Claude Fable 5: The Most Capable Model Yet, and When It's Worth the Price
Claude Fable 5 is the most capable model Anthropic has ever released to the public. It also costs exactly double Claude Opus 4.8. Both facts land on June 9, and together they set up the question worth asking before you wire it into anything. Fable 5 is good. The real question is when it earns the bill.
I build systems that run unattended for hours, so a model that finishes a hard job on the first pass is worth real money to me. I also pay the bill, so a model that costs twice as much per token has to prove it. This is the honest read on where Fable 5 wins, where it does not, and the math that decides between them.
The short version: Claude Fable 5 is the new top of Anthropic’s lineup, with the best published scores in coding, hard reasoning, and vision. It runs $10 per million input tokens and $50 per million output, exactly twice Opus 4.8. It thinks on every request and a hard task can run for minutes before it answers, and it can decline some requests outright, so you keep a fallback ready. It is a specialist, not a default. Reach for it when a correct first answer is worth more than the token bill, and route everything routine to a cheaper model.
What Claude Fable 5 actually is
Fable 5 sits at the top of Anthropic’s lineup, a tier above the Opus class. It carries a 1 million token context window and writes up to 128,000 tokens in a single response. It is built for the work that breaks smaller models: long autonomous runs, first-try builds of well-specified systems, dense documents, and messy screenshots.
The part you feel in the invoice is the price. Here is the lineup, flat:
| Model | Context | Max output | Input / 1M | Output / 1M |
|---|---|---|---|---|
| Claude Fable 5 | 1M | 128K | $10 | $50 |
| Claude Opus 4.8 | 1M | 128K | $5 | $25 |
| Claude Sonnet 4.6 | 1M | 64K | $3 | $15 |
| Claude Haiku 4.5 | 200K | 64K | $1 | $5 |
Two numbers stand out. Fable 5 is exactly twice the price of Opus 4.8 on both input and output. It is also less than half what the previous frontier tier cost, so the ceiling got higher and cheaper at the same time. The gap between Fable 5 and Haiku 4.5 is tenfold on input and tenfold on output. That spread is the whole reason routing matters, and I will come back to it.
The benchmarks, and what they do not tell you
Anthropic put Fable 5 at the top of nearly every test it published. The coding number is the one most builders care about, so start there.
| Benchmark | Claude Fable 5 | Claude Opus 4.8 |
|---|---|---|
| SWE-Bench Pro (agentic coding) | 80.3% | 69.2% |
| FrontierCode (hardest split) | 29.3% | 13.4% |
| GDP.pdf (vision, no tools) | 29.8% | 22.5% |
On SWE-Bench Pro, Anthropic’s agentic coding test, Fable 5 posts 80.3 percent against Opus 4.8 at 69.2. The rest of the field sits further back, with the strongest competitors near 58 and 54 percent on the same test. On the hardest reasoning split it roughly doubles Opus, and on document vision with no tools it leads there too.
A benchmark measures the ceiling, not your bill. Two things matter more in practice. Fable 5 thinks on every request, with no switch to turn that off, and a single hard task can run for minutes before it answers. You design around that with streaming and a job queue, not a blocking call your user sits and watches. The capability is real, but it shows up as fewer retries and less of your own time, not as a faster API call.
The one thing you design around
Fable 5 ships with safety classifiers that can decline a request. When that happens the API returns a normal success, not an error, with the response marked as a refusal. Benign work near security or biology can trip it, so you check for the refusal and fall back to Opus, which carries no such classifier.
def ask(messages):
resp = client.messages.create(model="claude-fable-5", messages=messages, max_tokens=16000)
if resp.stop_reason == "refusal":
# Fable declined. Opus has no classifier, so retry there.
resp = client.messages.create(model="claude-opus-4-8", messages=messages, max_tokens=16000)
return resp
That is the whole pattern. The API can also run the fallback for you, but the shape is the same. A refused request that lands on Opus is a far better experience for your user than a dead end, and you are not billed for a request Fable refuses before it writes anything. Build the fallback once and you stop thinking about it.
When double the price pays for itself
Every feature I build passes one cost test. What is the cheapest path to a correct result, and how much of the work can that path carry? Most work has a cheap correct path. Fable 5 is for the slice that does not.
Say a model has to untangle a knotted service overnight and have it building clean by morning. Run that on Opus and it might take three passes and an hour of your cleanup. Run it on Fable and it lands in one. The token bill doubled. The finished job got cheaper, because the expensive part was never the tokens. It was the second attempt, the third attempt, and your morning.
| Reach for Claude Fable 5 | Stay on Opus 4.8 or cheaper |
|---|---|
| Long autonomous runs that must finish clean | Interactive chat and quick edits |
| A hard build you want right on the first try | Work Opus already handles |
| Dense documents and messy screenshots | Plain text and clean inputs |
| Jobs where a wrong answer is expensive | High-volume, cost-sensitive routing |
Per token, Fable 5 is the priciest model Anthropic sells. Per finished job, on the jobs it is built for, it is often the cheapest, because it removes the retries and the handholding a weaker model bills you for in your own time. On everything routine it is a waste, and a keyword rule or Haiku does the same work for a tenth of the price.
Reach for it when a right answer on the first try is worth more than the tokens it burns. The rest of the time, the most capable model is the wrong one.
Frequently asked questions
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s most capable widely released model, generally available from June 9, 2026. It has a 1 million token context window, writes up to 128,000 tokens per response, and is built for hard reasoning and long-horizon agentic work. At launch it tops nearly every public benchmark Anthropic reported.
How much does Claude Fable 5 cost?
It is $10 per million input tokens and $50 per million output tokens. That is exactly double Claude Opus 4.8, which runs $5 and $25, and less than half what the previous frontier tier cost. For context, Sonnet 4.6 is $3 and $15, and Haiku 4.5 is $1 and $5.
Claude Fable 5 vs Opus 4.8: which should I use?
Use Opus 4.8 for most work. Reach for Fable 5 on the hardest jobs: long autonomous runs that must finish clean, first-try builds where a wrong answer is expensive, and dense document or vision work. On SWE-Bench Pro it scores 80.3 percent against Opus at 69.2, but that edge only pays off when the job is hard enough to need it.
What is the catch with Fable 5’s safety classifier?
Fable 5 can decline a request. The API returns a normal response marked as a refusal rather than an error, and benign work near security or biology can trigger it. Plan for it: check for the refusal and fall back to Opus 4.8, which has no such classifier. You are not billed for a request Fable refuses before it writes anything.
Is Claude Fable 5 worth it for a small team?
As a default, no. As a specialist, often yes. Keep routine work on a cheaper model or a plain keyword rule, and call Fable 5 for the slice that genuinely needs it. Priced per finished job rather than per token, it earns its keep exactly when being right the first time matters.