Claude Fable 5.1: Pricing, Benchmarks and Sri Lanka Impact

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026. They are the same underlying model shipped with different levels of safeguards: Fable 5.1 is generally available, and Mythos 5.1 is restricted to trusted-access programmes for cybersecurity and life-sciences work. For teams in Sri Lanka the interesting part is not the benchmark table — it is the pricing structure, and the fact that one of the two models is not available here at all.

The pricing change is in cache reads, not the headline rate

Per-token rates did not move. Fable 5.1 still costs $10 per million input tokens and $50 per million output tokens. What changed is the cache-read price, which dropped by 75% to $0.25 per million tokens. Anthropic puts the practical effect at roughly 25% cheaper for typical workloads and up to 45% cheaper for highly agentic work.

That distribution matters. A single question-and-answer call barely benefits, because it reads little from cache. A long-running agent that re-reads the same system prompt, tool definitions and file context on every turn benefits enormously, because cache reads dominate its bill. If your workload is a chatbot, expect a modest saving. If it is a coding or document-processing agent that loops, this is a material cut.

• Input: $10 per million tokens (unchanged from Fable 5).
• Output: $50 per million tokens (unchanged from Fable 5).
• Cache reads: $0.25 per million tokens — down 75%.
• Net effect: about 25% cheaper typical, up to 45% for agentic loops.

Where the capability gains actually landed

The published comparison against Fable 5 is uneven, and the shape of it tells you what the model was tuned for. The largest jumps are on long-horizon, tool-using, computer-operating tasks. General knowledge moved very little.

• Terminal-Bench-Science 0.1: 52.6%, up from 24.7% — more than doubled.
• AutomationBench: 31.4%, up from 17.1%.
• Terminal-Bench 4.0: 55.8%, up from 42.0%.
• OSWorld 2.0 (strict): 41.7%, up from 36.1%.
• Humanity's Last Exam (no tools): 60.9%, up from 57.8% — a small move.
• CursorBench 3.2.0: 73.4%, up from 70.5%.

Read that as a model built for agents rather than for answers. If you are using Claude to draft copy or answer questions, 5.1 will feel much like 5. If you are running it as an operator that executes multi-step work against real systems, the difference is the point of the release.

Fewer false refusals, which is a real operational change

Anthropic reports that Fable 5.1 declines 85% less often on benign requests about elementary biology and medical questions, and that cyber safeguards trigger around 60% fewer interventions per session than Fable 5. For anyone who has had a legitimate clinical or security question refused mid-workflow, that is the most immediately noticeable change. The model can also now identify software vulnerabilities for defensive work, and Anthropic says robustness against prompt injection improved.

One addition cuts the other way for some teams: new anti-distillation mechanisms prevent manual context editing across multi-turn conversations for new API accounts. If your integration rewrites conversation history between turns — a common pattern for trimming context by hand — test it before you migrate.

Mythos 5.1 is not available in Sri Lanka

Fable 5.1 is broadly available: Claude.ai, Claude Code, Claude Enterprise and the Claude Platform, plus Amazon Web Services, Google Cloud and Microsoft Azure. Mythos 5.1 is not. It is gated behind two verification programmes — a Cyber Verification Program for defensive security professionals and a Life Sciences Verification Program run in partnership with the US government — and access is currently limited to US organisations.

This is worth stating plainly because Sri Lankan teams have already been through one access disruption with this model family. Fable 5 was blocked and then restored in July 2026, and the pricing and access questions that followed were the substance of that story. The lesson repeats here: build so that the specific model is a configuration value, not an assumption baked through your codebase.

What a Sri Lankan team should actually do this month

1. If you run agentic workloads, re-model your monthly bill against the new $0.25 cache-read rate before assuming the 25% headline figure applies to you.
2. Verify your prompt cache is hitting at all — a broken cache turns the whole saving into zero.
3. If your integration edits conversation history between turns, test it against 5.1 on a new API account before migrating.
4. Do not design around Mythos 5.1. It is not available to Sri Lankan organisations today.
5. Keep the model ID in configuration, not scattered through your code, so the next availability change is a one-line edit.

Where Capricon fits

Capricon builds and maintains operational software for Sri Lankan businesses, and model choice is a cost line inside that work rather than a headline. If you are running AI inside ERP, customer operations or document processing and want the bill modelled honestly against your actual token pattern, see AI integration services or contact us for a scoped review.