Three releases this week share a quieter kind of news than usual. A model got faster and cheaper. An agent got a new way to call its own tools. And a toolkit discovered that one workflow setting means three different command-line flags depending on which coding agent runs it. None of the three is a capability jump. All three are about the exact wiring underneath, which is where the real work has moved.
What happened
Claude Fable 5.2 and Mythos 5.2 shipped on November 9, with the small model getting cheaper and the specialist one getting harder to reach. Anthropic’s announcement prices Fable 5.2 at $8 per million input tokens and $40 per million output tokens ($16/$80 in fast mode), a cut from Fable 5.1. For Mythos 5.2, the number that matters is security, not price. In Anthropic’s internal CVE Rediscovery evaluation, it “found 10.2% more vulnerabilities than Claude Opus 5.5 and 14.4% more than Claude Mythos 5.1.” Access to that capability is gated, not open: cybersecurity professionals apply through the Cyber Verification Program, and biomedical researchers go through a separate Life Sciences Verification Program. The same day, Claude Code v2.1.321 made Fable 5.2 its default Fable model, with 1M context and $0.20 per million tokens for cache reads.
Claude Code added Code Mode, an experimental switch that changes how an agent calls its own tools. Version 2.1.323, shipped November 12, lets Claude “write JavaScript to call its tools” instead of invoking them one function call at a time. It is opt-in, and an organization can turn it off. This is harness engineering made visible: the tool-calling loop itself is now something a team configures, not a fixed property of the agent.
Spec Kit’s v1.4.0 translated one workflow setting into three unrelated command-line flags, and quietly admitted one of them has no equivalent. The release maps a single reasoning_effort and context_size setting to --reasoning-effort/--context in Copilot CLI, to -c model_reasoning_effort=/-c model_context_window= in Codex CLI, and in Claude Code to --effort alone. Context size has nothing to translate to there. Per the pull request, “the step fails with a hint to choose a long-context model” instead. It’s a small admission with a real point: routing a task to the right engine means routing around flags that don’t exist yet, not just around prices and benchmarks.
How to read it
None of these three is a reason to switch anything today. Together they point at where the actual engineering happened this week: not in what a model can do, but in the layer that decides how it is told to do it, which model answers which call, and what happens when a setting one CLI understands has no meaning in another. That layer rarely gets its own announcement. This week it got three.
What to watch
Fast mode for Opus 4.8 is deprecated as of October 30 and is removed on November 30, 2026, after which requests return an error instead of falling back to standard Opus 4.8. Sixteen days out, it is still the clearest deadline on the calendar.