GPT-5.5 vs DeepSeek V4. While you were sleeping, the world turned upside down again

Yesterday OpenAI released GPT-5.5, and today the Chinese from DeepSeek responded with their V4.
I perfectly understand that you've already read 150 identical enthusiastic translated posts about GPT-5.5 in other channels, and maybe even learned about DeepSeek. Everyone has already discussed everything, liked the benchmarks, and forgotten.
But let me still highlight the most important points.

GPT-5.5
The key metric here is not MMLU (who cares about it in 2026?), but Expert-SWE (73.1%). This means the model can dig through your legacy code for 20 hours straight, debug, rewrite tests, and eventually produce a working PR, not a bunch of apologies. It consumes fewer tokens for the same tasks than 5.4. OpenAI realized that endlessly expanding context without control is a dead end.

"Losing access to GPT-5.5 feels like losing a limb" — quote from the release.

DeepSeek V4: Chinese crowbar against American sanctions 🇨🇳
While Sam Altman talks about AGI and sells expensive Enterprise subscriptions, the Chinese are open-sourcing a model with 1 Trillion parameters and a context of 1 million tokens.

Thanks to aggressive MoE (Mixture-of-Experts), out of this trillion, only ~49B parameters are activated during inference for the Pro version, and a mere 13B for the Flash version.

They don't have access to top-tier Nvidia chips, so they rely on pure mathematics. The Engram memory architecture and Manifold-Constrained Hyper-Connections (mHC) allow them to train and run this behemoth on local hardware like the Huawei Ascend 950PR.

As a result: the API costs pennies (from $0.03 to $0.30 per million input tokens). This is tens of times cheaper than Claude 4.7 and GPT-5.4.

Model on HF

Waiting for a release from Sber? 😁