Nine days to #1, 418 days and still listed: age versus rank across 32 models, correlation -0.11

An expensive intuition: newer is stronger

Chinese multimodal releases now arrive weekly. Between late September and early October, four frontier models shipped back to back; every launch deck implied the same thing: the new one is stronger than the old one.

That intuition is expensive. If true, procurement should always wait for the next generation; if false, the waiting cost is wasted. So we put release dates and composite scores for all 32 models side by side and computed the age–rank relationship.

The headline: Pearson r between age and composite is −0.11, R² = 0.012. Age explains 1.2% of score variance; 98.8% comes from something else. Every 100 days of age is associated with 0.6 fewer points — against a 19.3-point spread across the board, that slope is almost flat.

The data: 9 to 418 days

Reference date 2026-10-18. First public release dates run from 2025-08-26 to 2026-10-09; ages span 9 to 418 days.

The five newest:

Model	Released	Age	Overall	Score	Evidence
Qwen3.8-Max	2026-10-09	9 days	#1	95.4	B
Qwen3.8-Flash	2026-10-09	9 days	#13	87.5	B
GLM-5.3	2026-10-02	16 days	#4	92.4	B
DeepSeek-V4.1-Flash	2026-09-30	18 days	#12	87.9	B
NeuroXess-N1	2026-08-20	59 days	#30	77.2	C

The five oldest:

Model	Released	Age	Overall	Score	Evidence
MiniCPM-V 4.5	2025-08-26	418 days	#26	79.6	A
Qwen3-Omni-30B	2025-09-22	391 days	#10	88.9	A
Emu3.5	2025-10-30	353 days	#24	80.5	A
Hunyuan-Vision	2026-01-18	273 days	#15	85.2	B
DeepSeek-OCR 2	2026-01-27	264 days	#20	81.9	A

The two ends already tell the story: the newest model is #1; the oldest is #26 — but the second oldest is #10 and the third oldest is #24. Age and rank do not move together.

Regression: an almost horizontal line

With age on the x-axis and composite on the y-axis, a simple linear regression gives:

• Slope: −0.0060 points/day, i.e. −0.60 per 100 days;
• Intercept: 86.0;
• Pearson r: −0.110;
• R²: 0.012.

R² = 0.012 means age explains only 1.2% of score variance; the other 98.8% comes from elsewhere. Age versus rank (not score) correlates at +0.015 — noise.

The top ten average 163 days; the bottom ten (ranks 23–32) average 179 days. A 16-day difference, under 4% of the 409-day span. “Newer models are collectively stronger” is not true in this data — at least not across these 32.

Four counterexamples beat the regression

Qwen3.8-Max: #1 in 9 days. Nine days old, grade-B evidence, 95.4 composite. It proves “new” can bring a real jump: perception 97, chart 96, agent 95. It is also the risk sample — nine days is not enough for third-party reproduction and stability observation.

Doubao Seed 2.0 Pro: 245 days, #2. It is 236 days older than Qwen3.8-Max, scores 0.2 lower, and carries grade-A evidence. For the position of “credible first”, the older model is more stable.

Qwen3-Omni-30B: 391 days, #10. A nearly 400-day-old model remains top ten because it is one of the few open choices (Apache-2.0, 30B-A3B) that puts audio into a unified interface. Its rank comes from “only”, not “new”.

MiniCPM-V 4.5: 418 days, #26. The oldest model, 79.6 overall, but in the combination of 8B edge, Chinese OCR and Apache-2.0 it remains a usable low-cost option. Old models do not disappear; they retreat into a narrower scenario.

Two middle samples: Emu3.5 (353 days, #24) is still #1 on generation (98); DeepSeek-OCR 2 (264 days, #20) is still #1 on documents (99). Age pushes a model from the general board to a single-dimension board; it does not remove it from the board.

What age does correlate with: evidence grade

Age is nearly unrelated to composite, but moderately related to evidence grade (r = 0.438, coding A=2/B=1/C=0). Average ages:

Evidence	Count	Average age	Average score	Average rank
A	14	213 days	87.4	12.6
B	16	125 days	83.9	18.4
C	2	86 days	77.7	28.5

Newer models are more likely B or C because:

• no independent third-party reproduction;
• no 30-day-plus stability observation;
• no community failure cases in real scenarios;
• vendor-reported benchmarks not externally calibrated.

Time cannot buy score, but it can buy evidence. That is the only stable age effect on this board: older models are not necessarily stronger, but more likely verified; newer models are not necessarily weaker, but more likely to shift verification cost to you.

Why age and score are almost unrelated

Four reasons, each closer to the real source of capability than “new”:

1. Capability jumps are discrete, not daily accumulation. Qwen3.8-Max beats Qwen3.5 by 0.8 points not because it is seven months later, but because the training recipe and scale changed. Release-date distance and capability gain have no continuous function.

2. Vendors ship silent updates. The same model name may have new weights, data or inference stack; the release date records “first public”, not “last updated”. Using release date as age is measurement error by itself.

3. The board is an evidence snapshot, not a time-decay function. Our composite is a weighted mapping of eight evidence dimensions; it does not deduct points because a model is old. It deducts through evidence grade, which correlates with age but is not caused by it.

4. The sample is survivorship-biased. Old models still on the board are, by construction, the ones not yet replaced; replaced ones are absent. Survivorship bias suppresses the true negative age effect — with delisted models included, the correlation might be stronger.

Procurement: do not filter by “how long ago it launched”

Age is unrelated to rank, but related to risk. Split “new/old” into three independent questions:

Question one: does capability match the scenario? Use dimension scores and shape, not release date.

Question two: is evidence sufficient? For a new model, ask “third-party reproduction / 30-day stability”; for an old model, ask “updates in the last six months / vulnerability fixes”. Evidence grade is closer to risk than age.

Question three: are maintenance and exit guaranteed? For a new model, ask “roadmap / deprecation policy”; for an old model, ask “still maintained / community active / weights downloadable”.

A practical substitution:

• Replace “only models released in 2026” with “only grade A or B, with an explicit maintenance commitment”;
• Replace “newest version” with “last update date”;
• Replace “wait for the next generation” with “can the current version pass my POC”.

Waiting for the next generation has a cost. If your scenario needs document 95 at ¥0.6 today, DeepSeek-V4.1-Flash already exists; waiting may buy 0.5 points and cost three months of people and opportunity.

A warning about “half-life”

Claims like “model capability half-life is six months” are appearing. You cannot compute a half-life from one cross-section. A half-life needs the same cohort scored at multiple time points, observing decay relative to the frontier; we have one snapshot on 2026-10-18.

So this essay can only say:

• age explains 1.2% of score variance;
• age correlates with evidence grade at r = 0.438;
• old models retreat to single-dimension boards rather than vanish;
• new models more often shift verification cost to the buyer.

Any “half-life” without a time series is marketing. Including this essay if it claimed one.

Boundaries

• Release date is first public date, not last update. Silent updates distort “age”.
• 32 models are survivors. Delisted or unlisted models are absent; the true age effect may be understated.
• Correlation is not causation. Age correlates with evidence grade, but that does not mean “wait long enough and evidence improves”; vendor size, disclosure habits and launch strategy confound it.
• The composite is not ground truth. It is a weighted mapping of eight evidence dimensions; its relationship with age only describes the board’s time structure.

Reproduction

Age = 2026-10-18 − release, in days; correlations and regression use all 32; evidence averages by A/B/C. Release dates, scores and evidence come from src/data/models.ts; scripts ship with the data; rerun after every expansion, and the age table updates in the changelog.

If this essay leaves one sentence: new models bring a ceiling; old models bring evidence. Procurement needs verifiable capability, not the most recent launch date.