Nex-N2.5: free agentic coders on OpenRouter that drive a real PC
Pro and Mini are free today; the 1.6T Max leads the family. Vision-native computer use, Apache-2.0 weights rolling out, 262K context, tool calling.
Copy markdown
Free right now on OpenRouter
Nex-N2.5-Pro and Nex-N2.5-Mini are both live and free for a limited time on OpenRouter, with tool calling and adjustable reasoning effort. 262K-token context and ~58 tok/s on Pro. Point Cursor, Claude Code, or your own harness at a single endpoint and try them today.
The scores: Pro trades blows with GPT-5.6
Nex-N2.5-Pro posts 82.7 on Terminal-Bench 2.1, 61.2 on SWE-Bench Pro, 82.2 on OSWorld-Verified and 89.7 on BrowseComp — roughly on par with GPT-5.6 Sol on agentic work while trailing Claude Opus 5 on pure coding. Mini is the cheap runner (73.4 / 43.8); Max tops the chart.
Computer use is the actual headline
These are vision-native agents that operate a real browser and desktop: explore a repo, make multi-file edits, run commands, then look at the screen to verify the result. Build this weekend — a self-checking agent that drives an app end-to-end instead of guessing from text output.
The 1.6T Max, and the hardware tax
Nex-N2.5-Max is a 1.6-trillion-parameter text-only MoE that leads the family (86.1 Terminal-Bench, 65.7 SWE-Bench Pro, 92.6 BrowseComp). Self-hosting scales from 2×H100 for Mini to 16×H200 across two nodes for Max, via their nexagi/sglang fork.
The catch: 'open,' with an asterisk
Nex calls the family Apache-2.0, with weights promised on Hugging Face and ModelScope — but the Pro card still reads 'weights coming soon,' so today you get the hosted API, not a download. The free OpenRouter tier is limited-time, so lock in your evals now.
Elsewhere: OpenAI ships prompt-cache diagnostics
OpenAI made Prompt Cache Diagnostics generally available in the Responses API for GPT-5.6 and later, so you can see exactly where a request missed the cache and stop paying full price to re-send the same context every turn.