LG opens K-EXAONE 2.5 API for free, but teams should test before migrating
A free, compatible API makes K-EXAONE 2.5 worth a serious evaluation—not an unplanned production migration.
What opened
LG AI Research opened free access to K-EXAONE 2.5 on November 9. It uses an OpenAI-compatible chat interface, while the text-only model supports Korean and English, tool calling, structured outputs and a 512K context window.
LG reports competitive results against DeepSeek V4.1 Flash and GLM 5.3 across four public agentic benchmarks. Those figures are vendor-run measurements, however, not independent proof of production behaviour.
What remains unknown
The documentation says the service is currently free, but gives no end date, future price or numeric rate limits. LG describes the release as a technical-validation model and is already concentrating on a third-generation system targeted for completion in January 2027.
A team can therefore test quality without an API bill today, but it cannot yet model long-term cost, capacity or service stability with confidence.
What we would do
Virtual Arc would add K-EXAONE 2.5 as another provider, not another dependency. We would replay real tasks first, then shadow production traffic and compare completed outcomes, latency, failures and tool execution.
Migration becomes rational only if the model beats the incumbent on your workload and LG publishes terms under which that advantage can persist.
LG has not launched a production price war; it has opened a free test bench. That distinction matters. K-EXAONE 2.5 arrives through an OpenAI-compatible API with a 512K context window, tool calling and structured outputs, while LG reports competitive results across agentic coding, search and tool-use benchmarks. But LG also calls it a technical-validation model, publishes no end date for free access and provides no numeric rate limits. We would not move a production agent to it on headline economics or vendor-run benchmarks. We would place it behind an existing multi-provider gateway, replay a representative task set and shadow live traffic without affecting users. We would measure cost per completed task once pricing appears, alongside latency, retries, tool-call accuracy and failure recovery. If it wins on the work that matters, the OpenAI-compatible surface should reduce migration effort. Until then, free access is useful evidence-gathering capacity, not a foundation for a product roadmap.
Turn insight into a working system.
We build a provider-neutral evaluation and shadow-routing layer that runs K-EXAONE 2.5 beside your current model on real tasks without putting production traffic at risk. We add cost, latency, quality and failure tracing so a later switch is evidence-led. Talk to us if your team runs agentic or coding workflows this model could touch.