Открытые против закрытых: GLM-5.2 и DeepSeek V4-Pro отстают уже на 4–7 месяцев

July 20, 2026, the UK AI Security Institute (AISI) published the first public measurement of the gap between open and closed models: in cybersecurity tasks it has narrowed to 4–7 months, down from 6–10 months for most of 2025.

How far behind are open models

The open models GLM-5.2 and DeepSeek V4-Pro scored in AISI's cyber tests at the level of closed models released 4–7 months earlier. This is the regulator's first attempt to measure the distance not by eye, but against a fixed set of tasks.

• GLM-5.2 (Zhipu, June 2026) matched Claude Opus 4.6 from Anthropic (February 2026) — a gap of about 4 months
• DeepSeek V4-Pro reached the level of Claude Opus 4.5 (November 2025)
• The measurement was conducted on a fixed set of AISI cybersecurity tasks
• A year earlier, for most of 2025, the gap held at 6–10 months
• On long, multi-step tasks the lag is more noticeable than on narrow cyber tests

"Recent open models GLM-5.2 and

DeepSeek V4-Pro perform on par with closed frontier models released 4–7 months earlier," — from the AI Security Institute's analytical report.

Why this matters for defense

The main shift isn't so much speed as price. According to AISI, a single CyberRange test run with a 100 million token quota costs about $85 on Claude Opus 4.5 or 4.6, about $46 on GLM-5.2, and just $1.19 on DeepSeek V4-Pro.

A difference of tens of times means that comparable cyber capabilities become available for a fraction of the cost — and end up in open models, which typically have fewer safeguards than closed ones.

"Advanced capabilities are reaching less-safeguarded open models faster than before," — the AI

Security Institute states.

Kimi K3 and Hassabis's plan

Jack Clark reinforces the same trend in the Import AI newsletter with two more signals. The first is the announcement of Kimi K3 from Moonshot AI: a model with 2.8 trillion parameters, whose weights and research paper the company promises to release in the coming weeks. According to Import AI's description, in a single 48-hour autonomous run K3 designed, optimized, and verified a chip, while also building the MiniTriton compiler.

The second signal is a regulatory proposal from Demis Hassabis, co-founder of DeepMind. He proposes creating an industry Standards Body modeled on the US financial regulator FINRA: companies voluntarily submit a new model for review 30 days before release, and over time the voluntary regime would be enshrined in law.

"The

Standards Body would be responsible for developing evaluation protocols and interacting with relevant federal agencies," — Demis Hassabis, co-founder of DeepMind, in a post on X.

What it means

Open models are no longer "a year behind." A gap of 4–7 months combined with inference costs tens of times lower shifts the balance: cutting-edge capabilities spread to less-safeguarded models faster, and that's exactly why proposals like Hassabis's industry regulator are now on the table.

Frequently asked questions

What is the gap between open and closed models?

It's the amount of time by which the best open models lag behind closed ones on specific tasks. According to AISI's measurement from July 20, 2026, in cybersecurity it stands at 4–7 months, down from 6–10 months a year earlier.

How much cheaper are open models?

In AISI's CyberRange test, a run on DeepSeek V4-Pro cost $1.19, on GLM-5.2 about $46, and on the closed Claude Opus 4.5 and 4.6 about $85 for the same 100 million token quota.

What is Demis Hassabis proposing?

Creating a Standards Body modeled on FINRA: voluntary submission of models for review 30 days before release, with these rules eventually being converted into law.