Fable 5.1's Cache Cut Changes the Economics of Agent Code Review

Anthropic shipped Claude Fable 5.1 on September 1, 2026, and the headline number is not a benchmark. Cache reads dropped 75% to $0.25 per million tokens. For agent-based code review, that single price change moves whole-repo context from an expensive edge case to the obvious default, and teams that treat this as a minor pricing update will find their review tooling competitively behind by the time they revisit the decision.

What the price change actually means for review loops#

Anthropic's launch post is clear about where the savings concentrate. Input and output token prices are unchanged at $10 and $50 per million tokens, respectively. The cut applies only to cache reads. For typical workloads, that produces about a 25% cost reduction. For context-heavy, tool-heavy agentic work, the reduction reaches approximately 45%, per Anthropic's own figures from four weeks of August 2026 usage data.

Code review is almost entirely cache reads in an agent loop. The repository state, the style guides, the dependency manifest, the security policy , those load once and persist. Every subsequent call reads from cache. The implication is direct: the cost floor for keeping an entire mid-sized repository warm across every PR just collapsed by roughly three-quarters on the cache portion of the bill. What was a deliberate architectural decision , whether to cache aggressively or re-tokenize per call , is now a straightforward answer.

The benchmark gap between Fable 5.1 and Mythos 5.1#

Fable 5.1 scores 73.4% on CursorBench 3.2.0, the highest recorded on that benchmark. On Terminal-Bench 4.0, Fable 5.1 reaches 55.8%. Mythos 5.1 reaches 60.9% on the same Terminal-Bench 4.0 evaluation. That 5.1-point gap is not a different model. It is the same weights with guardrails configured differently, and what Anthropic calls the gap between them is explicitly attributed to tasks where the earlier, less precise cyber safeguards intervened and produced a zero.

The practical read: Mythos 5.1's benchmark advantage comes from fewer false-positive blocks on security-adjacent coding tasks, not from superior reasoning. Fable 5.1 with updated safeguards blocks 60% fewer benign queries than Fable 5 did, but Mythos 5.1 blocks fewer still, because vetted partners operate under a different policy profile entirely.

Tiered access and what it means for security tooling#

Mythos 5.1 is not available generally. Access routes through Project Glasswing, the Cyber Verification Program, and the Life Sciences Verification Program, all currently restricted to a set of US organizations. Anthropic's Claude Security product, which scans codebases for vulnerabilities and suggests patches, is now powered by Mythos 5.1.

This creates a real split. Two organizations running otherwise identical code review pipelines, both using Claude, may be running models with materially different behavior on security-adjacent code analysis. Penetration testing, exploit generation, and binary-based vulnerability scanning are still redirected to Opus-class models even for Fable 5.1 users. Mythos 5.1 partners operate under different rules. Whether that gap is meaningful for a given team's review coverage depends on what their code touches , but it is a gap worth knowing exists.

What the bank CISO group built, and why it matters for regulated teams#

Eight members of the Analysis and Resilience Center for Systemic Risk, whose roster includes CISOs from Goldman Sachs, Morgan Stanley, Citi, Bank of America, and Wells Fargo, spent months with Anthropic working through a specific question: what would it take to run frontier models inside a systemically important bank? Scott DePasquale, the center's president and CEO, described the group as defining "who holds the data, who holds the keys, what automated review can and cannot see, and under what conditions a human is ever permitted to look."

The answer is Enterprise Frontier Safeguards. Misuse-detection data lives in the customer's AWS S3, Azure Blob, or Google Cloud Storage account, under the customer's encryption keys. Automated systems analyze traffic for signals including stolen credentials and attempts to develop offensive cyber capabilities. When a flag fires, the alert goes to the customer. Munish Kumar Sharma, CISO at Wells Fargo, is quoted directly: "our logs stay in a Wells-managed environment under Wells-managed keys."

The open question the dossier raises directly: Anthropic does not specify how long the rolling detection window runs. That omission matters if an attacker knows to space activity across sessions. For teams configuring SIEM integrations, the detection window length affects correlation logic, and the current answer from Anthropic is that the number is not published.

Anthropic developed EFS with more than 100 customers, covering a quarter of the Fortune 100 and every US global systemically important bank. It is rolling out in phases beginning later this fall. Until EFS is available, eligible customers can run Fable 5.1 with zero data retention.

The defender-side configuration checklist#

Four things to act on now, based on what the dossier confirms:

1. Cache budget review. If code review agent costs are currently modeled against Fable 5 cache read prices, the September 1 price change invalidates those models. At $0.25/MTok for cache reads, whole-repo context loading becomes the cheaper default path on most repository sizes.

2. Mythos 5.1 eligibility check. Teams doing defensive security work , vulnerability discovery, not exploit generation , should evaluate whether they qualify for the Cyber Verification Program. The 5.1-point Terminal-Bench gap between Fable and Mythos is specifically on security-adjacent coding tasks. The CVP is the current access path.

3. EFS readiness assessment. If the organization is in financial services, healthcare, telecom, law, or public sector, EFS will matter for compliance. The storage and key management model maps cleanly onto existing cloud infrastructure. The alert queue responsibility transfers to the internal team, which requires staffing that queue.

4. Detection window uncertainty. Until Anthropic publishes the length of the rolling detection window, treat it as an unknown in threat modeling. Attackers spacing activity to evade short windows is a documented pattern. Internal monitoring should not assume Anthropic's detection covers what it cannot confirm it covers.

The verification gap autonomous tools fill#

The bottleneck the cache price change removes is computational cost, not review quality. Keeping an entire repository warm in context makes whole-codebase analysis cheap. It does not make it accurate or safe to act on without verification. Hyrax runs 13 verification steps between a candidate fix and a submitted PR precisely because the cost of analysis falling does not reduce the cost of a wrong fix shipping. The dossier notes Fable 5.1 scored 73.4% on CursorBench 3.2.0. The 26.6% it does not solve needs a gate.

The code review problem described in the bottleneck post is unchanged by cheaper tokens: generating candidate changes became cheap, validating them did not. Fable 5.1 changes where the cost floor sits. It does not change what happens if verification is skipped.

Hyrax is live at hyrax.dev.