Claude Fable 5.1 Pricing: What Changes for Your AI Budget

Claude Fable 5.1 pricing is the detail worth reading past the benchmark headlines: Anthropic's September 2026 release cuts cache costs 75% and typical workload costs by roughly a quarter, while pushing agentic coding accuracy from 42% to as high as 60.9%. If your product runs on Claude — or you're deciding whether it should — the pricing shift matters more than the leaderboard jump.

What launched: Claude Fable 5.1 and Claude Mythos 5.1

Anthropic shipped two models built on the same underlying weights, differing only in safeguard configuration, according to the official announcement. Claude Fable 5.1 is generally available across the API, AWS, Google Cloud, and Microsoft Azure. Claude Mythos 5.1 runs with a different safeguard profile and is currently restricted to vetted organizations through Anthropic's Cyber Verification and Life Sciences Verification programs — access aimed at cybersecurity and life-sciences research, not general product use.

For anyone building on Claude, Fable 5.1 is the one that matters.

Claude Fable 5.1 pricing: what actually got cheaper

The headline numbers, per Anthropic's announcement:

• Input tokens: $10 per million.
• Output tokens: $50 per million.
• Cache reads: cut 75%, down to $0.25 per million tokens.
• Typical workloads: roughly 25% cheaper overall.
• Agentic tasks (the ones that read and re-read a lot of context): up to 45% cheaper.

That cache-read cut is the number worth sitting with. Agentic features — the ones re-reading the same tool definitions and context on every step — are exactly where cache pricing dominates the bill. A 75% cut there changes the math on features that were previously too expensive to run at scale.

The capability jump behind the price cut

Cheaper alone wouldn't be news. Anthropic's benchmark numbers show Fable 5.1 also getting substantially more capable than Fable 5:

• Agentic coding: 42.0% → 55.8–60.9%.
• Agentic scientific research: 24.7% → 52.6%.
• Computer use (OSWorld 2.0, partial credit): 72.9% → 77.9%.

Anthropic also reports the safety classifiers got quieter without getting looser: roughly 85% fewer false positives on elementary biology topics and about 60% fewer cyber-related refusals per session. For teams that had tuned prompts or added workarounds around overly cautious refusals, that's worth re-testing.

What this means if you're running (or scoping) an AI feature

Three practical takeaways for engineering and product teams:

1. Re-run your cost model before assuming last quarter's numbers hold. A 75% cache-read cut changes which architecture is the cheap one — a design that avoided caching to save money six months ago may now be the expensive path.
2. Re-evaluate model choice for agentic features specifically, not the whole stack. The biggest capability gains are in agentic coding and tool use — if you shipped a narrower model there to control cost, that trade-off may no longer hold.
3. Get the number, not the vibe. Before committing a roadmap slot to "we should upgrade," run your actual token volumes through a calculator — LLMScout's Claude cost calculator is a fast way to compare your current spend against the new pricing (Fable 5.1 isn't in its database yet, but Opus 5, Sonnet 5, and Haiku 4.5 pricing there is a useful baseline).

Where this fits in a real engagement

Model pricing changes every few months; the architecture decisions that determine whether a cost cut actually reaches your budget — caching strategy, context management, which features route to which model — don't change nearly as often, and that's where the real savings live. We treat that evaluation as part of scoping any applied AI engagement, not a follow-up conversation after launch.

If a pricing shift like this is reason enough to revisit your AI cost architecture, talk to our team — it's a short conversation, and cheaper to have before the next invoice than after.