Anthropic Claude Fable 5.1: 75% Cheaper Cache Reads and 52.6% on Science Bench

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, three months after the Fable 5 line shipped in June. The two share one underlying model separated by different safeguard layers. Fable 5.1 is publicly available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.1 is restricted to vetted US organizations in Project Glasswing. The biggest commercial headline is a 75% cut to cache read pricing, from $1.00 to $0.25 per million tokens, while base input and output rates stay at $10 and $50 per million.

What happened

Detail	Value
Release date	September 1, 2026
Context window	1M tokens
Max output tokens	128K
Cache read price (new)	$0.25 per million tokens
Cache read price (old)	$1.00 per million tokens
Base input price	$10 per million tokens (unchanged)
Base output price	$50 per million tokens (unchanged)
Batch processing	$5 and $25 per million tokens
Terminal-Bench-Science 0.1 (Fable 5.1)	52.6%
Terminal-Bench-Science 0.1 (Fable 5)	24.7%
Terminal-Bench-Science 0.1 (Opus 5)	29.0%
Terminal-Bench-Science 0.1 (GPT-5.6 Sol)	22.4%
Availability (Fable 5.1)	Public via claude-fable-5-1 API
Availability (Mythos 5.1)	Restricted, Project Glasswing only

Both models run with adaptive thinking always on and share a 1M token context window with 128K max output. Fable 5.1 is the generally available version. Mythos 5.1, which sits behind stricter controls, is limited to vetted US organizations and is not accessible through standard API sign-up.

How do the benchmarks compare?

Terminal-Bench-Science 0.1 is an agentic scientific research benchmark where models run code, analyze data, and complete multi-step experiments autonomously. Fable 5.1 scores 52.6%, roughly double Fable 5’s 24.7% and well above Opus 5 at 29.0%. Anthropic notes a standard error of 3.5 to 4.5 percentage points per model, so the exact ranking matters less than the size of the gap.

Other benchmark results worth noting:

• Terminal-Bench 4.0: Fable 5.1 at 55.8%, Mythos 5.1 at 60.9%
• CursorBench 3.2.0: 73.4%
• Humanity’s Last Exam: 60.9% without tools, 65.0% with tools
• OSWorld 2.0 strict: 41.7%
• AutomationBench: 31.4%
• GDPval-AA v2: 1853

The 5.1-point gap between Fable 5.1 and Mythos 5.1 on Terminal-Bench 4.0 (55.8% vs 60.9%) represents the direct performance cost of the additional safeguard interventions applied to Mythos. Anthropic publishing that number is an unusually candid disclosure.

What the cost cut actually means

Cache reads let repeated context (system prompts, documents, long conversation history) load from a stored cache rather than re-processing each time, which is central to any agentic or document-heavy workflow. The new rate of $0.25 per million tokens is 0.025 times the base input price, compared with 0.1 times on every previous Claude model.

According to Anthropic, this translates to roughly 25% lower cost on typical workloads and up to 45% on context-heavy agentic ones. If you run AI integrations that pass large system prompts or reference documents on every call, the savings are real and immediate with no code changes required.

Three breaking changes your team needs to fix

Anthropic documents three API changes that will break existing agent pipelines:

1. Forced tool use is removed. Setting tool_choice to any or tool now returns a 400 error. Switch to auto with strict tool use or structured outputs instead.
2. Thinking blocks are model-bound. Fable 5.1 can read thinking blocks from earlier models, but no earlier model can read its own blocks back from Fable 5.1. Any router or fallback setup that switches down to a smaller model loses the reasoning chain silently.
3. Editing earlier conversation turns invalidates thinking blocks. Injecting or deleting per-turn reminders, or rebuilding the system or tools array mid-conversation, now errors. This check is enforced for accounts created on or after August 31, 2026. The recommended fix is turn-scoped system messages and server-side context editing.

For teams building workflow automations with multi-step Claude agents, all three of these deserve a deliberate audit before upgrading.

Known regressions to watch

Anthropic also documents real regressions, which is worth noting. Parallel tool calling is more variable: agent loops that previously batched several tool calls per turn may now issue one call at a time, which slows throughput. At low effort settings, the model answers from memory more often instead of running tools. It also prefers full-file rewrites over targeted edits, which matters for coding agents where diff size is a cost driver.

Safeguard and research highlights

Cyber safeguards now allow vulnerability discovery but block exploit development. According to Anthropic, this cut safety interventions in Claude Code by roughly 60% per session. Biology safeguards fire 85% less often on benign requests. Penetration testing, exploit generation, and binary-based vulnerability scanning still redirect to Opus.

On the research side, Mythos 5.1 designed protein binders with roughly a 50% hit rate across 12 targets, against an industry norm of 10 to 15%. Fable 5.1 built a Venus elevation map at 2 to 3 km resolution and sped up seven open-source genomics models by up to 2.5x using custom GPU kernels.

All output now carries a statistical text watermark and C2PA credentials on files. Content provenance is not opt-in.

Our take

The cache read price drop is the most immediately actionable change here. For any business already using Claude at scale with long system prompts or large document contexts, this is a cost reduction that lands on day one with no engineering work. That is the kind of update worth acting on fast.

The benchmark leap on Terminal-Bench-Science is impressive, but Anthropic’s own error bars (3.5 to 4.5 points) mean you should test on your actual task rather than assume the ranking holds. The regressions on parallel tool calling and whole-file rewrites are specific enough to matter in production pipelines. Any team running Claude agents should build a regression test suite before cutting over, not after.

The Fable vs. Mythos split is the most interesting structural choice. Publishing the performance delta between the two models (essentially the safeguard tax) sets a precedent we would like to see more labs follow. It gives buyers a real number to weigh against access restrictions. You can track further AI model developments across our AI news coverage.

If you want help auditing your Claude agent pipelines for the breaking changes or sizing the cache cost savings for your workload, get in touch with the Lumien team.

Source: Marktechpost