Level on its own scoreboard

LG's announcement says 2.5 performs on par with DeepSeek V4.1 Flash and GLM 5.3, and its model card publishes the table behind that claim. On SWE-Bench Pro, the coding test, 2.5 scores 59.21, ahead of DeepSeek's 58.41 and behind GLM 5.3's 64.03. On BrowseComp it scores 78, ahead of DeepSeek's 75.8 and a point behind GLM's 79. On AutomationBench, in its open-set evaluation, it scores 74.6, just behind DeepSeek's 74.9 and ahead of GLM's 71.6. On Claw-Eval it scores 77.29, ahead of DeepSeek's 75.78 and fractionally behind GLM's 77.64.

Counted plainly from LG's reported scores: 2.5 leads DeepSeek on three of four benchmarks and GLM on one of four, and its clearest deficit is on the coding test, where GLM leads it by almost five points. Every number in the table is LG's own measurement on public datasets, run under LG's own conditions. The jump from the previous version is large on the same table, most starkly AutomationBench, 16.9 to 74.6. A gain that size is exactly the kind of figure that needs an outside run before it settles. The sources reviewed provide no independent evaluation.

Free, with the terms still blank

The documentation checked on 9 November 2026 describes access as currently free and specifies a 524,288-token ceiling per call, counted as input tokens plus max_tokens. Those pages give no free-access end date and no subsequent pricing, and the rate-limit page publishes no numbers — only the policy: limits per second, minute and day, reported in response headers. The endpoint reference documents an OpenAI-compatible chat API; the Models page lists tool calling and structured outputs among the model's features, although the endpoint reference documents no tool-calling parameters yet. Aju Press reports that LG plans to use NVIDIA's Dynamo inference software to improve inference efficiency and expand service scale.

There are also no weights for 2.5. LG has published weights for earlier K-EXAONE models on Hugging Face. This time access runs through LG's API, and the terms beyond "currently free of charge" are the ones LG has not published yet.

That should be read alongside how LG describes the model itself. LG calls 2.5 a technical validation model, further trained from the version released in August 2026, and Im Woo-hyung said the team is concentrating on a third-generation model aimed at frontier-level performance by January 2027 (Aju Press). A validation model offered free, with the successor already announced, is an invitation to test — on a timeline LG controls and has not shared.

Sources

  1. [1] LG AI Research, "K-EXAONE 2.5 API Available for Free!", Medium, 9 November 2026Read source
  2. [2] EXAONE API Platform documentation, Models page, checked 9 November 2026Read source
  3. [3] EXAONE API Platform documentation, Changelog, Rate Limits and K-EXAONE 2.5 endpoint reference, checked 9 November 2026Read source
  4. [4] Aju Press, "LG AI Research Opens Free Access to 'K-ExaOne 2.5' API", 9 November 2026Read source