DeepSeek V4 Benchmark Leak Fuels Talk of a New Coding King

Earlier today a screenshot circulating on X claimed DeepSeek V4 achieved an 83.7% score on the SWE‑Bench Verified coding test, allegedly surpassing Claude Opus 4.5 (80.9%) and GPT‑5.2 High (80.0%). The post also listed other leaked metrics such as ~90% on HumanEval, >80% across SWE‑Bench domains, 1M token context, and a cost 20‑40× lower than OpenAI.

Netizens quickly questioned the authenticity. The image showed an AIME 2026 score of 99.4%, a value mathematically impossible according to the official AIME scoring system. Epoch AI director Jaime Sevilla refuted the claim, stating the referenced FrontierMath Tier 4 scores are from a dataset only accessible to OpenAI and the DeepSeek team, and that DeepSeek V4 has not been evaluated.

“In the official scoring system, the highest AIME score is 119/120 (99.2%) or 120/120 (100%). 99.4% cannot occur.”

Despite the dubious leak, DeepSeek’s official app announced a gray‑scale test of a new long‑text model supporting a 1‑million‑token context, aligning with the “repo‑level reasoning” capability mentioned in the leaked data.

The Overchat.ai page, likely a pre‑release SEO placeholder, described three core features of the upcoming model:

Repo‑level Reasoning : understands entire project architecture, not just individual code files.

Engram Conditional Memory System : aims for near‑infinite context retrieval when handling 1M‑token inputs.

Mixture‑of‑Experts (MoE) architecture upgrade : builds on V3’s efficiency while specializing for coding tasks.

Both the leaked screenshot and Overchat’s description indicate a planned release around mid‑February 2026 (approximately February 17, coinciding with the Lunar New Year). The article concludes that the 83.7% SWE‑Bench figure is likely false, the 1M‑token model test is genuine, and DeepSeek is poised to launch a new model during the holiday period.