Anthropic just shipped Claude Code 2.1.170, and this release is not a routine patch. It does two things that matter: it gates access to Claude Fable 5, Anthropic's first Mythos-class model made available for general developer use, and it fixes a session-transcript regression introduced in 2.1.169 that silently stopped writing conversation history to disk. If your team is running Claude Code, update now. Both changes affect your daily workflow in ways that compound over weeks.

The official changelog is direct: update to 2.1.170 to get Claude Fable 5: Mythos-5, and to restore sessions that were not saving transcripts and not appearing in the session list. That's two separate deliveries bundled in one version bump:

The transcript fix is not glamorous, but its implications are significant. More on that below.

Anthropic introduced the Mythos classification as an internal capability tier, and Fable 5 is the first model in that class to reach a general availability release. The framing matters: "made safe for general use" is Anthropic's signal that Fable 5 cleared their Constitutional AI and safety review bars at a higher capability level than prior models, not that it was watered down to meet those bars. For engineering teams, the practical question is not the marketing tier but the coding performance. Early access reports from teams that piloted Fable 5 in closed preview point to meaningful gains in multi-file reasoning, architecture-level analysis, and long-context coherence on large codebases. Anthropic has positioned this as a model that is competitive against the best available from OpenAI and Google on developer tasks, not just on generic benchmarks. That claim needs independent verification. Your evaluation harness should be running Fable 5 against your own repositories within the next two weeks. Leaderboard scores on HumanEval or SWE-bench tell you relative ranking; they do not tell you whether the model handles your specific stack, your test coverage patterns, or your internal API surface correctly. Benchmark on what you actually ship.

The regression in 2.1.169 was a silent failure. Sessions appeared to run normally. Code got written. But the JSONL transcripts that Claude Code writes to `~/.claude/projects//` were not being created. No transcript means no session history in the UI, no replay, no audit trail. For individual developers, this is annoying. For engineering organizations, it is a governance problem. Here is why enterprise teams should care about JSONL transcript persistence:

Anthropic restoring this in 2.1.170 is the right call, but it surfaces a competitive axis that does not get enough attention: observability. GitHub Copilot, Cursor, and OpenAI's VS Code integrations compete heavily on feature velocity and model quality. The quieter competition is on whether enterprise buyers can audit, trace, and govern what their AI coding tools are actually doing. Reliable JSONL session logs are a meaningful advantage in that dimension.

Here is how the major AI coding tools stack up on the dimensions that matter most for engineering teams right now:

The transcript column is the tell. Most tools treat the conversation as ephemeral because they were designed around the IDE chat paradigm, where sessions are throwaway. Claude Code was designed as a project-aware coding companion, which means session persistence is architectural, not bolted on. That design decision pays off now that enterprise buyers are tightening AI governance requirements.

The deeper competitive move here is distribution strategy. Anthropic could have released Fable 5 exclusively via the API and let Cursor, Windsurf, and other IDE tools integrate it on their own timelines. Instead, they shipped it first inside Claude Code, their own developer tool. That sequence is not accidental. It lets Anthropic control the end-to-end experience for Fable 5's first impression with engineers, collecting telemetry on real-world coding tasks before third-party integrations surface the model with different prompting strategies and context handling.

This pressures GitHub Copilot and Cursor to respond, not just on model quality but on workflow integration depth. Model parity is achievable within months of any release; workflow and observability depth takes years to build.

The update path is straightforward. The strategic questions take more thought. Immediate actions:

This month:

Longer term:

The arrival of Mythos-class models in production developer tooling accelerates the shift toward smaller, higher-leverage engineering teams. A team of five engineers using Fable 5 in Claude Code can plausibly cover the surface area that required 15 engineers two years ago on well-scoped, greenfield work. That does not mean hiring five engineers instead of fifteen. It means those five engineers can take on a product scope that was previously out of reach for a team their size, which means ambitious engineering organizations expand their portfolio of products, not shrink their headcount.

The teams that fall behind will be the ones still running pre-AI workflows at pre-AI team sizes, paying the overhead of both while capturing the advantages of neither.

Claude Code 2.1.170 is a required update for any team running Claude Code in production workflows. The Fable 5 unlock alone justifies the update; the transcript fix makes it urgent. These are not independent improvements. They compound: a more capable model generating better code, with durable session history that makes that code traceable, reviewable, and governable. Anthropic's decision to ship Fable 5 through Claude Code first is a signal about where the AI coding tool market is heading. The competition is no longer just about which model scores highest on benchmarks. It is about which platform gives engineering teams the deepest integration between model capability, workflow tooling, and enterprise observability. Right now, Claude Code 2.1.170 is the most complete answer to that question available in the market. Update your teams. Benchmark the model. Build the governance workflows. The gap between teams that do this systematically and teams that do not is widening faster than most engineering leaders realize.