One week, one task: can AI actually finish it?
A hypothetical laptop purchase connects the key developments from 5–11 October: understandable answers, multi-step work, suitable models and checkable results. This retrospective was produced on 13 October.
ChatGPT can answer with diagrams and tools
On 7 October, OpenAI began rolling out GPT-6 with Intelligent UI in ChatGPT. Answers can combine text with diagrams, buttons or small calculators. The update concerns Chat; it does not change the models behind Work or Codex. Rollout starts with Plus, Pro, Business and Enterprise. Free and Go are scheduled to follow the next day. Workplace settings may affect access.
Our interpretation: Changing an example can make explanations easier to understand.
Contract workflows become an agent evaluation
OpenAI and Ironclad presented research tasks for contracting software. An agent might configure approval rules and test different requests. Scoring measures specific criteria. A higher average score does not mean error-free contract work.
Our interpretation: The complete process matters more than individual clicks.
Claude Haiku 5.5 targets frequent, short tasks
Anthropic released Claude Haiku 5.5 for fast, repeated tasks such as summaries and request classification. The company reports lower costs than Haiku 4.5. Actual improvements depend on the task and settings.
Our interpretation: Small tasks do not always need the largest model.
Mistral Large 4 enters public preview
On 6 October, Mistral released an API preview of Large 4. The model accepts different kinds of input, including text and images. Model weights were announced for the end of October; they had not been released with this announcement. Mistral reports progress in coding and business tasks. These are vendor claims, not Applied by Hand test results.
Our interpretation: An API preview differs from a model you can run yourself.
Google introduces a continuous workplace agent
On 8 October, Google introduced the Gemini agent. It is designed to answer questions, create documents, run code and work with connected business systems. Google says cloud tasks can continue after a laptop closes. The announcement describes shared memory, permission controls and spending limits. It does not guarantee that every multi-step assignment will be completed correctly.
Our interpretation: Access permissions and checking results become especially important.
Anthropic launches cyber defence programmes
Anthropic announced its Cyber Mission to support critical infrastructure and open-source security. The new OSS Scanner offers open-source projects regular, free security scans. Findings still need verification, prioritisation and fixes.
Our interpretation: Finding a flaw is only the start of fixing it.
AI text gets an invisible provenance signal
On 5 October, OpenAI described a text watermark: a subtle statistical pattern in word choices that may indicate its origin. API customers can opt in for selected models. A gradual rollout is planned for eligible ChatGPT and Codex text in the EU. Short, edited or translated text can be harder to detect. Detection does not establish accuracy or identify the author.
Our interpretation: A provenance signal cannot replace checking the claims.
NemoClaw strengthens recovery of agent environments
On10 October, NVIDIA announced NemoClaw v0.0.132. It helps run agents in isolated environments called sandboxes. The update improves recovery and checks before rebuilding. It more precisely validates ownership of environments and images. These checks help protect shared resources during cleanup. The release describes safeguards, not a guarantee of error-free agents.
Our interpretation: Recovery and cleanup also need checkable boundaries.