Claude Opus 5, Now Available in Model ML

Claude Opus 5 is now available in Model ML.

Ahead of its release, we tested Opus 5 across the workflows our clients rely on the most: quantitative research, multi-document analysis, and multi-step financial execution.

Here’s what we found, and how Opus 5 compares to the other frontier and open-source models we run:

Quantitative research

On number-heavy analysis over market data, including multi-company screens, regressions, and filing reconciliations, Opus 5 is on par with GPT-5.6 Sol and slightly ahead of Fable 5. While open-source models still don’t perform at the level of closed-source models in this category, the gap is closing: GLM 5.2 and DeepSeek v4 Pro are starting to do credible quant work at a fraction of the cost.

Multi-document intelligence

Many of the most valuable questions in financial analysis cannot be answered from a single source. They require synthesizing information distributed across data rooms, cross-filings, and research memos. On this class of task, Opus 5, Fable 5, and Gemini 3.5 Flash performed at the top of the field. GLM 5.2 trailed only marginally, and its substantially lower cost makes it a credible option for teams balancing analytical performance against operational spend.

Financial workflow execution

Finance professionals often have to undergo multi-step tasks such as comps pulls, leverage bridges, and guidance checks. In this class of task, Opus 5 is at the level of GPT-5.6 Sol and Fable 5. On the other hand, this is another category where open-source models are also starting to show strong results, with DeepSeek v4 Pro and Kimi K2.6 close behind.

Cost and token efficiency

Opus 5 and Fable 5 are the most expensive models to run. That premium buys analysis depth, which the quant and multi-document results above justify. However, GPT-5.6 Sol delivers comparable performance while using up to 66% fewer tokens on the same tasks. The strongest open-source models we ran on our own GPUs, GLM 5.2 and DeepSeek v4 Pro, now reach usable accuracy at a few cents per task.

Why model-agnostic still wins

Our findings show that no single model wins on both price and performance at once. That is precisely why Model ML remains model-agnostic, routing each task to whatever model is best suited to it. As new models ship, that routing layer means our clients benefit from every improvement without having to commit to one vendor.