Claude Code 2.1.170 Makes Fable 5 the Default Question for Hard Agent Work
Claude Code 2.1.170 is a model release wearing the clothes of a CLI update. The changelog is short: Claude Fable 5 is now available in Claude Code, and a transcript-saving bug is fixed for sessions launched from VS Code’s integrated terminal or shells that inherited Claude Code environment variables. That is not a long list. It is still a consequential one, because it moves Anthropic’s newest generally available high-end model into the place where developers actually ask hard questions about messy repositories.
The release landed on June 9, 2026 at 17:23 UTC. During research, the anthropics/claude-code repository sat at roughly 131,326 stars, 21,279 forks, and 8,668 open issues, with ten binary/checksum assets already accumulating a few hundred downloads. The paired Anthropic announcement describes Fable 5 as a “Mythos-class model” made safe for general use, exceeding any model Anthropic has previously made broadly available. The model docs list claude-fable-5 with a 1 million-token context window, 128k max output, adaptive thinking always on, and pricing of $10 per million input tokens and $50 per million output tokens.
That combination changes how Claude Code should be used. Fable 5 is not just “the new default, please enjoy burning money faster.” It is a more expensive capability tier for work where model judgment matters more than raw edit speed: migrations across a large codebase, debugging that requires reading many files, architecture review, dense spec interpretation, security-sensitive refactors, and the backlog item everyone keeps postponing because no human wants to build a complete mental model of the system first.
The model is the feature, but routing is the product
Anthropic’s launch claims are not subtle. The company says Fable 5 is available on Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. It also says Mythos 5 remains limited to approved Project Glasswing customers, while Fable is the version brought into general availability with safeguards. Those safeguards matter: when classifiers detect certain cybersecurity, biology/chemistry, or distillation requests, Fable 5 falls back to Claude Opus 4.8 instead of answering directly. Anthropic says the fallback triggers in less than 5% of sessions on average and users are informed.
That is a reasonable safety architecture, but it makes model routing part of the developer experience. If you are doing legitimate defensive security work, exploitability analysis, reverse engineering, or secure-code review, you may be in the exact zone where a fallback appears. Teams should treat that as observable runtime state, not invisible policy magic. Log when fallbacks happen. Compare output quality. Decide which tasks still work under Opus 4.8 and which need a different workflow, approval path, or human review. The worst version of this future is a CI job or security automation assuming it received Fable-level analysis when the runtime silently changed capability underneath it.
The community reaction was unusually strong. The Hacker News thread around Fable 5 had roughly 1,493 points and 1,168 comments during research, which is not normal release-note gravity. Simon Willison reported using Fable 5 in Claude Code and Claude.ai on difficult tasks he had postponed, including turning a MicroPython WASM library experiment into a full CPython WASM wheel. He also estimated he had spent $82.92 in Fable-priced API tokens while still inside his Claude Max allowance. That is the whole story in miniature: the model may unlock work that was previously not worth attempting, but the economics are no longer background noise.
Other practitioners described solving reverse-engineering tasks that Opus 4.8 and GPT-5.5 Codex had not cracked, reviewing dense 50-page specs, and seeing more surgical diffs. The skeptical responses were healthy: concerns about aggressive classifiers downgrading benign work, influencer-style early access, IP and open-source provenance, and whether generated interfaces are actually production quality rather than nicely arranged screenshots. That skepticism is not anti-AI reflex. It is what mature adoption looks like when the demos are impressive enough to deserve real scrutiny.
Use Fable like a senior reviewer, not a prettier autocomplete
The practitioner move is selective routing. Keep cheaper models for normal edits, test generation, formatting, small bug fixes, and boilerplate. Use Fable when the cost of a wrong answer, incomplete context, or weak judgment exceeds the cost of the model. A senior engineer does not review every typo fix; they review the migration plan, the security boundary, the unclear architecture decision, and the patch that touches five subsystems. Fable should be treated the same way.
That means teams need run labels and cost telemetry. “Claude worked on this” is not enough. Was it Sonnet, Opus, or Fable? Did it route through a fallback? How many tokens did the session consume? Was the task launched interactively, through a background agent, or through CI? Without those answers, model upgrades become vibes with invoices attached.
The transcript fix in 2.1.170 deserves more credit than it will get. Claude Code sessions launched from the VS Code integrated terminal, or shells inheriting Claude Code environment variables, could fail to save transcripts and therefore disappear from --resume. For long-horizon agent work, state continuity is not a convenience. It is the workflow. If the transcript disappears, the agent loses memory, the developer loses auditability, and any post-hoc review becomes guesswork. A model with a 1 million-token context window is only useful if the runtime can preserve the work around it.
Upgrade to Claude Code 2.1.170 if you want Fable 5 access or if your team launches sessions through VS Code terminals. Then do the less glamorous work: reserve Fable for high-ambiguity tasks, track usage separately, test safeguard fallbacks on security-adjacent workflows, and verify that --resume sees new sessions after the upgrade. The release is exciting, but the winning teams will not be the ones that simply point every prompt at the newest model. They will be the ones that know when the expensive reviewer belongs in the room.
Sources: Anthropic Claude Code v2.1.170 GitHub release, Anthropic Claude Fable 5 and Claude Mythos 5 announcement, Claude models overview, Hacker News discussion