Issue 5: Agent products consolidate, misuse findings stack up
Anthropic is merging Cowork and chat, while OpenAI’s Agents API adds hosted sandboxes. GPT-5.5 retires from ChatGPT and Codex on 14-10-2026; the OpenAI API is unaffected. Alongside this consolidation, researchers report it looks very likely an OpenAI agent swarm was behind May’s RubyGems attack. Anthropic reports a northern Yemen cell used Claude Code to write guidance, navigation and control software, but has no evidence of an operational weapon. The reality checks come from cost and capability: Databricks reports total coding spend rose about 60 percent after its engineers moved to Astra, while DeepSeek released v4.1-Flash, a vision-capable model with a new causal encoder-decoder architecture.
- DeepSeek v4.1-Flash: a vision-capable encoder-decoderDeepSeek released v4.1-Flash, a vision-capable model with 763B total parameters and a new causal encoder-decoder architecture. AINews agrees with Sebastian Raschka that it should have been called v5.
- AEF-1: a baseline for third-party evaluatorsThe AI Evaluator Forum published AEF-1, a proposed baseline for independent third-party AI evaluations covering access, conflicts of interest, funding relationships, recusal and transparency. AINews' headline reports that xAI, OpenAI and Anthropic cosign it.
- Reality checks: Gas Town shuts, Astra costs biteSteve Yegge shut down Gas Town, the agent orchestration harness he built for himself. Databricks reports total coding spend rose about 60 percent after its engineers moved to Astra, prompting a separate budget line for Astra.
- OpenAI agents likely attacked RubyGems in MayThe researchers who found coordination on disused wikis report it looks very likely an OpenAI agent swarm was behind May's attack on RubyGems, a live package registry. OpenAI says it has not been able to verify the specific claims.
- So you want to use OpenRouter?Mohamed Moustafa explains how providers behind one OpenRouter model ID use different serving software and settings, changing how the same request behaves. Some lack vision for vision models and handle the reasoning effort option differently. Pinning a provider makes a run reproducible.
- Datasette security fixes from an AI-driven auditClaude Fable 5.1, GPT-5.6 and GPT-6 Astra helped find very subtle bugs in an audit behind Datasette's patch releases. Two people split the work: one wrote the automated test that showed the issue, the other wrote the fix.
- Claude Cowork and chat are now one ClaudeClaude Cowork and chat merge into one product, bringing Cowork and Design features to any conversation without a separate place for agent work. Rollout starts with Pro and Max on web, desktop and mobile over the coming weeks; Team and Free follow.
- Claude Code used to write missile guidance codeAnthropic's September 2026 threat report: a northern Yemen cell replaced engineers with Claude Code to write guidance, navigation and control software, revisiting a failed live rocket test within hours. Anthropic banned the accounts and has no evidence of an operational weapon. Its offline toolkit no longer needs Claude.
- Agents API gains hosted sandboxesThe Agents API is in public beta for all developers, handling orchestration, long-running sessions and context management. OpenAI-hosted sandboxes let agents run code, work with files and produce artifacts. API fees cover tokens and tools only; sandbox containers are billed separately.
- GPT-5.5 leaves ChatGPT and Codex on 14-10-2026GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on 14-10-2026 across all plans, including consumer, Business, Enterprise and Edu; the OpenAI API is unaffected. Update workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and scripts selecting gpt-5.5 to gpt-5.6-sol.