Results as of 2026-09-22T20:31:36Z: 13 models, 145,509 answers

README.md

@@ -26,13 +26,13 @@ configs:
 
 # VLADBench-reeval
 
-We re-evaluated the VLADBench benchmark (Li et al., 2025, [arXiv:2503.21505](https://arxiv.org/abs/2503.21505)) against current SOTA VLMs under the original scoring criteria and prompts. Please see  [Eventual-Inc/VLADBench](https://github.com/Eventual-Inc/VLADBench) for our implementation. The report with interactive figures is on the [Eventual blog](). 
+We re-evaluated the VLADBench benchmark (Li et al., 2025, [arXiv:2503.21505](https://arxiv.org/abs/2503.21505)) against current SOTA VLMs under the original scoring criteria and prompts. See [Eventual-Inc/VLADBench](https://github.com/Eventual-Inc/VLADBench) for the code, and the [companion site](https://eventual-inc.github.io/VLADBench/) for interactive results.
 
 ## Cost vs Score
 
-![Cost vs Score: TOTAL against sweep spend for every model, with the Pareto frontier](cost-vs-score.png)
+![Cost vs Score: TOTAL against sweep cost for every model, with the Pareto frontier](cost-vs-score.png)
 
-Up and to the left is better. The dashed line is the cost-performance frontier which is defined by measuring the best TOTAL score available against it's cost. 
+Up and to the left is better. The dashed line is the cost-performance frontier which is defined by measuring the best TOTAL score available against its cost. 
 
 ## Leaderboard
 
@@ -40,22 +40,21 @@ Up and to the left is better. The dashed line is the cost-performance frontier w
 | # | Model | Score | Sweep cost | Input $/frame | Output $/query | Est. $/hour of video |
 |---|---|---|---|---|---|---|
 | 1 | GPT-6 Astra | 83.60 | $542.16 | $0.01377 | $0.00113 | $50.68 |
-| 2 | Gemini 3.8 Flash | 75.93 | $16.93 | $0.00005 | $0.00077 | $0.55 |
-| 3 | GPT-5.6 Sol | 75.56 | $106.31 | $0.00274 | $0.00007 | $10.03 |
-| 4 | Qwen 3.8 Max | 75.38 | $89.26 | $0.00189 | $0.00183 | $7.72 |
-| 5 | Muse Spark 1.3 | 73.10 | $64.03 | $0.00016 | $0.00349 | $2.20 |
-| 6 | GPT-5.6 Luna | 65.29 | $10.69 | $0.00028 | $0.00001 | $1.01 |
-| 7 | MiniMax M3 | 64.35 | $11.27 | $0.00005 | $0.00058 | $0.44 |
-| 8 | Gemini 2.5 Flash Lite | 64.04 | $5.53 | $0.00003 | $0.00032 | $0.24 |
-| 9 | Qwen 3.6 35B A3B | 62.96 | $2.67 | $0.00006 | $0.00000 | $0.22 |
-| 10 | Qwen 3.8 27B | 61.64 | $5.56 | $0.00010 | $0.00003 | $0.38 |
-| 11 | Gemma 4 31B | 59.66 | $0.49 | $0.00001 | $0.00000 | $0.04 |
+| 2 | Claude Opus 5.5 | 82.73 | $205.28 | $0.00480 | $0.00159 | $18.18 |
+| 3 | Gemini 3.8 Flash | 75.93 | $16.93 | $0.00005 | $0.00077 | $0.55 |
+| 4 | GPT-5.6 Sol | 75.56 | $106.31 | $0.00274 | $0.00007 | $10.03 |
+| 5 | Qwen 3.8 Max | 75.38 | $89.26 | $0.00189 | $0.00183 | $7.72 |
+| 6 | Muse Spark 1.3 | 73.10 | $64.03 | $0.00016 | $0.00349 | $2.20 |
+| 7 | GPT-5.6 Luna | 65.29 | $10.69 | $0.00028 | $0.00001 | $1.01 |
+| 8 | MiniMax M3 | 64.35 | $11.27 | $0.00005 | $0.00058 | $0.44 |
+| 9 | Gemini 2.5 Flash Lite | 64.04 | $5.53 | $0.00003 | $0.00032 | $0.24 |
+| 10 | Qwen 3.6 35B A3B | 62.96 | $2.67 | $0.00006 | $0.00000 | $0.22 |
+| 11 | Qwen 3.8 27B | 61.64 | $5.56 | $0.00010 | $0.00003 | $0.38 |
+| 12 | Gemma 4 31B | 59.66 | $0.49 | $0.00001 | $0.00000 | $0.04 |
 
-See the code for how we calculated the 
+See the code for how we calculated the metered video costs.
 
 
-Also in the data but left out of the figure and this table: Reka Edge (39.78, 7B edge model scoring 39.8; kept in the data, left out of the headline figures as an outlier). `models.featured` marks them.
-
 The same table from the parquet files with [Daft](https://www.daft.ai):
 
 ```python
@@ -80,30 +79,29 @@ leaderboard.show(models.count_rows())
 
 ## Re-Evaluating VLADBench in 2026
 
-We re-evaluated 12 current vision-language models on VLADBench with our own harness, the original paper's  scoring criteria and prompts. The benchmark is unchanged using the same questions, frames, gold references, scorer, and per-task weights. 
+We re-evaluated 13 current vision-language models on VLADBench with our own harness, the original paper's scoring criteria and prompts. The benchmark is unchanged using the same questions, frames, gold references, scorer, and per-task weights. 
 
 Here is where our setup differs from the paper's:
 
 - **Not all models support video inputs**. Frame sequences are passed as video where supported and ordered images otherwise. Most tasks only provide 2 or 3 frames so we expect the temporal perception impact to be smaller than 8-16 frame clips. Alibaba ended up rejecting clips under four frames, and OpenAI only accepts images, so those models received the frames as ordered images. Input differences are labeled as such. 
-- **Reasoning is set to its lowest available level**. Several providers do not expose a means to disable reasoning all together, so the lowest reasoning settings is used.
+- **Reasoning is set to its lowest available level**. Several providers do not expose a means to disable reasoning altogether, so the lowest reasoning setting is used. GPT-6 Astra is the exception: it ran at low, though its endpoint also accepts minimal.
 
 We think these are reasonable trade-offs given the structural differences in how models are served today. 
 
 ## Scoring
 
-The overall score is a question-weighted average. VLADBench is comprised of 28 tasks, each with its own score. To roll scores from all tasks, we average the task scores with each task weighted by how many questions it has, so a task with 795 questions counts about four times as much as one with 200. The paper does not say how it rolled up its own numbers, but this rule reproduces the group averages and overall scores printed in its Table 10, so we use it too. Those rows are labelled MEAN and TOTAL, as in the paper.
-one reproduces the published rows for the paper's own models.
+The overall score is a question-weighted average. VLADBench is comprised of 28 tasks, each with its own score. To roll scores from all tasks, we average the task scores with each task weighted by how many questions it has, so a task with 795 questions counts about four times as much as one with 200. The paper does not say how it rolled up its own numbers, but this rule reproduces 241 of the 250 group averages printed in its Table 10 to within 0.1, so we use it too. We run the paper's own scoring code unchanged. Those rows are labelled MEAN and TOTAL, as in the paper.
 
 ## Tables
 
 
 | Table         | Rows          | One row is                                                                                           |
 | ------------- | ------------- | ---------------------------------------------------------------------------------------------------- |
-| `models`      | 12    | a model: label, lab, provider string, protocol declarations, spend, latency summary, TOTAL           |
+| `models`      | 13    | a model: label, lab, provider string, protocol declarations, spend, latency summary, TOTAL           |
 | `tasks`       | 28     | a task: category, group, scorer function, the paper's component weights, sample and question counts  |
 | `questions`   | 11,193 | a released question: prompt, pinned frame URLs, gold reference, country                              |
-| `answers`     | 134,316   | one model's answer to one question: text, finish reason, tokens, cost, latency, protocol hash        |
-| `task_scores` | 336    | one model on one task: the paper's scoring criteria applied to our answers, components and composite |
+| `answers`     | 145,509   | one model's answer to one question: text, finish reason, tokens, cost, latency, protocol hash        |
+| `task_scores` | 364    | one model on one task: the paper's scoring criteria applied to our answers, components and composite |
 
 
 ## Models
@@ -114,6 +112,7 @@ one reproduces the published rows for the paper's own models.
 | | Alibaba | Qwen 3.8 27B | `qwen/qwen3.8-27b` | 27B dense | video | off | 8192 | 2 |
 | Alibaba | Qwen 3.6 35B A3B | `qwen/qwen3.6-35b-a3b` | 35B MoE, 3B active | video | off | 8192 | 0 |
 | Alibaba | Qwen 3.8 Max | `qwen/qwen3.8-max-0902` | 2.4T MoE | images | low | 8192 | 0 |
+| Anthropic | Claude Opus 5.5 | `anthropic/claude-opus-5.5` | undisclosed | images | low | 8192 | 0 |
 | Google | Gemma 4 31B | `google/gemma-4-31b-it` | 30.7B dense | video | off | 8192 | 0 |
 | Google | Gemini 2.5 Flash Lite | `google/gemini-2.5-flash-lite` | undisclosed | video | low | 8192 | 0 |
 | Google | Gemini 3.8 Flash | `google/gemini-3.8-flash` | undisclosed | video | low | 8192 | 0 |
@@ -127,14 +126,22 @@ one reproduces the published rows for the paper's own models.
 
 Gemini 3.8 Flash also started under it; its cut answers were re-asked under the live 8192 guard and the rest carried over (11,137 answers, `models.carried_from_512`). Full declarations and footnotes are in `docs/PROTOCOL.md` in the repository.
 
-## Scoring
-
-We follow the same scoring criteria as the original paper, running its own scoring code unchanged, and roll tasks up the way the paper does by weighting each task by its number of questions. Reference the repository for the code. 
-
 ## Cost and latency
 
 `answers.cost_usd` is the provider-billed cost per answer as returned by OpenRouter, and `models.cost_usd` sums it over the sweep. Prices are as billed on the dates in `answers.timestamp`. `answers.elapsed_seconds` is wall-clock time per request as seen by the harness; `models` carries the 25th, 50th, 75th, and 95th percentiles.
 
+## Known issues
+
+- **Bounding boxes are scored in pixels.** Three tasks ask for a box and the prompt doesn't say what units to use. OpenAI models answer in pixels. Qwen, Muse, Gemma, MiniMax, and Reka answer on a 0 to 1000 grid, and Gemini answers on that grid with y before x, so most of their boxes miss. The paper's scorer rescaled boxes for some model families by name; we don't. Read on each model's own grid, Qwen 3.8 Max goes from 75.4 to 78.3.
+- **Vehicle cut-in references are almost all "yes".** 172 of the 174 yes/no references are "yes", so answering yes to everything scores about 99% on that component. The question also asks whether the vehicle has "the intention to cross the road", which reads oddly for a cut-in.
+- **Some references aren't among the options.** On Traffic Light, 7 of 795 references aren't in the list the question offers, like "Malfunction" and "Manualt".
+- **Weather and Light labels are ambiguous.** Every model lands within a few points of the others, and 17% and 20% of the questions stump every model. Labels like "cloudy" versus "overcast" and "dawn&dusk" don't always match what the frames show.
+- **Descriptive answers are exact match.** Vehicle and VRU behavior are graded against one short reference, so a reasonable answer in different words scores zero.
+- **Two Qwen 3.8 27B answers hit the completion guard.** They're scored as returned, which by our own protocol means that run isn't protocol-complete.
+- **Every model ran once.** Differences under about two points could be run-to-run noise.
+
+The Caveats tab on the [companion site](https://eventual-inc.github.io/VLADBench/#caveats) walks through each of these with the per-question answers.
+
 ## Citation
 
 ```bibtex
@@ -150,4 +157,4 @@ We follow the same scoring criteria as the original paper, running its own scori
 ## Source data
 
 Questions and frames were sourced from the released VLADBench annotations at revision `1895f222` of
-[depth2world/VLADBench.](https://huggingface.co/datasets/depth2world/VLADBench)  This dataset adds model answers and scores only. Answer text is published as returned by each provider.
+[depth2world/VLADBench](https://huggingface.co/datasets/depth2world/VLADBench). This dataset adds model answers and scores only. Answer text is published as returned by each provider.

answers.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:128936a9193aa5a942cc83eb1e16d92da06770470b8a61d8b48dc054b7842ce2
-size 8637772
+oid sha256:172327e5ad033bd6d54487f2e7ca04340d8aea1eca93487a7fb2bc75aa87621d
+size 9305620

models.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fb57bcff3ccc029eb554925e4d77122d111fc35123a13cca7b5ba6b46468096a
-size 10354
+oid sha256:25bf4114918010644ae17e478195c36571effb506fde6a3f6a8cc3693af2a8b5
+size 10484

task_scores.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c3b77189b6b23a88c632d5237beb98162ac58b857f3e4f8b3609155f9ea5c03f
-size 11984
+oid sha256:2b6a81b8abcc51606413ed9c35780ff0f7c5d8cd36de233b0731dff92e833f12
+size 12584