AI digest: 2026-10-21

Highlights

• Claude Code 2.1.305: upgrade if you use Bedrock, Vertex, or a gateway, because it fixes a 2.1.304 regression that broke every request there. It also adds sandbox.allowNestedUserNamespaces for Chromium and podman inside the Linux sandbox, and a /usage view where f applies the fixes it suggests.
• Dashboard Design skill: a new MIT skill and plugin for building and critiquing analytics dashboards. Install it with /plugin marketplace add abdiraimov-otabek/dashboard-design-skill.
• OpenAI 12-hour prompt cache: prompt_cache_options.ttl: "12h" keeps a prefix warm for a whole day, at 2.5x input to write. Gap filled: Ultrafast costs 6x Standard.

New sources

Agents and tooling

• Claude Code releases 2.1.305 and 2.1.306: gateway regression fix, sandbox user namespaces, /usage fixes, MCP allowlist change.
• Dashboard Design: a Claude Code skill for analytics dashboards: dashboard design skill installable as a plugin, from today's GitHub trends.
• Inference hooks (Claude docs): server-side allow/deny on every Claude Enterprise prompt is now GA; also API user profiles and org webhooks.

Frameworks and libraries

• Pydantic AI v2.58.0: run and request timeouts, task steering, ChatGPT sign-in provider.
• Ollama v0.40.7: template-source reporting, wider speculative decoding, pplx-decider model.

Models

• Prompt caching: 12-hour lifetime (OpenAI API docs): 12-hour cache on the Responses API, 2.5x write cost.
• Voxtral 3: Mistral's speech-to-text, text-to-speech, and voice-agent stack; licence not stated.

Topic pages updated

• tooling/coding-agents: Claude Code 2.1.305 and 2.1.306.
• tooling/mcp: allowedMcpServers change for managed servers.
• tooling/skills: Dashboard Design skill.
• practices/security: inference hooks GA, sandbox user namespaces.
• models/claude: inference hooks, user profiles, org webhooks.
• practices/cost-and-caching: 12-hour OpenAI cache. Ultrafast, Fast, Batch, and Flex prices added (gap fill; open question answered).
• models/gpt: Ultrafast prices, 12-hour cache, Agents API Ultrafast and sandbox skills.
• frameworks/landscape: Pydantic AI 2.58.0, Ollama 0.40.7.
• models/open-weight: Voxtral 3 and its licence question.