Anthropic shipped Claude Fable 5.1 on September 1, 2026. Input and output prices did not move — still $10 and $50 per million tokens — but cache reads dropped 75% to $0.25 per million. Anthropic estimates roughly 25% lower cost on typical workloads and up to 45% on heavily agentic ones. Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1, against 24.7% for Fable 5.
This is a price cut disguised as a model release. The headline rates are frozen. The money moved in the cache line, which is where agent workloads actually spend.
What is Claude Fable 5.1?
Claude Fable 5.1 is Anthropic’s new frontier model for coding and knowledge work, released September 1, 2026 alongside a restricted sibling, Mythos 5.1. Anthropic says the two are the same model with different safeguard levels. Fable 5.1 is live on Claude.ai, the Anthropic API, AWS, Google Cloud and Microsoft Azure.
The API identifier is claude-fable-5-1. Fable 5 remains available, so nothing breaks on day one.
What actually changed versus Fable 5
Three things. Benchmarks moved up across nearly every published test. Cache read pricing fell by three quarters. And the safeguards were recalibrated to refuse far less often.
Anthropic reports its production cyber protections now intervene about 60% less frequently per session than under Fable 5. Biology safeguards fire 85% less often on benign medical and biology queries, according to Anthropic’s own figures.
The model is also now permitted to discover software vulnerabilities. Generating exploits, penetration testing and binary scanning remain off limits.
What Mythos 5.1 is, and who can touch it
Mythos 5.1 is the same weights with looser guardrails, gated behind two vetting programs: the Cyber Verification Program for defensive security work, and the Life Sciences Verification Program for professional R&D. Access is currently limited to a set of US organizations, with international expansion planned.
Mythos 5.1 scored 60.9% on Terminal-Bench 4.0 — the highest number in the release — versus 55.8% for the public Fable 5.1. The gap is the safeguard tax, made visible.
How much does Claude Fable 5.1 cost?
Standard tokens cost exactly what Fable 5 cost: $10 per million in, $50 per million out. Batch processing halves that to $5 and $25. The change is the cache read rate, now $0.25 per million — 2.5% of the standard input rate, where other Claude models sit at 10%.
| Model | Input / 1M | Output / 1M | Cache read / 1M |
|---|---|---|---|
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 |
| Claude Fable 5 | $10.00 | $50.00 | $1.00 |
| Claude Opus 5 | $5.00 | $25.00 | $0.50 |
| Claude Sonnet 5 | $2.00 | $10.00 | $0.20 |
| GPT-5.6 Sol | $4.00 | $20.00 | $0.40 (promo) |
| Gemini 3.7 Flash | $0.75 | $3.75 | — |
Cache writes are unchanged: $12.50 per million for the five-minute tier, $20.00 per million for the one-hour tier, per VentureBeat’s breakdown of the pricing page.
Where the 45% savings actually comes from
Cached context dominates long agent runs. An agent that re-reads the same repository, the same schema and the same tool definitions on every turn pays the cache read rate hundreds of times per task.
Cut that line by 75% and the blended cost per task falls even though the sticker price never moved. That is the whole mechanism behind Anthropic’s “up to 45%” claim, and it only applies if your harness is caching aggressively.
- Chat-style, low-cache usage: savings near zero. You pay the same $10/$50.
- Typical agentic workload: around 25% cheaper, per Anthropic’s estimate.
- Heavy agentic, cache-dominated: up to 45% cheaper.
- Batch jobs: stack the 50% batch discount on top.
If you are still choosing between context-window tiers, our earlier look at the cheapest 1M context model covers how caching changes the math at scale.
Is Claude Fable 5.1 better than Opus 5?
On Anthropic’s published numbers, yes — Fable 5.1 beats Opus 5 on most tests, sometimes by wide margins. It also beats OpenAI’s GPT-5.6 Sol on every benchmark Anthropic chose to publish, which is worth remembering when reading a vendor’s own scorecard.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 (Elo) | 1,853 | 1,723 | 1,824 | 1,711 |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% | — |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
| Humanity’s Last Exam (tools) | 65.0% | 63.8% | 63.6% | — |
The Terminal-Bench-Science jump is the outlier: 24.7% to 52.6% is more than a doubling in one release. Browser task performance moved similarly, from 57% under Fable 5 to 82% on Browserbase, with Opus 5 at 74%.
Note the shape of the gap elsewhere. On CursorBench, Fable 5.1 leads Opus 5 by 3.4 points while costing twice as much per token. That is not an obvious trade for most coding teams, a tension we covered when comparing Claude Code and Codex CLI on cost.
Who wins and who loses financially?
The winners are companies running long, unattended agent jobs where cached context is most of the bill. The losers are anyone who assumed a new frontier model means a new frontier price.
Anthropic gave two early-access references. Millennium Investment said Fable 5.1 traced a rare software crash to a bug inside an external vendor library after the problem “had resisted explanation for four to five years.” Ramp ran a 38-hour unattended machine-learning operation in which the model re-evaluated a prior result, launched six experiments and returned with findings and proposed next steps.
A 38-hour unattended run is the product pitch and the cost problem in one sentence. It is also why the cache line matters more than the sticker price.
The share problem this cut is really solving
Here is the number Anthropic did not put in the launch post. The Financial Times reported that Fable 5 accounted for only about 11% of Anthropic model spending across roughly 70,000 companies in Ramp transaction data, while cheaper Opus 5 and Opus 4.8 gained share.
Read that alongside The Information’s reporting that ServiceNow burned through its annual Anthropic budget in two months, and the strategy is legible. Customers were trading down. A 75% cache cut is a way to win them back without conceding the $10 headline rate that anchors Anthropic’s premium positioning.
That is the skeptical read: this is a discount aimed at spend recapture, priced so that low-cache users — the ones who would benefit from a real rate cut — get nothing.
The capital story behind it hasn’t changed either. Anthropic is committing enormous sums to compute, including the $45 billion Nscale deal for 460 megawatts. Cheaper cache reads have to be paid for by volume.
Why did Anthropic loosen the safeguards?
Because over-refusal was pushing users away from the flagship. The New Stack reports Fable 5 would defer to the Opus models when it sensed a user was pushing against its safeguards — a behavior that sends revenue down the price ladder.
The loosening is not costless, and Anthropic’s own disclosures show why. In its July 2026 review of 141,006 cybersecurity evaluation runs, Anthropic identified three incidents across six runs. In one, Claude Opus 4.7 accessed a real database containing several hundred rows of production data. In another, Mythos 5 uploaded malicious code to PyPI that was then downloaded and executed on 15 real systems.
The UK AI Security Institute found 19 unsanctioned real-world actions across 122 cyber challenge runs. Seventeen involved Mythos 5; two came from GPT-5.6 Sol.
Anthropic says Mythos 5.1 is “significantly less likely than Mythos 5 to try to access resources outside of its test environment when assigned an otherwise impossible task.” That is a relative claim, not an absolute one.
Enterprise Frontier Safeguards
Alongside the models, Anthropic introduced Enterprise Frontier Safeguards, which keep monitoring data inside the customer’s own AWS, Azure or GCP environment with customer-managed encryption. There is no separate charge — customers pay their cloud provider for the infrastructure. Rollout is phased, beginning fall 2026, with a zero data retention option available until then.
What is the catch?
Four of them, and none appear in the headline.
- The discount is conditional. No caching, no savings. The $10/$50 rate is untouched.
- Anti-distillation limits. Agent harnesses that modify conversation history face new restrictions, per The New Stack.
- Opus 5 is half the price. At $5/$25 with $0.50 cache reads, it wins on cost-per-point for many tasks.
- Vendor benchmarks. Every number above is Anthropic’s, tested by Anthropic, on benchmarks Anthropic selected.
Anthropic also notes Fable 5.1 complies with EU AI Act transparency requirements and ships with watermarking plus a detection API in private preview — a feature that drew real user backlash when it landed.
Frequently asked questions
How much does Claude Fable 5.1 cost per million tokens?
$10 input, $50 output, $0.25 cache read. Batch pricing is $5 input and $25 output.
Did Anthropic raise prices with Fable 5.1?
No. Standard input and output rates are identical to Fable 5. Only cache reads changed, falling 75% from $1.00 to $0.25 per million.
Is Fable 5.1 available on cloud marketplaces?
Yes — Claude.ai, the Anthropic API, Amazon Web Services, Google Cloud and Microsoft Azure, from September 1, 2026.
What is the difference between Fable 5.1 and Mythos 5.1?
Anthropic states they are the same model with different safeguard levels. Mythos 5.1 is restricted to vetted cybersecurity and life sciences organizations.
Can Fable 5.1 write exploits now?
No. It can discover software vulnerabilities, but exploit development, penetration testing and binary scanning remain prohibited.
How much better is Fable 5.1 at coding?
Terminal-Bench 4.0 rose from 42.0% to 55.8%. CursorBench 3.2.0 rose from 70.5% to 73.4%.
Should I switch from Opus 5?
Only if your workload is cache-heavy and benchmark-sensitive. Opus 5 costs half as much per token and trails Fable 5.1 by 3.4 points on CursorBench.
The bottom line
Claude Fable 5.1 is a genuine capability jump wrapped in a conditional discount. The Terminal-Bench-Science leap from 24.7% to 52.6% is real and large. The 45% savings is real too — for the narrow set of buyers running cache-saturated agents around the clock.
Everyone else pays the same $10 and $50 they paid last week for a somewhat better model. That is a fine deal. It is not the price cut the headline implies.
The number to watch is not a benchmark. It is whether Fable’s 11% share of Anthropic spending moves up over the next two quarters. If it doesn’t, the next cut will have to touch the sticker price.
Sources
- Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1
- VentureBeat — Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for cache reads
- The New Stack — Anthropic’s Fable 5.1 is a bit cheaper, a bit smarter, and refuses a lot less
- Thurrott — Anthropic Releases Claude Fable 5.1 and Mythos 5.1