Coding models, five benchmarks compared

17 October 2026. Five sources side by side, a capability ranking, and what $100 of API spend actually buys from each model. Coverage decides which models can be ranked at all, so every chart states how many sources measured the model.

Latest bots — new on the agents board

Three always-on agents landed in the last four weeks; all three now sit on the agents directory and the news feed:

Also new on the model board: Gemini 4 Carbon — an Artificial Analysis v5 score of 59, the highest independent figure on record; no DeepSWE run yet.

Capability, five sources combined

Left: models covered by three or more of the five sources, so a composite can be computed. Right: the rest — one or two sources only — shown at their raw DeepSWE score. Two measures on one 0–100 scale, so the panels are divided and styled apart; the amber qualifying bar applies to the left panel only.

Composite — 3 to 5 sources 25 models

Solid bars: composite percentile. Amber line = qualifying bar, 53.2.

  • Claude Opus 5.593.3
  • GPT-6 Astra87.9
  • GPT-6.1 Sol85.2
  • Claude Sonnet 5.583.3
  • Claude Fable 572.5
  • Gemini 4 Argon70.0
  • GPT-5.6 Sol67.0
  • Muse Spark 1.365.6
  • Claude Opus 565.5
  • GPT-6 Sol60.8
  • Grok 4.758.7
  • GPT-5.6 Terra57.3
  • GLM-5.353.2
  • Gemini 3.8 Flash47.3
  • Grok 4.644.1
  • Kimi K343.5
  • GLM-5.3-Flash43.2
  • GPT-6 Luna40.7
  • GPT-5.6 Luna37.9
  • Claude Opus 4.836.2
  • Grok 4.533.4
  • Muse Spark 1.228.1
  • Gemini 3.7 Flash27.5
  • Claude Sonnet 523.8
  • Qwen3.8 Max22.2

Too new to composite — 1 to 2 sources 19 models

Hollow bars: raw DeepSWE score, not a composite. These are not ranked against the panel on the left.

  • Muse Wonder Full79.0
  • Muse Wonder78.8
  • Ember-175.2
  • DeepSeek V4.1 Flash‡74.2
  • MiMo-V2.6-Pro71.9
  • MiMo-V2.6-Flash67.9
  • GPT-5.567.0
  • Mistral Large 461.7
  • Qwen3.8 Flash / Flash-Next58.7
  • Muse Spark 1.153.0
  • GPT-5.452.0
  • Gemini 3.6 Flash49.0
  • GLM-5.244.0
  • Qwen3.8-27B42.2
  • Laguna S 2.140.4
  • Gemini 3.5 Flash37.0
  • Kimi K2.7 Code31.0
  • Claude Sonnet 4.630.0
  • Gemini 3.1 Pro12.0

Cost efficiency — what $100 actually buys

Quality-weighted tokens = tokens × (composite ÷ 53) × coverage factor. The composite is the five-source percentile from the chart above; 53 is its median, so a model sitting at the median scores a coefficient of exactly 1.00. Coverage factor: 1 source ×0.50, 2 ×0.75, 3 ×1.00, 4 ×1.25, 5 ×1.50 — a stated prior, not a measurement: a figure reproduced by more independent runs is worth more, and a figure one lab reports is worth less. ‡ Targeted-training penalty: a headline number that survives only in the maker's own harness is halved before any of this is computed. DeepSeek V4.1 Flash is the first case: 74.2 under mini-SWE-agent but 65.6 under Codex, and 30.0 on Terminal-Bench 3.0 where Opus 5 posts 43.3 — a score tuned to the task shape the benchmark rewards. The dagger marks the model, the coefficient shows the halved composite, and the raw table number stays visible everywhere else. Columns: weighted tokens, raw tokens before any adjustment, the coefficient, and how many sources cover the model. Hollow bar = factor below neutral, so a row held down by thin evidence reads differently from one held down by price.

Models sold only as API — DeepSeek, GLM, Qwen, MiMo — have no subscription tier at all, so $100 of API spend is the only denominator that covers every vendor on the same terms.

Coverage reorders the board. Laguna S 2.1 buys the most tokens of anything here at 800M, but its composite is 9.1 and only one source measures it, so the factor cuts it to 69M and 8th place. The multiplier works the other way too: Claude Opus 5.5 is covered by all five sources and has the highest composite on the board, so its 12.5M raw tokens count as 33M.

  • GPT-6 Luna384M500M0.773/5
  • GLM-5.3-Flash344M422M0.823/5
  • MiMo-V2.6-Flash306M571M0.541/5
  • Qwen3.8 Flash280M435M0.642/5
  • DeepSeek V4.1 Flash‡211M382M0.552/5
  • MiMo-V2.6-Pro168M184M0.912/5
  • Gemini 3.8 Flash74M67M1.124/5
  • Laguna S 2.169M800M0.091/5
  • GLM-5.366M53M1.254/5
  • Muse Spark 1.362M50M1.243/5
  • Claude Sonnet 5.559M25M2.365/5
  • GPT-6.1 Sol50M25M2.014/5
  • Grok 4.746M33M1.384/5
  • Gemini 3.7 Flash43M67M0.654/5
  • Qwen3.8-27B42M95M0.442/5
  • Mistral Large 442M97M0.432/5
  • GPT-6 Sol36M25M1.434/5
  • Grok 4.635M33M1.044/5
  • Claude Opus 5.533M12M2.645/5
  • GPT-5.6 Luna32M44M0.723/5
  • Muse Spark 1.227M50M0.533/5
  • Qwen3.8 Max25M40M0.635/5
  • Grok 4.521M33M0.633/5
  • GPT-5.6 Terra19M18M1.083/5
  • Claude Opus 519M10M1.855/5
  • Muse Wonder15M12M1.232/5
  • Ember-114M17M0.861/5
  • GPT-5.6 Sol14M9M1.584/5
  • Kimi K314M17M0.823/5
  • GPT-6 Astra12M5M2.495/5
  • Claude Sonnet 511M25M0.453/5
  • Muse Spark 1.110M50M0.191/5
  • GLM-5.29M68M0.131/5
  • Claude Fable 59M5M1.714/5
  • GPT-5.57M9M0.792/5
  • Claude Opus 4.87M10M0.683/5
  • Kimi K2.7 Code3M74M0.041/5
  • Gemini 3.1 Pro3M22M0.122/5
  • Claude Sonnet 4.62M17M0.132/5
  • GPT-5.42M11M0.171/5

4 models have a DeepSWE score but no public price, so they cannot be priced: Muse Wonder Full, Gemini 4 Argon, Gemini 3.6 Flash, Gemini 3.5 Flash. Qwen3.8-Flash-Next is not on this list because it is the same measurement as Qwen3.8 Flash — the weights release and the production API name of one model, double-listed upstream and counted once here. That includes the DeepSWE leader Muse Wonder Full, which Meta sells only inside Muse Code.

Tokens = $100 ÷ blended API list price per 1M tokens, at a 3:1 input:output ratio. Nothing on this tab is adjusted for capability — it answers only how much text $100 buys, which is the input the weighted tab needs. Sorted by tokens. The second column is the blended price, the last is how many of the five sources cover the model.

The same $100 buys 800M from Laguna S 2.1 and only 5M from Claude Fable 5 — a 160× spread. That gap is the whole reason the weighted tab reorders anything: a model can be far cheaper per token and still lose once capability is priced in.

  • Laguna S 2.1800M$0.1251/5
  • MiMo-V2.6-Flash571M$0.1751/5
  • GPT-6 Luna500M$0.2003/5
  • Qwen3.8 Flash435M$0.2302/5
  • GLM-5.3-Flash422M$0.2373/5
  • DeepSeek V4.1 Flash382M$0.2622/5
  • MiMo-V2.6-Pro184M$0.5442/5
  • DeepSeek V4 Pro 0813184M$0.5440/5
  • Mistral Large 497M$1.0322/5
  • Qwen3.8-27B95M$1.0502/5
  • Kimi K2.7 Code74M$1.3601/5
  • GLM-5.268M$1.4621/5
  • Gemini 3.8 Flash67M$1.5004/5
  • Gemini 3.7 Flash67M$1.5004/5
  • GLM-5.353M$1.9004/5
  • Muse Spark 1.350M$2.0003/5
  • Muse Spark 1.250M$2.0003/5
  • Muse Spark 1.150M$2.0001/5
  • GPT-5.6 Luna44M$2.2503/5
  • Qwen3.8 Max40M$2.4755/5
  • Grok 4.733M$3.0004/5
  • Grok 4.633M$3.0004/5
  • Grok 4.533M$3.0003/5
  • GPT-6.1 Sol25M$4.0004/5
  • Claude Sonnet 5.525M$4.0005/5
  • GPT-6 Sol25M$4.0004/5
  • Claude Sonnet 525M$4.0003/5
  • Gemini 3.1 Pro22M$4.5002/5
  • GPT-5.6 Terra18M$5.6253/5
  • Ember-117M$6.0001/5
  • Kimi K317M$6.0003/5
  • Claude Sonnet 4.617M$6.0002/5
  • Muse Wonder12M$8.0002/5
  • Claude Opus 5.512M$8.0005/5
  • GPT-5.411M$8.7501/5
  • Claude Opus 510M$10.0005/5
  • Claude Opus 4.810M$10.0003/5
  • GPT-5.59M$10.7502/5
  • GPT-5.6 Sol9M$11.2504/5
  • GPT-6 Astra5M$20.0005/5
  • Claude Fable 55M$20.0004/5

5 (Muse Wonder Full, Gemini 4 Argon, Qwen3.8-Flash-Next, Gemini 3.6 Flash, Gemini 3.5 Flash) of the board publish no list price and cannot appear here: Muse Wonder Full, Gemini 4 Argon, Qwen3.8-Flash-Next, Gemini 3.6 Flash, Gemini 3.5 Flash.

Only four subscription tiers have a measured quota, which is why this pane is short. The bars are effective dollars per benchmark-sized coding task; the right-hand column names the model the plan's quota was spent on. Monthly figures use 4.33 weeks, not four — a month averages 4.33 weeks, and rounding to four understates every plan by 8%.

Where these quota figures come from. No vendor publishes a quota-to-token conversion rate — that is why the site's own membership table can price only four of its nine products. So a subscription cannot be priced from published data at all; it has to be measured. The figures below are a community measurement, not a vendor disclosure and not our own test: finopsllm.com logged their real weekly usage from 11 September to 15 October 2026 and valued it at API list prices. Sample: one team, five tiers. Read them as an order of magnitude, not an invoice.

The weeks are full caps, not partial ones. The source runs Claude Max to 100% of the weekly limit every week, so no uplift is applied — the $658–$873 range already is the whole cap. Anthropic's “50%” figure is a different thing: on Max, Fable may consume at most 50% of your weekly limits, which caps how much Fable work a week can hold, not how full the week is. Codex windows that ended at 98–99% were scaled to 100% by the source.

The Pro $200 tier changes on 30 October. Until 29 October it is 20× Plus, about 5,154 tasks a week, which makes it the cheapest per task on this chart at $0.0090. From 30 October it drops to 10×, about 1,377 tasks, and lands at $0.0335 — level with Pro $100 and 3.7× worse than it is today. Both rows are in the table; only the current one is plotted.

  • ChatGPT Pro $200 (20x, to 29 Oct)$0.0090GPT-6.1 Sol
  • ChatGPT Plus$0.0179GPT-6.1 Sol
  • Claude Max 5x$0.0264Claude Opus 5.5
  • ChatGPT Pro $100$0.0336GPT-6.1 Sol
PlanPriceAPI value / wkTasks / wk $ / taskAPI $ per $1FlagshipStatus Quota measured by
ChatGPT Plus$20$67258$0.017914.5GPT-6.1 Solcurrentcommunity log, finopsllm.com
Claude Max 5x$100$761875$0.026433.0Claude Opus 5.5currentcommunity log, finopsllm.com
ChatGPT Pro $100$100$179688$0.03367.8GPT-6.1 Solcurrentcommunity log, finopsllm.com
ChatGPT Pro $200$200$13405154$0.009029.0GPT-6.1 Sol20x, until 29 Octcommunity log, finopsllm.com
ChatGPT Pro $200$200$3581377$0.03357.8GPT-6.1 Sol10x, from 30 Octcommunity log, finopsllm.com

Qwen3.8-Flash-Next is 125B of weights with 6B active per token, so it computes like a small model and loads like a large one. Strata keeps the weights in system RAM and treats VRAM as a cache for hot experts. That inverts the usual advice: VRAM decides how fast it runs, system RAM decides whether it runs at all.

GPU VRAMSystem RAMQuantDecode ContextWhat it actually means
8 GB16 GBnone—— Not viable. “8 GB runs 125B” is a video title; the thread that measured it settles on 12 GB + 64 GB as the floor. A 35B-A3B at Q4 is the honest pick for this card.
10 GB64 GBnone—— No measured report exists at this tier. Treat it as 12 GB with less headroom for the expert cache, which is what runs out first at long context.
12 GB64 GBQ2_093 tok/s32K The official reference machine (RTX 5070). This is the headline number, and it is the most aggressively compressed tier — not what you want to code with.
12 GB64 GBIQ3_S53 tok/s32K The quality tier: 93.26 against 93.12 for BF16 on the vendor's own three-part eval, at under a quarter of the size.
12 GB96 GBIQ3_S~1,400 prefill~60K This site's own machine. One request at a time; the client's 60-second connect deadline caps practical prompts near 60K tokens well before the model's 80K.
16 GB64 GBIQ3_S40–60 tok/s256–262K Where it gets comfortable. Two independent reports: a 5060 Ti at 40+ and a 5080 laptop at 45–60 with the full context window. Long context eats the VRAM that was caching experts — the 5080 was down to about 4 GB for that.
48 GB (2 cards)128 GBIQ3_S105.8 tok/s— Dual GPU is where speed comes from. The weights are 83.6 GB, so this is still mostly a memory-bandwidth story, not a compute one.
32 GB32 GBanyOOM— A 5090D with only 32 GB of system RAM ran out of memory. VRAM without RAM still fails.

Two things the marketing rounds off. The 93 tok/s figure is the Q2_0 tier; the quality tier runs at 53. The tier below it (IQ3_XXS) is reported to loop and over-think, so it is not a shortcut to lower VRAM.

Where it actually falls short. Three limits matter more than the speed numbers, because they decide whether this replaces a subscription at all.

  • One task at a time. Concurrency is 1 — a second request queues behind the first, and the weights are read from system RAM on every token, so there is no bandwidth left for a second stream. It cannot serve a team, a CI job, or an agent that fans out into sub-tasks.
  • Small working context. About 60K tokens is practical on 12 GB: the client's 60-second connect deadline binds well before the model's own limit. No whole-repo analysis, no long document passes, no multi-file refactors.
  • VRAM is the wall, and 12 GB sits below it. On 12 GB the weights are re-read from system RAM for every token, which is all the bandwidth one stream can absorb. 48 GB changes the picture: a measured RTX 4090 48G run held a 28 GB KV-cache pool and served 32 concurrent streams at 32K context — 150 tok/s aggregate, 85–100 tok/s for a single stream, first token under 0.7 s. A 96 GB card holds the weights outright. 24 GB is the awkward middle: at 32K context and 16 streams the KV cache alone passes 20 GB, so concurrency has to drop and decode falls back to 70–80 tok/s. So yes, 48–96 GB of VRAM does break the limit — it just costs more than the API it replaces.

Power only looks like a wash at Chinese rates. At 53 tok/s and a ~400 W whole-system draw, one million generated tokens takes 5.2 hours and burns about 2.1 kWh. At ¥0.6/kWh that is ¥1.26 per million tokens, roughly level with Qwen3.8-Flash's blended API price of ¥1.28. In Australia it is not level. At AU$0.35/kWh — a mid-range residential rate — the same 2.1 kWh costs about AU$0.74, roughly ¥3.4, which is two to three times the API price before any hardware is counted. A 48 GB card draws more than a 12 GB one, so the gap widens further.

Add the card and the picture inverts. A 12 GB machine is roughly ¥12,000; over three years that is ¥11 a day. Run it at a million tokens a day — five hours of continuous generation, which nobody actually does — and hardware alone adds ¥11 per million tokens, about nine times the API price. The honest break-even is not a token count. It is a constraint: data that cannot leave the machine, an API that is rate-limited, or a need to work offline.

Put together, the commercial verdict holds even after the VRAM path is priced in: 12 GB serves one request at a time and cannot see a whole repository; 48 GB fixes both, but the card alone costs more than years of API calls and, in Australia, the power bill on top of it already exceeds the API price. It is a tool for a specific constraint — data that cannot leave the machine, an API that is rate-limited, offline work — not a cheaper way to do the same job.

Can it reach 1M context? No — and the 8K you may have seen is a config default, not a model limit. The deployment tutorials ship with context: 4096 or max_model_len: 8192 because those settings fit any card. The model itself is rated to 128K, and the community has run 262K on two 24 GB cards with an int8 KV cache, passing needle-in-a-haystack at all depths.

1M is not a setting anyone forgot to change. It would need roughly four times the KV cache of that 262K run, and decode already drops from 105–160 tok/s at short context to 60–73 tok/s at 16–30K as the cache grows. Extrapolate that curve and a 1M run lands in single-digit tokens per second on hardware that does not exist in consumer form. The ceiling is architectural, not configuration.

Sources: the Strata reference machine figures, the vendor's GSQ-RCO eval, measured prefill/decode curves across 1K–30K context, and community reports collected 1–8 October 2026. The 12 GB / 96 GB row is measured here.

How to read this page. Only 25 of the 44 distinct models can carry a composite, because a model has to appear in three or more of the five sources first. The DeepSWE leader Muse Wonder Full appears in exactly one, so it sits in the right-hand panel at its DeepSWE score. Blending sources does not make a ranking more objective; it makes it later.

The two cost tabs answer different questions and must not be confused. The token tab is arithmetic only: $100 divided by list price, no capability, no coverage gate. The weighted tab multiplies that by (composite ÷ 53) × coverage factor, where the factor runs 0.50 / 0.75 / 1.00 / 1.25 / 1.50 for one to five sources. That factor is a stated prior, not a measurement, and it is the only place on this page where a judgement enters a number. Subscription figures are a community log, not vendor disclosure and not our own test: no vendor publishes a quota-to-token rate, so a plan can only be priced by measuring real usage and valuing it at list price.

Each source, ranked on its own

DeepSWE 1.1 45 models, all shown

Measured: runs dated 1 Sep – 16 Oct 2026, checked 16 Oct
79.0
MetaMuse Wonder Full
78.8
MetaMuse Wonder
77.9
GoogleGemini 4 Argon
75.4
MetaMuse Spark 1.3
75.2
OpenAIGPT 6.1 Sol
75.2
Fireworks AIEmber-1
74.2
AnthropicClaude Opus 5.5
74.2
DeepSeekDeepSeek V4.1 Flash
74.1
OpenAIGPT 6 Astra
73.7
GoogleGemini 3.8 Flash
73.0
OpenAIGPT 5.6 Sol
71.9
XiaomiMiMo V2.6-Pro
71.0
xAIGrok 4.7
71.0
AnthropicClaude Sonnet 5.5
70.0
OpenAIGPT 5.6 Terra
70.0
AnthropicClaude Fable 5
69.0
Moonshot AIKimi K3
68.8
OpenAIGPT 6 Sol
68.8
AnthropicClaude Opus 5
67.9
XiaomiMiMo V2.6-Flash
67.0
OpenAIGPT 5.6 Luna
67.0
OpenAIGPT 5.5
66.9
Z.aiGLM 5.3
66.6
OpenAIGPT 6 Luna
65.9
xAIGrok 4.6
65.3
GoogleGemini 3.7 Flash
63.4
Z.aiGLM 5.3-Flash
61.7
Mistral AIMistral Large 4
59.3
MetaMuse Spark 1.2
59.0
AnthropicClaude Opus 4.8
58.7
AlibabaQwen 3.8-Flash-Next
58.7
AlibabaQwen 3.8 Flash
56.6
AlibabaQwen 3.8 Max
54.0
AnthropicClaude Sonnet 5
54.0
xAIGrok 4.5
53.0
MetaMuse Spark 1.1
52.0
OpenAIGPT 5.4
49.0
GoogleGemini 3.6 Flash
44.0
Z.aiGLM 5.2
42.2
AlibabaQwen 3.8-27B
40.4
PoolsideLaguna S 2.1
37.0
GoogleGemini 3.5 Flash
31.0
Moonshot AIKimi K2.7 Code
30.0
AnthropicClaude Sonnet 4.6
12.0
GoogleGemini 3.1 Pro

Long-horizon software engineering. The site's current primary ranking.

FrontierSWE v2 15 models, all shown

Measured: snapshot 30 Sep 2026, refreshed quarterly
65.5
OpenAIGPT 6 Astra
62.3
AnthropicClaude Opus 5.5
61.9
AnthropicClaude Sonnet 5.5
55.1
GoogleGemini 4 Argon
52.0
AnthropicClaude Opus 5
47.0
AnthropicClaude Fable 5
32.2
OpenAIGPT 5.6 Sol
30.2
Z.aiGLM 5.3
29.5
xAIGrok 4.7
25.9
Moonshot AIKimi K3
25.3
xAIGrok 4.6
20.3
GoogleGemini 3.7 Flash
19.6
GoogleGemini 3.8 Flash
15.8
AlibabaQwen 3.8 Max
12.0
MetaMuse Spark 1.2

Ultra-long-horizon, 20-hour budget per task. Far from saturation, so it separates the top.

Terminal-Bench 4.0 22 models, all shown

Measured: runs dated 2 Sep – 29 Sep 2026
64.8
AnthropicClaude Opus 5.5
61.8
AnthropicClaude Sonnet 5.5
58.2
OpenAIGPT 6 Astra
58.2
OpenAIGPT 6.1 Sol
53.9
AnthropicClaude Opus 5
49.4
OpenAIGPT 6 Sol
44.5
AnthropicClaude Fable 5
41.8
Z.aiGLM 5.3
37.6
xAIGrok 4.7
37.3
OpenAIGPT 5.6 Sol
35.8
Z.aiGLM 5.3-Flash
27.0
AlibabaQwen 3.8 Max
23.6
AnthropicClaude Opus 4.8
21.5
OpenAIGPT 5.6 Terra
20.3
xAIGrok 4.6
19.1
GoogleGemini 3.8 Flash
17.3
OpenAIGPT 5.6 Luna
16.4
OpenAIGPT 6 Luna
14.5
MetaMuse Spark 1.3
12.4
xAIGrok 4.5
12.4
AnthropicClaude Sonnet 5
11.2
GoogleGemini 3.7 Flash

Scores model+harness pairs, so it partly measures the agent, not the model.

LiveBench 2.0 33 models, all shown

Measured: rolling, checked 16 Oct 2026
98.3
AnthropicClaude Fable 5
96.3
AnthropicClaude Opus 5.5
96.0
OpenAIGPT 6.1 Sol
94.7
AnthropicClaude Sonnet 5.5
94.3
xAIGrok 4.5
93.7
OpenAIGPT 6 Astra
93.1
OpenAIGPT 6 Sol
90.8
AnthropicClaude Opus 5
90.2
MetaMuse Spark 1.3
89.7
OpenAIGPT 5.6 Sol
89.1
xAIGrok 4.6
87.4
AlibabaQwen 3.8 Flash
86.8
AlibabaQwen 3.8-27B
86.2
MetaMuse Wonder
86.2
OpenAIGPT 5.6 Terra
85.4
DeepSeekDeepSeek V4.1 Flash
85.1
OpenAIGPT 5.5
84.5
xAIGrok 4.7
84.5
OpenAIGPT 6 Luna
82.8
Z.aiGLM 5.3
82.2
AnthropicClaude Sonnet 5
82.2
GoogleGemini 3.8 Flash
81.5
XiaomiMiMo V2.6-Pro
80.6
Z.aiGLM 5.3-Flash
80.6
MetaMuse Spark 1.2
80.5
GoogleGemini 3.7 Flash
77.3
AnthropicClaude Opus 4.8
73.6
OpenAIGPT 5.6 Luna
66.7
GoogleGemini 3.1 Pro
66.6
AnthropicClaude Sonnet 4.6
60.3
AlibabaQwen 3.8 Max
60.0
Mistral AIMistral Large 4
59.3
Moonshot AIKimi K3

Rotating questions resist contamination, but the coding column is saturated: the top eight sit above 90.

Code Arena 8 models, all shown

Measured: snapshot 13 Oct 2026, 878,407 votes
1811
AnthropicClaude Opus 5.5
1780
OpenAIGPT 6 Astra
1774
AnthropicClaude Sonnet 5.5
1760
OpenAIGPT 6.1 Sol
1690
AnthropicClaude Opus 5
1686
OpenAIGPT 6 Sol
1678
GoogleGemini 4 Argon
1673
AlibabaQwen 3.8 Max

The only blind, human-voted signal here, so it is the natural anti-gaming anchor. Caveat: the newest models carry the fewest votes (top entries ~1,800-2,600), so their Elo is still provisional.