LG Opens K-EXAONE 2.5 to Anyone With an API Key, Free, but Its Terms Stop at Noncommercial Use

LG AI Research opened K-EXAONE 2.5 to outside developers on Nov. 9, publishing it on the EXAONE API Platform with no charge for API calls. Developers, startups, enterprises and research institutions can request a key through docs.exaone.ai and use the model for service development and research, an availability claim rather than a roadmap item: the endpoint is live today.

The model itself is what LG calls a technical validation model, trained further on top of the version it released in August. LG positions it for coding, web retrieval and workflow automation, which places a Korean corporate lab in the same agentic lane that Chinese open-weight labs have been crowding since the summer.

LG published four internal benchmark scores, all computed on public datasets and none independently audited: BrowseComp 78.0 for navigating the web and finding information, Claw-Eval 77.3 for problem-solving in workplace environments, AutomationBench-Openset 74.6 for automating tasks through SaaS tools, and SWE-Bench Pro 59.2 for real-world software development and modification. LG says the results put it alongside frontier open-weight models and make it competitive with Chinese frontier models including DeepSeek-V4.1-Flash and GLM-5.3. The comparison targets have no scores in the post's text, LG published no sample counts, decoding settings or agent scaffolding, and agentic benchmark scores notoriously do not transfer across harnesses, so the numbers should be read as the company's own claim rather than a ranking.

The terms are where the free offer narrows. LG's platform conditions limit use to noncommercial purposes and prohibit embedding the API in products that let third parties call the model, which keeps the endpoint a development and evaluation channel rather than a production substitute. Rate limits and service-level commitments are not published, so a team building against it has to discover capacity by testing.

Free still moves the economics for a specific class of work. Subagent fan-out, browsing loops, bulk classification and first-pass code review generate large volumes of low-stakes calls where the marginal quality gain of a frontier model rarely justifies its marginal cost, and a hosted lane priced at zero changes that arithmetic immediately for exactly those calls. BrowseComp and AutomationBench are the two benchmarks that map most directly onto them, which is presumably why LG leads with both.

The costs move rather than disappear. Adopting the lane means an integration, an internal evaluation set, a failover path and a policy decision about sending prompts to a third-party platform owned by a conglomerate in another jurisdiction, and all of that work lands on the adopting team. LG's own wording, a technical validation model rather than a production flagship, is a fair warning about how much support to expect.

What it means: this is a distribution play, not a price war. LG is buying developer attention in the agentic segment with free tokens while keeping commercial use behind a gate it has not yet priced, a pattern that mirrors how Korean and Chinese labs have used open weights to get their models into Western toolchains. Whether any of it converts depends on two unglamorous things: published rate limits, and whether the SWE-Bench Pro number holds when someone else runs the harness.