Nex-N2.5 Benchmarks: How It Stacks Up Against Opus 5 and GPT-5.6
Nex-N2.5-Max trails Claude Opus 5 on coding benchmarks but leads open models on BrowseComp web-agent tasks. Full score breakdown.
What is Nex-N2.5 and how does it compare to closed frontier models?
Nex-N2.5 is an open-weight family of agentic models from Nex-AGI, released in three sizes (mini, Pro, and Max) and benchmarked against closed models including Claude Opus 5, GPT-5.6 Sol, and Kimi-K3. On coding benchmarks like SWE-Bench Pro and Terminal-Bench 2.1, Nex-N2.5-Max trails Claude Opus 5 by a meaningful margin. But on BrowseComp, a web-agent benchmark, Nex-N2.5-Max posts the highest score in the entire comparison table, ahead of every closed model tested.
TL;DR
- Nex-N2.5-Max is Nex-AGI’s largest model in the family, built on a 1.6-trillion-parameter Mixture-of-Experts foundation, and it’s the company’s first full post-training run at that scale.
- On SWE-Bench Pro, Nex-N2.5-Max scores 65.7 versus Claude Opus 5’s 79.2, a gap of over 13 points that shows real-world software engineering tasks remain the toughest category for the open model.
- On BrowseComp, a web-browsing agent benchmark, Nex-N2.5-Max scores 92.6, edging out Claude Opus 5 (90.8), GPT-5.6 Sol (90.4), and Kimi-K3 (91.2), the only benchmark in the suite where it takes the top spot outright.
- The three model sizes (mini, Pro, Max) show a clean scaling pattern: mini and Pro build on Nex-N2’s multimodal foundations for computer use and web browsing, while Max is text-only and optimized for raw reasoning and coding depth.
- On multimodal and computer-use benchmarks like OSWorld-Verified and OSWorld-2, Nex-N2.5-Pro competes respectably but Qwen3.8-Max and Claude Opus 5 hold the top scores.
- All three sizes are released as open weights on Hugging Face and ModelScope, with hosted access also available through OpenRouter, though Nex-N2.5-Pro’s weights were still pending release at benchmark publication.
- The benchmarks come from Nex-AGI’s own evaluation suite using its NexAU and NexCUA harnesses, plus scores pulled from official leaderboards where available, so figures should be read as vendor-reported rather than independently reproduced.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How does Nex-N2.5 perform on coding benchmarks?
Coding is where the gap between Nex-N2.5-Max and closed frontier models is most visible. On Terminal-Bench 2.1, Max scores 86.1 against Claude Opus 5’s 89.1, GPT-5.6 Sol’s 88.8, and Kimi-K3’s 88.3, a tight cluster where Max sits at the bottom but isn’t far off.
The gap widens on SWE-Bench Pro, a benchmark built around real-world software engineering pull requests. Max scores 65.7, while Claude Opus 5 leads at 79.2. GPT-5.6 Sol (64.6), Kimi-K3 (63.3), and GLM-5.3 (64.6) all land close to Max, suggesting Opus 5 has a distinct edge in this category rather than the whole open-model field falling short by a similar margin.
DeepSWE v1.1 tells a similar story: Max hits 65.6, trailing Claude Opus 5’s 73.7 and GPT-5.6 Sol’s 72.7. Across all three coding benchmarks, the pattern holds. Nex-N2.5-Max is competitive with the mid-tier of frontier and open models but doesn’t catch Claude Opus 5, which leads every coding benchmark in the table.
The smaller sizes scale down predictably. Nex-N2.5-Pro scores meaningfully lower than Max across the board (61.2 on SWE-Bench Pro, for instance), and Nex-N2.5-mini drops further still (43.8). This is a straightforward capacity trade-off: bigger model, better coding performance, roughly in line with what you’d expect from parameter scaling.
Where does Nex-N2.5 actually lead?
BrowseComp is the standout result. This benchmark measures how well an agent can browse the web to complete tasks, and Nex-N2.5-Max’s score of 92.6 is the best in the table, ahead of Claude Opus 5, GPT-5.6 Sol, and Kimi-K3. Nex-AGI notes it applies a “Summary” context-compaction strategy once token usage crosses 60% of the context window during these runs, which is a detail worth flagging since context-management strategy can meaningfully affect long-horizon browsing scores.
Nex-N2.5-Max also comes close to the top on AutomationBench (50.2 versus Claude Opus 5’s 50.3, a near-tie) and ties Kimi-K3 for the top score on Toolathlon Verified at 76.5 (Claude Opus 5 also ties at 76.5). These are agentic workflow benchmarks that test multi-step task execution and tool use rather than raw code generation, and it’s in this category that Nex-N2.5-Max looks most competitive with closed frontier models.
Where it falls short again is Job Bench and GDPval-AA v2, both benchmarks aimed at knowledge-work and productivity tasks. Claude Opus 5 leads both by wide margins (65.7 vs. 53.6 on Job Bench; 1831 vs. 1713 on GDPval-AA v2), suggesting the gap between Nex-N2.5 and Opus 5 is more pronounced on open-ended professional tasks than on narrower tool-use benchmarks.
How does Nex-N2.5 handle multimodal and computer-use tasks?
Multimodal and computer-use capability is central to Nex-N2.5’s pitch since mini and Pro are built to operate computers and browsers using visual feedback, treating vision as an interface for verifying task outcomes rather than just an input channel. Nex-N2.5-Max, notably, is text-only, so multimodal comparisons run against mini and Pro instead.
One coffee. One working app.
You bring the idea. Remy manages the project.
On OSWorld-Verified, a benchmark for computer-use agents, Nex-N2.5-Pro scores 82.2, close to Claude Opus 5 (83.4) and GPT-5.6 Sol (83.2), but Qwen3.8-Max leads the field at 86.1. On the harder OSWorld-2 benchmark, the gap widens: Pro scores 56.4 against Claude Opus 5’s 68.3, a difference of nearly 12 points.
Nex-N2.5-Pro does better on grounding-specific tasks. On OSWorld-G, which tests an agent’s ability to accurately locate and interact with UI elements, Pro scores 87.4, the best result in that row, ahead of Claude Opus 5 (76.8), GPT-5.6 Sol (77.7), and Qwen3.8-Max (84.9). This suggests Nex-N2.5’s visual grounding, the mechanical accuracy of clicking and interacting with interface elements, is a relative strength even where broader task completion (OSWorld-2) lags.
On Vision2Web and SWE-MM, GPT-5.6 Sol and Claude Opus 5 respectively hold clear leads, and Nex-N2.5-Pro sits behind both.
Is Nex-N2.5 worth using over a closed frontier model?
The answer depends on what the task looks like. For pure coding work, especially anything resembling real-world software engineering pull requests, Claude Opus 5 has a clear and consistent lead over Nex-N2.5-Max across every coding benchmark in the comparison. Teams optimizing purely for code-generation accuracy have little reason to switch.
For web-browsing agents and certain tool-use workflows, Nex-N2.5-Max’s BrowseComp score and its near-ties on AutomationBench and Toolathlon Verified make it a legitimate open-weight alternative, particularly for teams that want to self-host or fine-tune rather than rely on an API.
The open-weight angle is the other half of the calculation. Nex-N2.5-Max, Pro, and mini are all released with weights on Hugging Face and ModelScope (Pro’s weights were still pending at the time of this benchmark release), alongside hosted access via OpenRouter. That gives developers the option to run the models on their own infrastructure, which closed models don’t offer regardless of benchmark scores. Deployment isn’t trivial, though: Nex-N2.5-Max requires a multi-node setup (the documented configuration uses 16 H200 GPUs across two nodes), while Pro and mini run on single nodes with 8 and 2 H100 GPUs respectively.
Frequently Asked Questions
What sizes does Nex-N2.5 come in?
Three: Nex-N2.5-mini, Nex-N2.5-Pro, and Nex-N2.5-Max. Mini and Pro build on Nex-N2’s multimodal foundations for computer use and browsing. Max is a text-only model built on a 1.6-trillion-parameter Mixture-of-Experts foundation.
Does Nex-N2.5-Max beat Claude Opus 5 on any benchmark?
Yes, on BrowseComp, a web-browsing agent benchmark, where Nex-N2.5-Max scores 92.6 against Claude Opus 5’s 90.8. It also ties Claude Opus 5 and Kimi-K3 on Toolathlon Verified at 76.5.
How big is the gap between Nex-N2.5 and Claude Opus 5 on coding?
On SWE-Bench Pro, the gap is about 13.5 points (65.7 vs. 79.2). On Terminal-Bench 2.1, it’s closer, about 3 points (86.1 vs. 89.1).
Are the Nex-N2.5 benchmark scores independently verified?
No. The scores come from Nex-AGI’s own evaluation harnesses (NexAU for coding, NexCUA for computer-use and browser tasks), supplemented by figures pulled from other providers’ published leaderboards where available. They should be treated as vendor-reported results.
Can I run Nex-N2.5 myself?
Yes, model weights for mini and Max are available on Hugging Face and ModelScope, with Pro’s weights listed as forthcoming at release time. Nex-AGI also provides a prebuilt Docker image with a customized sglang backend, though Max requires a multi-node, multi-GPU setup while Pro and mini can run on a single node.