# Claude Opus 5.5 Tops Code2Video-AA; GPT-6.1 Sol Is 90% Cheaper

> Artificial Analysis' new code-to-video benchmark reveals a sharp quality-versus-cost split between frontier models and open-weights systems.

*Published 2026-10-27 · By Marcus Vance*

Code2Video-AA is an independent benchmark from Artificial Analysis that scores AI models on their ability to produce short product and brand marketing videos from design briefs by writing HyperFrames HTML code.

A new independent benchmark for code-to-video marketing content shows Claude Opus 5.5 leading the field with a 1536 Elo score, while GPT-6.1 Sol finishes 167 points behind at roughly 90% lower cost per task, according to results published Oct. 26, 2026, by Artificial Analysis.

The benchmark, Code2Video-AA, evaluates 27 AI models across 114 design briefs drawn from the public split of HeyGen's Code2Video Bench. Each model receives a short product or brand marketing brief and must produce a single HyperFrames HTML composition, which is rendered to video and judged on five axes: Engagement, Prompt Intent, Composition, Temporal Consistency, and Craft.

## What is Code2Video-AA and how is it scored?

Artificial Analysis describes Code2Video-AA as its independent implementation of HeyGen's Code2Video Bench, a set of design briefs for short product and brand videos. Each model designs the video in code, rather than generating pixels directly with a text-to-video diffusion system.

HeyGen Research introduced the underlying task in its Code2Video benchmark, defining it as a workflow in which an LLM agent generates code that a rendering framework, such as HyperFrames, deterministically renders into the final video. The deterministic step is central: because the same code always produces the same video, benchmark scores reflect the model's design ability rather than the renderer's stochastic behavior.

Artificial Analysis' implementation introduces a judge model to compare outputs. The results are derived from more than 35,000 pairwise verdicts collected by a Gemini 4 Argon judge, with Elo ratings reported alongside 95% confidence intervals. Using a single judge model keeps the comparison consistent across all 27 models, though it also means the rankings inherit whatever biases the judge model has.

The five scoring axes are designed to cover the dimensions that matter in short marketing video. Engagement measures how likely a viewer is to keep watching; Prompt Intent measures whether the video follows the brief; Composition evaluates layout and framing; Temporal Consistency checks whether elements remain stable across frames; Craft assesses overall production polish. A model must balance all five to reach the top of the leaderboard.

Code2Video-AA sits at the intersection of two fast-moving capabilities: long-context code generation and deterministic rendering. HyperFrames, the rendering framework used by the benchmark, is HTML-based, which means the model's task overlaps with familiar web code patterns while still requiring spatial and temporal reasoning. Models that write clean, well-structured HTML tend to score higher because the renderer has fewer opportunities to introduce visual artifacts.

- The model receives one of 114 design briefs for a short product or brand video.
- The model writes a single HyperFrames HTML composition describing the video.
- The HTML is rendered deterministically to a video file.
- A Gemini 4 Argon judge compares pairs of videos across five axes.
- Pairwise verdicts are aggregated into Elo scores with 95% confidence intervals.

## How large is Claude Opus 5.5's lead on Code2Video-AA?

Claude Opus 5.5, evaluated in its max configuration with default fallback, scores 1536 Elo, 93 points ahead of the next model and first on all five judge axes, according to Artificial Analysis. The model's lead is consistent across axes, which is notable because some models excel at prompt adherence but produce poorly composed frames.

Claude Sonnet 5.5 follows at 1443, with GPT-6 Astra at 1427 and GPT-6.1 Sol at 1369, according to the benchmark page. The ordering places two Anthropic models and two OpenAI models in the top four, with no open-weights model inside the top group.

Pairwise judging with a single judge model is a deliberate methodological choice. Elo scores derived from pairwise comparisons are more stable than absolute scores, and using one judge removes inter-judge variance, but it also means the leaderboard is a measure of what Gemini 4 Argon prefers. A different judge model could produce a different ordering, especially in close contests such as the 1-point separation between the two MiMo variants.

| Model | Code2Video-AA Elo | Published cost per task |
| --- | --- | --- |
| Claude Opus 5.5 (max with fallback) | 1536 | $3.08 |
| Claude Sonnet 5.5 (max with fallback) | 1443 | Not disclosed |
| GPT-6 Astra (max) | 1427 | Not disclosed |
| GPT-6.1 Sol (max) | 1369 | $0.31 |
| DeepSeek V4.1 Flash (max) | 1053 | Not disclosed |
| MiMo-V2.6-Flash | 1052 | Not disclosed |
| MiMo-V2.6-Pro | 1052 | Not disclosed |

## What does the cost comparison show?

GPT-6.1 Sol (max) costs about $0.31 per task, compared with $3.08 for Claude Opus 5.5, a roughly 90% reduction for a 167-point Elo deficit. Artificial Analysis' published summary frames the tradeoff directly: Claude Opus 5.5 leads by 93 points, but costs about 10 times more than GPT-6.1 Sol, which trails by 167 points.

Across the 27 models evaluated, cost per task spans more than 300x, from $0.01 to $3.89, according to Artificial Analysis. That spread is unusually large for a single task class, and it means the value ranking of models changes depending on a team's budget and quality threshold.

The cost difference compounds with volume. A marketing team that renders thousands of videos per month is making a procurement decision, not a one-off benchmark comparison, and the gap between $0.31 and $3.08 per task can reach tens of thousands of dollars monthly. That math favors cheaper models even when their Elo scores are lower, as long as outputs clear a brand's quality bar.

Cost also shapes how enterprises can use the benchmark. Because Code2Video-AA requires a full HTML generation pass per brief, expensive frontier models are costly to evaluate at scale, and the same is true for a company running its own private briefs. The 114 public briefs keep the benchmark manageable, but a serious internal evaluation will multiply that cost by the number of prompts and model configurations tested.

## Why do open-weights models lag so far behind?

Open-weights models trail by at least 483 Elo points on Code2Video-AA, with DeepSeek V4.1 Flash (max) at 1053 and MiMo-V2.6-Flash and MiMo-V2.6-Pro both at 1052, according to Artificial Analysis. The gap is large enough that an open-weights model would need a substantial improvement in code generation to reach the top tier.

The code-to-video task exposes weaknesses that text benchmarks often miss. Producing a complete HyperFrames composition requires the model to maintain a coherent visual plan across dozens of elements, keep animation timing consistent, and encode visual intent in HTML rather than in natural language. Open-weights models have narrowed gaps in code generation generally, but the motion-design layer adds constraints that current open-weights systems handle less reliably.

The open-weights gap is not static. The two MiMo-V2.6 variants and DeepSeek V4.1 Flash are separated by only a point, suggesting that the field of open-weights contenders is clustered rather than dominated by a single approach. Future releases could shift the cluster upward, but the current snapshot shows a clear separation from closed frontier models.

The closed-model lead also reflects training priorities. Anthropic and OpenAI have invested heavily in agentic coding and long-horizon planning, while many open-weights releases focus on general chat and instruction following. Code-to-video rewards the kind of sustained, structured output that agentic coding models are explicitly trained to produce.

## What are the market implications for enterprises choosing a model?

Code2Video-AA gives procurement teams a cost-quality frontier rather than a single winner. Marketing organizations that need polished, on-brand motion graphics may prefer Claude Opus 5.5 despite its price; teams that produce large volumes of social-first video at lower budgets may accept the 167-point gap in exchange for $0.31 per task.

The benchmark also highlights a workflow difference from conventional text-to-video generation. Because the model writes code that a renderer executes, the output is deterministic and editable, which matters for brand consistency, version control, and review cycles. An agency can rerender an approved composition at different aspect ratios or resolutions without asking the model to invent a new video.

Model choice in code-to-video is not only about Elo. The judge-based scoring measures average quality across 114 briefs, but a brand's own briefs may concentrate on a narrower style, such as product close-ups or kinetic typography. Enterprises should validate the top models on their own prompts before locking in a configuration, and the benchmark provides a structured way to start that validation.

## What do the benchmark authors say about the results?

> Claude Opus 5.5 (max with fallback, 1536) leads Code2Video-AA by 93 points, but costs $3.08 per task, ~10x more than GPT-6.1 Sol (max, 1369), which trails by 167 points.Artificial Analysis

Artificial Analysis published the result as a cost-performance comparison rather than a simple leaderboard announcement. The organization's launch article states that Claude Opus 5.5 (max with fallback) leads with 1536 Elo, is 93 points ahead of the next model, and is first on all five judge axes, while also noting the $3.08 per-task cost.

The emphasis on cost is notable for an independent benchmark organization. By publishing per-task prices alongside Elo scores, Artificial Analysis is asking readers to weigh quality against expense, a framing that matches how enterprise buyers actually select models.

The benchmark's independence is part of its value. Rather than relying on vendor-reported scores, Code2Video-AA runs the models through a consistent pipeline and reports the methodology alongside the results, which makes the leaderboard reproducible and comparable across releases.

## What comes next for code-to-video benchmarks?

Code2Video-AA is a snapshot as at Oct. 26, 2026, and the benchmark is designed to be rerun as new models are released, according to Artificial Analysis. The open-weights gap of at least 483 points could narrow if the next generation of open-weights models improves HTML composition generation and temporal consistency.

The benchmark's methodology is also likely to evolve. Pairwise judging with a single judge model can be extended to more judges or to human evaluation, and the 114 briefs could be expanded to cover more industries and video lengths. Any such change would reset the leaderboard, so teams should track methodology versions when comparing future results.

Buyers should also watch pricing stability. Per-task costs are tied to model providers' API pricing, which can change without notice, and a 10x price gap could narrow if Anthropic or OpenAI adjust rates. The benchmark's cost figures are a point-in-time input to a procurement decision, not a durable contract term.

For now, the practical takeaway is that cost and quality are decoupled in code-to-video work. Claude Opus 5.5 is the quality leader, GPT-6.1 Sol is the value pick among closed models, and open-weights models have a clear mountain to climb.

## Sources

1. [Launch article for Code2Video-AA showing Claude Opus 5.5 (max with fallback) leading at 1536 Elo, 93 points ahead of the next model and first on all five judge axes, with results as at Oct. 26, 2026.](https://artificialanalysis.ai/articles/code2video-aa)
2. [Code2Video-AA benchmark page reporting Claude Opus 5.5 at 1536, Claude Sonnet 5.5 at 1443, and GPT-6 Astra at 1427, with 114 design briefs and 27 models evaluated.](https://artificialanalysis.ai/evaluations/code2video-aa)
3. [GPT-6.1 Sol (max) trails Claude Opus 5.5 by 167 Elo points at 1369 Elo while costing $0.31 per task versus $3.08 for Claude Opus 5.5.](https://x.com/ArtificialAnlys/status/2114845067026571365)
4. [HeyGen's Code2Video benchmark frames the code-to-video task as an LLM agent generating code that a rendering framework such as HyperFrames deterministically renders into the final video.](https://www.heygen.com/research/introducing-code2video-benchmark)

---
Source: https://aiintelreport.com/frontier-models/claude-opus-5-5-code2video-aa-gpt-6-1-sol-cost
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt