On September 1, 2026, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. It has been roughly three months since Claude Fable 5 launched, and this time there is no government directive and no surprise shutdown in the story.
The headline is different: the sticker price did not move, but the cost of actually running the model dropped sharply. On top of that, Fable 5.1 beats last month’s Claude Opus 5 on every benchmark Anthropic published. Let’s look at the details.
What Is Claude Fable 5.1?
Fable 5.1 is Anthropic’s most capable generally available model. It sits in the “Mythos class” one tier above the Opus family and is called with the model id claude-fable-5-1.
As with Fable 5, the same brain ships in two packages:
- Claude Fable 5.1 (
claude-fable-5-1): the version everyone can use, with safety classifiers in place. - Claude Mythos 5.1 (
claude-mythos-5-1): the identical model with safeguards tuned for cybersecurity and life-sciences work. Available only to verified organizations under Project Glasswing.
In Anthropic’s own words, the gap between the two is not capability. It is where the safety filters step in.
Benchmark Results
Anthropic’s table compares Fable 5.1 against its predecessor, Opus 5, and OpenAI’s GPT-5.6 Sol. The jump is clearest on agentic tests:
| 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% |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
| OSWorld 2.0 (partial) | 77.9% | 72.9% | 75.4% | no data |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% | no data |
| Humanity's Last Exam (no tools) | 60.9% | 57.8% | 56.6% | no data |
| GDPval-AA v2 (score) | 1853 | 1723 | 1824 | 1711 |
The standout row is Terminal-Bench-Science. On this measure of agentic research work, Fable 5.1 more than doubles its predecessor: 24.7% to 52.6%. On Terminal-Bench 4.0, which measures agentic coding, the gain is over 30%. Mythos 5.1 goes higher still at 60.9%, because some of the cyber filters that fire in the Fable build are not present there.
Knowledge work follows the same pattern. On GDPval-AA v2, Fable 5.1 scores 1853, ahead of both Opus 5 (1824) and Fable 5 (1723). The gains are not limited to code.
Real-World Results
Beyond synthetic tests, the early-customer results Anthropic shared are more concrete:
- Millennium: the investment firm had a rare crash in its internal systems that engineers had failed to explain for years, and no model had cracked either. Fable 5.1 found the cause.
- Browserbase: completed 82% of the hardest browser-agent tasks. Opus 5 managed 74% on the same set, Fable 5 only 57%.
- MongoDB: finished a complex prototype in three days, running the optimization work unattended.
- Canva: built a rhythm game with real music and on-beat gameplay.
- Jane Street: reported that it solves more coding problems and set a new best on trading-intuition evaluations.
What these have in common is telling: they are all long, multi-step, unsupervised jobs. That is exactly where Fable 5.1 stakes its claim, not on answering one question well but on finishing a task that runs for hours.
Pricing: Same Sticker, Different Bill
This is the most practical change in the release. Input and output prices are identical to Fable 5, but cache reads dropped to a quarter of what they were:
| Item | Fable 5.1 | Fable 5 | Change |
|---|---|---|---|
| Input (1M tokens) | $10.00 | $10.00 | Unchanged |
| Output (1M tokens) | $50.00 | $50.00 | Unchanged |
| Cache read (1M tokens) | $0.25 | $1.00 | 75% cheaper |
A cached token now costs just 2.5% of the normal input price. Anthropic estimates the total drops around 25% for typical workloads and by up to 45% for heavily agentic work.
Why does one line item matter that much? In a long agent session the same system prompt, the same tool definitions and a growing conversation history get resent on every turn. That repeated prefix is served from cache, and across a 50-turn task it accounts for the bulk of the spend. Cutting it by 75% lands straight on the invoice.
If you want the numbers for your own scenario, I added Fable 5.1 to the LLM cost calculator . Enter your token counts and compare it against Opus 5 and the rest.
For Developers: What Changed in the API
If you are moving from Fable 5 to Fable 5.1, there are three breaking changes. None of them fail quietly, they return errors:
- Forced tool use is gone.
tool_choice: {"type": "any"}and{"type": "tool", "name": ...}now return a 400. Useautoplus an explicit instruction naming the tool, and addstrict: trueto the tool definition to keep arguments schema-valid. If you only forced a call to get JSON back,output_config.format(structured outputs) is the right tool. - Thinking blocks are bound to the model that produced them. Replay them to a different model and they are silently dropped (and not billed). On the same model, pass them back unchanged.
- Editing history invalidates thinking. This control is called “preserved thinking”: rewriting earlier turns while keeping the thinking blocks now errors. Accounts created on or after August 31, 2026 get a 400 outright. In short, build your harness so history is append-only.
There are welcome additions too: per-message effort (changed without resetting the cache), single-turn clear_at: "next_user_message" system messages, and a thinking.display: "updates" mode that returns short progress notes between tool calls.
A basic call looks like this:
import anthropic
client = anthropic.Anthropic()
response = client.beta.messages.create(
model="claude-fable-5-1",
max_tokens=16000,
# Thinking is always on; effort controls the depth
output_config={"effort": "xhigh"},
# Server-side fallback when a request is declined
betas=["server-side-fallback-2026-07-01"],
fallbacks="default",
messages=[{"role": "user", "content": "Find the root cause of the bug in this codebase..."}],
)
print(response.stop_reason)
print(response.content[-1].text)
thinking: {"type": "disabled"} or the old budget_tokens field returns a 400. Control the depth with effort instead: low, medium, high, xhigh and max. For most coding work xhigh is the sweet spot.Two more details worth knowing: Fable 5.1 is not offered on Priority Tier, and organizations on zero data retention cannot use it without explicit authorization from Anthropic, otherwise requests come back as a 400.
A Concrete Calculation: What the Cache Discount Actually Saves
Percentages stay abstract, so let’s walk through a realistic agent session. Say a long coding task sends 20M input tokens in total and 95% of that is served from cache (a perfectly normal ratio for long sessions, since the system prompt and conversation history are resent every turn). Output is 200K tokens.
| Item | Fable 5.1 | Fable 5 |
|---|---|---|
| Uncached input (1M tokens) | $10.00 | $10.00 |
| Cached input (19M tokens) | $4.75 | $19.00 |
| Output (200K tokens) | $10.00 | $10.00 |
| Total | $24.75 | $39.00 |
Same job, same model power, 36% cheaper. Note that the entire saving comes from one line. The longer the session runs, the larger the cached share becomes and the bigger the discount gets. For short, one-shot usage you will feel almost no difference, because caching never really kicks in there.
Fable 5.1 or Opus 5?
This will be the most common question. The price gap is exactly 2x: Opus 5 runs at $5 input and $25 output per million tokens, Fable 5.1 at $10 and $50. A simple split:
- Where Opus 5 is enough: chat interfaces, classification, summarization, short code fixes, high-volume and latency-sensitive routes. On this work, Fable 5.1’s extra capability does not show up in the result as much as it shows up on the bill.
- Where Fable 5.1 earns its price: unsupervised agent runs that go on for hours, root-cause analysis, large migrations, research and scientific reasoning. The 2x gap on Terminal-Bench-Science lives exactly here.
A practical suggestion: keep Opus 5 as the default and route only the hard tasks you get stuck on to Fable 5.1. One caveat, though: caches are model-scoped. An architecture that bounces between two models forfeits part of the caching advantage from the start.
What to Watch When Writing Prompts
Switching to Fable 5.1 is not just a model-string change. Three behavioral shifts stand out in Anthropic’s own notes:
- Turns really can run long. On a hard task, a single request may take many minutes. Raise your timeouts, use streaming and show progress to the user. The
thinking.display: "updates"mode exists for exactly this. - Over-prescriptive prompts hurt quality. Long prompts written for older models that spell out every step backfire on Fable 5.1. State the task and the success criteria clearly, then leave the method to the model.
- The writing style changed. The model now uses fewer bullet points and less bold text, leaning toward flowing paragraphs. If any system of yours parses its output by formatting, check it.
Also think about effort per task. For routine work low or medium is usually enough and noticeably cheaper. For coding and long-horizon agentic runs, xhigh is a good default.
Safety: Fewer False Alarms, More Permission
The loudest complaint about Fable 5 was safety filters firing on harmless requests. Anthropic reports concrete improvements here:
- 60% fewer false positives in cybersecurity. The model is now also allowed to identify software vulnerabilities. Writing working exploit code is still off the table, but this is a real opening for defensive work.
- 85% fewer false alarms on biology and medical questions. A student asking a basic biology question no longer gets refused for nothing.
- Enterprise monitoring data can stay in-house. With Enterprise Frontier Safeguards, audit data can live in the customer’s own cloud under their own encryption keys.
Watermarking: An Invisible Signature in the Text
One more notable piece arrives with Fable 5.1. To comply with the EU AI Act, text produced by models released after August 2, 2026 carries an invisible watermark. Anthropic is also offering a detection API in private preview for regulators, law enforcement, media organizations, fact-checkers, researchers and educators.
I covered how the watermark works, what happens when you copy and paste the text, and where its limits are in Claude’s text watermark .
Where You Can Use It
As of launch day, Fable 5.1 is available on:
- Claude API (
claude-fable-5-1) - Claude.ai and the mobile apps
- Claude Code
- Claude Enterprise
- Amazon Bedrock, Google Cloud and Microsoft Azure
Mythos 5.1 is limited to verified US cybersecurity and life-sciences organizations.
Frequently Asked Questions (FAQ)
Q: How much does Claude Fable 5.1 cost? A: $10 per million input tokens and $50 per million output tokens. Cache reads are $0.25 per million, which is 75% cheaper than Fable 5.
Q: What is the difference between Fable 5.1 and Opus 5? A: Fable 5.1 is stronger but twice the price. Opus 5 runs at $5 input and $25 output per million tokens. On Terminal-Bench 4.0, Fable 5.1 scores 55.8% against Opus 5’s 52.3%. For routine work Opus 5 is still a sensible choice.
Q: Are Fable 5.1 and Mythos 5.1 the same model? A: Yes, the same model. Only the level of safeguards differs. Mythos 5.1 is restricted to approved organizations under Project Glasswing.
Q: What do I need to change in my code? A: Drop forced tool use (tool_choice: any or tool), make your conversation history append-only, and set thinking depth with effort.
Q: Can the watermark Fable 5.1 adds be removed? A: Anthropic offers no way to strip it, and the whole point is to keep the source of the text detectable. Rewriting the text from scratch may weaken the signal, but that counts as a deliberate circumvention attempt and runs against the EU AI Act’s transparency clause.
Q: Is Fable 5.1 available in Claude Code? A: Yes, it shipped on launch day across Claude Code, Claude.ai and Claude Enterprise.
Conclusion
Claude Fable 5.1 is a slightly unusual model announcement. The scores went up, of course, but the real shift is in the pricing architecture. A 75% cut to cache reads is what finally makes long agent sessions affordable, and Anthropic’s framing points the same way: not one-shot chats, but unsupervised jobs that run for hours.
The loosened safety filters matter just as much. Cutting false alarms by 60% to 85% is an improvement anyone using the model daily will feel directly.
So what do you think? Is Fable 5.1 worth twice the price, or is Opus 5 enough for the work you do? Let’s talk in the comments, I would genuinely like to hear your experience.