The Pulse – what changed in AI · 25.10.2026
What changed for people who build with AI, Thursday 22 – Sunday 25 October 2026.
Fast mode for Claude Opus 5.5 is open to all paid API tiers, now also on Google Cloud
models · verified · 2026-10-22
Add speed: "fast" to a request and the same model generates up to 2.5× more output tokens per second, at $8 / $40 per million input/output tokens, with its own rate limit and no Batch API support. Since October 22 it is also part of the Anthropic-managed Claude Platform on Google Cloud, billed on your Google Cloud invoice.
Source: Claude Platform is now available on Google Cloud Docs: Fast mode - Claude Platform Docs
Anthropic: try one agent loop with skills before splitting work into subagents
prompts · verified · 2026-10-22
Anthropic says that in enterprise deployments a single agent with skills and tools often beat subagent designs on quality, usually at lower cost and latency; one customer's multi-agent harness ran about 2× slower. Keep a subagent only if the folded version scores worse on your evals, run each eval task more than once, and use the open single-agent-loop-kit to test the change.
Source: Building agents: When to fold subagents into your main agent loop Kit: anthropics/single-agent-loop-kit
OpenAI's Decisions API is generally available
models · verified · 2026-10-22
POST /v1/decisions returns typed answers (yes/no, a choice or a score) about text, images or JSON instead of a full response. It runs only on gpt-6-luna, costs $0.10 per million uncached input tokens with no charge for output, and now takes up to 400 questions per request (200 in beta); OpenAI says it is about 10× faster than the Responses API.
Source: OpenAI API docs: Decisions
Claude Code 2.1.309: /rewind no longer undoes sed -i edits
harnesses · verified · 2026-10-22
sed -i now gets the standard Bash permission prompt and, like other Bash edits, is not restored by /rewind, so commit before you let the agent run it. In bypass-permissions mode in the Desktop app and VS Code, dangerous rm and rmdir commands are now denied outright, and claude ultrareview requires workspace trust (scripts pass --trust-workspace).
Source: Claude Code changelog: 2.1.309
New to building with AI? Start with the free first lesson: https://www.vibe-coding.academy/start
Sources
Auto-generated by Vibe Coding Academy on October 25, 2026, grounded in the real sources linked above. We review for accuracy, but please verify time-sensitive details against the primary sources.
Build Blueprint · Builder
Have an idea? Get the spec your AI agent can build from.
Describe any product and get a complete build blueprint — stack, data model, screens, APIs, and a ready-to-paste prompt for Claude Code or Cursor. Export to PDF.
Open the Blueprint