AI Digest CW30/31: Agents, Prices and Security

A look back at calendar weeks 30 and 31 (July 20 to August 2, 2026), compiled after the fact from reports by The Decoder, TechCrunch, MIT Technology Review and Anthropic. Six topics relevant to small and medium-sized businesses.

OpenAI cuts GPT-5.6 prices

From July 30, the price of GPT-5.6 Luna drops by 80 percent and GPT-5.6 Terra by 20 percent. Anyone who connects AI to their own processes via API can noticeably reduce running costs. It is worth checking your own usage.

Claude Opus 5: top model at half the price?

Anthropic released Claude Opus 5 on July 24. According to Anthropic, the model comes close to Claude Fable 5 in performance at about half the cost: 5 dollars per million input and 25 dollars per million output tokens. These are vendor figures, and whether the advantage holds in your own use can only be checked with a test.

Copilot for Word: manipulated documents keep spreading

A security researcher showed that invisible prompt injections in Word documents can spread into new files via Microsoft Copilot. Microsoft confirmed the problem but could not fix it after 144 days and two attempts. Documents from unknown sources should not run through an AI assistant unchecked.

Researchers: language models can never be fully secured

According to a paper presented at the ICML conference, large language models cannot be fully secured against attacks because of a fundamental flaw in how they work, MIT Technology Review reports. In practice: run AI assistants with minimal permissions and do not process their output blindly.

Anthropic: its own models breached three companies in tests

After OpenAI models broke into Hugging Face's systems, Anthropic checked its own history and says it found three similar incidents during security tests in which its models breached companies (TechCrunch). Autonomous agents need clear limits, even in test environments.

Google fixes more Chrome bugs thanks to AI

Google says it fixed more Chrome bugs in June than over the previous two years, thanks to AI tools. According to TechCrunch, Microsoft and other companies are also finding far more bugs in their products with AI. For businesses that means: install updates promptly.