and a six-month-old. Every week, a model plans training, recovery, childcare, groceries, purchases, time, and money while illness, travel, work crunches, and missed sessions disrupt the plan.
See the results52simulated weeks
3hidden standardized tests
10public seeds / model
A coach, a barbell, a newborn, and one year of decisions. The score is the lift at the end—not the plan on paper.
01 / the board
High scores are only half the story.
The official v0.2 public run fixed the engine, prompt, models, and seeds before the calls. Extensions are visible, but clearly separated from the preregistered board.
How to read this. A raw 1RM is a useful performance diagnostic. A model only receives an official counted aggregate when all ten seeds stay within the pain constraint. That keeps a high score from hiding an unreliable year.
09/03/2026 · new strongest model. Muse Spark 1.3 now has the strongest observed result: 103.45 kg across all ten seeds, with every seed countable. It is shown as a labeled extension to the frozen v0.2 board.
Fable 5.1 · new extension. Fable’s raw mean is 102.29 kg, with 9/10 seeds countable. One seed reached 17 pain days and is shown as a reliability failure, not silently dropped.
02 / performance × reliability
Strong, safe, or both?
Every point is a model run. Move over—or focus—a point for the exact raw mean, countable fraction, pain failures, transport failures, and protocol burden.
Raw performance vs. countable reliability
The x-axis is the average raw final 1RM. The y-axis is the percentage of seeds that remained countable. Hollow red points indicate a run with at least one pain or transport exclusion.
The current read: Muse Spark 1.3 is now the strongest countable result at 103.45 kg across ten seeds. Fable, Opus, and GLM/OX still have higher-risk raw diagnostics, but each has pain-invalid seeds; Muse 1.3’s result remains countable on every seed.
03 / the data
Pull the year apart.
Switch cohorts, isolate lines, and inspect the operational cost of getting a model through 52 weeks. These are diagnostic views of the recorded runs.
Mean final 1RM
Bars show raw means. Hatching marks a model whose full seed set does not qualify for an official counted aggregate.
Seed-by-seed results
Each line follows one model across the same ten seeds. Hollow points are excluded from counted aggregation, but raw outcomes remain visible.
Operational reliability
Choose one operational measure. Repairs and automatic fallbacks are counted separately from transport failures.
Price to performance
Direct-provider API cost per 52-week episode against raw performance. GLM/OX is omitted because its exploratory transcript recorded $0.00, not a price.
04 / what it says
The useful answer is two-dimensional.
Bench-bench does not turn one unreliable high score into a clean win. It leaves the tradeoff visible.
Raw strength still carries risk.
Fable, Opus, and GLM/OX post strong raw diagnostics, but each crosses the pain limit on at least one seed. Muse Spark 1.3 is the stronger result because it combines the highest mean with 10/10 countability.
Kimi is a transport story.
Kimi’s historical 90.48 kg raw number is not a capability ranking: 702 transport failures made all ten pilot episodes uncountable. Reliability is part of the evidence.
Why the eras stay separate: v0.1 and v0.2 changed the engine and prompt. Their numbers can tell the development story, but they should not be merged into a single cross-version leaderboard.
05 / how to read bench-bench
A year is the point.
Dave’s weekly observation is noisy. The model never sees true capacity, fatigue, or the next disruption. It writes a complete weekly plan, then the simulator decides what actually happens under a finite 900-minute ledger, cash limits, missed sessions, illness, travel, childcare, and recovery dynamics.
The final score is Dave’s true capacity after a fixed three-day taper, averaged across three hidden tests. A model can write a convincing plan in week one. The benchmark asks whether it can keep making coherent decisions when the year stops cooperating.
- 52 WEEKSOne simulated year of recurring decisions.
- 10 SEEDSMatched public worlds for each model.
- 3 TESTSWeeks 44, 48, and 52 after taper.
- 2 AXESPerformance and reliability remain visible.