Free agentic model ships with built-in visual feedback loop

Nex-N2.5-Pro launched on 8 September 2026 as an agentic model that accepts text and image input and runs a visual feedback loop directly through its API. Unlike models that only generate code, it is built to turn goals into verified outcomes: it explores codebases, makes multi-file edits, executes commands, launches applications, interfaces with browsers and desktops, and tests software from the user's perspective. When observed behaviour diverges from intent, the model diagnoses, fixes and re-tests — making it relevant for autonomous software engineering, GUI-based QA, computer-use automation, and research workflows where success is measured in the environment, not just in generated code.

Technical specs and what they mean in practice

The model is offered free with a 262 thousand-token context window, large enough to hold entire medium-sized codebases in a single conversation. Throughput sits at 32 tokens per second (P50, best among providers), with median latency of 1.72 seconds (P50, best provider). Reported availability is 99.97%, and a rerouting mechanism recovers from upstream provider errors in 59.68% of cases, provided request filters allow it. Per-provider availability data is accessible via API endpoints, with load-balancing and customisation options.

The visual feedback loop, not just a slogan

The core capability is that "visual feedback loop": the model does not just write code — it sees the result, screenshots, DOM state, terminal output, and compares it against the intent. When it detects a gap, it corrects iteratively. This is fundamentally different from the classic "write and run" approach where a model generates code once and hopes for the best. Here the loop is closed: act, observe, diagnose, fix, re-observe. For developers that means fewer manual fix cycles, but it also means the model needs access to an execution environment — browser, desktop, terminal — not just a text API.

Performance and limits, what was measured and what was not

The published numbers are infrastructure metrics (throughput, latency, uptime), not agentic capability benchmarks. No SWE-bench data, no computer-use task success rates, no comparison to Claude 3.5 Sonnet or GPT-4o with tool use. The 59.68% upstream error recovery is an infrastructure figure, not a measure of the quality of the model's own fixes. Anyone building on this for production needs to run their own evaluations on their specific tasks, and verify whether context-window limits and request filters allow the automatic rerouting to work when things break.

Bottom line

Nex-N2.5-Pro offers an interesting package: an agentic model with built-in computer-use capabilities, free pricing, and a large context window, all wrapped in a single API that also manages the visual feedback loop. The clear downside is the lack of transparency on the quality of the agency itself: no public benchmarks, no architecture details, and no SLA commitment beyond the infrastructure metrics. For teams looking to replace brittle automation scripts or accelerate manual QA, it is worth a try. For anyone who needs measurable guarantees, it is still too early.