How to Use Claude Fable 5: Refusals, Fallback, and Cost
How to use Claude Fable 5: the full benchmark table, the refusal and fallback contract that changes your code, mandatory retention, and what Opus 5 changed.
Anthropic released Claude Fable 5 on June 9, 2026, the first publicly available model from its Mythos family, at $10 per million input tokens and $50 per million output[1]. It is the most capable model Anthropic has shipped to the general public, and the steepest list price in the Claude lineup.
Learning how to use Claude Fable 5 means learning one thing the other Claude models do not make you handle: it can decline a request. Fable 5 ships with safety classifiers that Mythos 5 does not have, and a declined request comes back as a successful HTTP 200 with stop_reason: "refusal" rather than an error[1]. Your integration needs three new behaviors for that: refusal handling, a fallback path to another model, and an understanding of how refusals are billed.
This guide covers what Fable 5 actually does, how it differs from its restricted sibling Mythos 5, the refusal-and-fallback contract in detail, the export-control suspension that took it offline for eighteen days, and why the arrival of Claude Opus 5 changed the answer to "should I be paying for this?"
TL;DR
- List price is $10 / $50 per million tokens, twice Claude Opus 4.8[3]. On reAPI it runs $8.00 / $40.00, which is 80% of Anthropic's published rate.
- It can refuse. Fable 5 carries safety classifiers covering cybersecurity, biology and chemistry, and distillation attempts. A refusal is an HTTP 200 with
stop_reason: "refusal", and the response reports which classifier declined[1]. - Refusals are not billed when nothing was generated, and fallback credit refunds the prompt-cache cost of retrying on another model[1].
- Thinking cannot be disabled. Adaptive thinking is always on;
thinking: {"type": "disabled"}is unsupported. Useeffortto control depth[1]. - 30-day data retention is mandatory. Fable 5 and Mythos 5 are Covered Models and are not available under zero data retention[1].
- Claude Opus 5 changed the math. Anthropic describes Opus 5 as delivering frontier intelligence at half the cost of Fable 5[4].
What Claude Fable 5 is
Claude Fable 5 is Anthropic's most capable widely released model, built for the most demanding reasoning and long-horizon agentic work[1]. The API identifier is claude-fable-5.
The launch is two models, and conflating them is the easiest mistake to make.
| Claude Fable 5 | Claude Mythos 5 | |
|---|---|---|
| API model ID | claude-fable-5 | claude-mythos-5 |
| Availability | Generally available | Limited release through Project Glasswing only |
| Safety classifiers | Yes | No |
| Context / max output | 1M tokens / 128k | Same |
| List price | $10 / $50 per MTok | Same |
They share the same underlying capabilities and the same specs[1]. The difference is the classifier layer. So when a benchmark number is attributed to "Mythos 5 / Fable 5" on a cybersecurity or biology row, read it carefully: that ceiling belongs to the model without the brakes. The public Fable 5 you can actually call will decline on exactly those questions by design.
Benchmarks
Anthropic published a head-to-head against Mythos Preview, Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. The standout is agentic coding, where the lead is large rather than marginal[1].
| Benchmark | Mythos 5 / Fable 5 | Mythos Preview | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|
| Agentic coding (SWE-Bench Pro) | 80.3% | 77.8% | 69.2% | 58.6% | 54.2% |
| Agentic coding (FrontierCode Diamond, xhigh) | 29.3% | n/a | 13.4% | 5.7% | n/a |
| Agentic coding (Terminal-Bench 2.1) | 88.0%* | n/a | 82.7% | 83.4%† | 70.7%† |
| Knowledge work (GDPval-AA) | 1932 | n/a | 1890 | 1769 | 1314 |
| Knowledge work vision (GDP.pdf, no tools) | 29.8% | n/a | 22.5% | 24.9% | 16.7% |
| Spatial reasoning (Blueprint-Bench 2) | 38.6% | n/a | 14.5% | 36.2% | 26.5% |
| Tool use (AutomationBench) | 17.4% | n/a | 15.5% | 12.9% | 9.6% |
| Computer use (OSWorld-Verified) | 85.0% | 85.4% | 83.4% | 78.7% | 76.2% |
| Legal (Legal Agent Benchmark) | 13.3% | n/a | 10.4% | 2.1% | 0.0% |
| Reasoning (HLE, no tools) | 59.0%* | 56.8% | 49.8% | 41.4% | 44.4% |
| Reasoning (HLE, with tools) | 64.5%* | 64.7% | 57.9% | 52.2% | 51.4% |
| Biology (BioMysteryBench, hard) | 46.1%* | 29.6% | 40.0% | n/a | n/a |
| Biology (BioMysteryBench, human-solved) | 83.9%* | 82.6% | 80.4% | n/a | n/a |
| Cybersecurity (ExploitBench, Cap%) | 78.0%* | 69.0% | 40.0% | 34.0% | n/a |
| Health (HealthBench Professional) | 66.0%* | 64.7% | 56.9% | 51.8% | n/a |
† Terminal-Bench figures for GPT-5.5 and Gemini 3.1 Pro were run through Codex CLI and Gemini CLI respectively.
Anthropic's own methodology note under that table is the most important thing on it: reported scores land within a one to three percentage point difference for Mythos 5 and Fable 5, and the table shows the higher of the two. Starred rows show a larger gap "due to our blocking safeguards for cybersecurity and biology-related questions," and on those "Claude Fable 5 performs closer to Claude Opus 4.8 due to fallbacks"[1].
Read that carefully before quoting any starred number. The 78.0% cybersecurity figure and the 66.0% health figure are not what the public model delivers on those questions.
The unstarred rows are the gains the public model keeps. SWE-Bench Pro at 80.3% is more than ten points above Opus 4.8 and over twenty above GPT-5.5 and Gemini 3.1 Pro.
FrontierCode is the row that explains the price.
On that hardest subset, Fable 5 climbs from roughly 11% at low effort to about 31% at max, while Opus 4.8 tops out near 13% at xhigh and comes back down at max, and GPT-5.5 stays flat around 5% to 6% across its whole ladder. Fable 5 converts extra compute into real accuracy on hard problems in a way the others do not. You are paying for a model designed to keep working a problem longer.
Partner results from the launch are specific enough to be falsifiable: Stripe reported completing a 50-million-line Ruby codebase migration in roughly one day against a two-month manual estimate; Hex said Fable 5 was the first model to clear 90% on its core analytics benchmark; Hebbia recorded its highest finance-benchmark score. On vision, the model completed Pokémon FireRed using vision alone.
How to use Claude Fable 5 when it refuses
This is the part of Fable 5 that changes your code, and it is worth being precise because it is easy to describe wrongly.
Fable 5 does not silently reroute a risky request to a smaller model. When a classifier declines, the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response, not an error, and the response reports which classifier declined it[1]. Code that indexes the first content block unconditionally will break on this. Check the stop reason before reading content.
Retrying on another Claude model is your call, and there are three supported paths[1]:
- Server-side. Pass the
fallbacksparameter and let the API retry for you, either in its"default"mode using Anthropic's recommended models or with a list you name. In beta on the Claude API. - Client-side. Use the SDK middleware to retry from the client on any platform.
- Manual. Build the retry yourself, in any language.
The billing rules matter as much as the mechanics. You are not billed for a request refused before any output is generated, and when you retry on another model, fallback credit refunds the prompt-cache cost of switching so you do not pay it twice[1].
Three domains are fenced: cybersecurity, including exploitation and offensive work; biology and chemistry; and distillation attempts aimed at extracting Fable 5's capabilities to train another model.
What the export-control suspension revealed
Fable 5 has already been taken offline once, and the reason is worth knowing before you build a product on it.
On Friday, June 12, three days after launch, the US government applied export controls to Fable 5 and Mythos 5, requiring Anthropic to restrict access to foreign nationals whether inside or outside the United States. Because the order took effect immediately and Anthropic had no reliable way to verify nationality in real time, it suspended access to both models for all users[2].
The controls were lifted on June 30, and Fable 5 returned globally on July 1 across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with AWS, Google Cloud, and Microsoft Foundry re-enabled afterward[2]. Mythos 5 access was restored to a set of US organizations following government approval on June 26[2].
Eighteen days of total unavailability is a real operational fact, not a footnote. If Fable 5 is the only model on a production path, that path had no service for eighteen days. The practical lesson is the same one the refusal contract already teaches: this model needs a fallback, and the fallback should be configured before you need it.
The retention requirement
Fable 5 and Mythos 5 carry mandatory 30-day data retention and are not available under zero data retention, because both are designated Covered Models[1]. Organizations that previously held zero-retention agreements do not keep them for this model.
If you have data-residency or retention commitments to your own customers, this is the clause to read before adopting Fable 5. It is not negotiable per-account, and no amount of gateway configuration changes it.
Pricing, and what Opus 5 changed
Fable 5 lists at $10 per million input tokens and $50 per million output, twice Opus 4.8's $5 / $25[3]. That was the calculus in June.
It is not the calculus now. When Anthropic shipped Claude Opus 5 on July 24, it described the model as delivering frontier intelligence at half the cost of Claude Fable 5[4], and Opus 5 lists at $5 / $25, unchanged from Opus 4.8[3]. Opus 5 also has no classifier-refusal layer of Fable 5's kind, no mandatory 30-day retention, and a full effort ladder.
On reAPI the spread is wider still:
| Model | reAPI rate per 1M tokens |
|---|---|
| Claude Fable 5 | $8.00 input / $40.00 output |
| Claude Opus 5 | $2.40 input / $12.00 output |
| Claude Opus 4.8 | $4.00 input / $20.00 output |
Fable 5 costs more than three times what Opus 5 costs on the same gateway. That does not make Fable 5 pointless. It does mean the burden of proof moved: you now need a specific reason to reach for it, measured on your own tasks.
Fable 5, Opus 5, or Opus 4.8?
Reach for Fable 5 when the task is long-horizon and genuinely hard, and when you have measured that it wins: repo-scale migrations, multi-file refactors against a trustworthy test suite, overnight agent loops, dense analytical work where being wrong is expensive. The FrontierCode curve is the evidence, since accuracy keeps climbing with effort where other models flatten.
Start with Opus 5 for most new work. Anthropic's own framing puts it close to Fable 5's frontier intelligence at half the price, and it does not carry the refusal layer or the retention requirement.
Stay on Opus 4.8 for the broad middle where your evals already characterize its behavior, and where its cheaper fast mode matters.
One operational note that is easy to miss: a security or life-sciences team that needs full capability is worse off paying Fable 5 rates, because those are exactly the domains where the classifiers decline. Calling Opus 4.8 or Opus 5 directly, or applying to Project Glasswing for Mythos 5, is the honest path.
Autonomy and file-based memory
The capability Anthropic stresses most is duration. Fable 5 and Mythos 5 can work autonomously for longer than any previous Claude models, and the supporting examples are deliberately long-horizon: with file-based memory, Mythos 5's performance on Slay the Spire improved three times more than Opus 4.8's, and in life-sciences work it ran autonomously over a week to assemble single-cell data for millions of cells across 138 species.
Supported features at launch include effort, task budgets in beta, the memory tool, code execution, programmatic tool calling, tool-result clearing through context editing, compaction, and vision[1].
Two request-surface rules follow from the always-on thinking. thinking: {"type": "disabled"} is unsupported, so effort is the only lever for thinking depth. And the raw chain of thought is never returned: thinking.display defaults to "omitted", with "summarized" returning a readable summary instead[1].
Calling Claude Fable 5 alongside your other models
Read the last four sections together and one requirement falls out: Fable 5 needs a second model wired up next to it, not as an optimization but as a correctness requirement. It refuses on three domains by design. It was fully unavailable for eighteen days. And a model at a third of its price now covers most of what people bought it for.
reAPI exposes Claude Fable 5 through an OpenAI-compatible /v1/chat/completions endpoint, so the same client that calls Opus 5, GPT, and Gemini calls it too:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_REAPI_KEY",
base_url="https://api.reapi.ai/v1",
)
resp = client.chat.completions.create(
model="claude-fable-5",
messages=[{"role": "user", "content": "Plan and execute this multi-file refactor."}],
max_tokens=16000,
stream=True,
)Rates on the gateway are $8.00 per million input tokens and $40.00 per million output, which is 80% of Anthropic's published $10 / $50. The gateway also exposes the native Anthropic /v1/messages surface, so SDKs written for that format work unchanged. Details are on reapi.ai/docs/claude-fable-5 and reapi.ai/models/claude-fable-5.
Switching the model string when a refusal comes back is a three-line change when both models sit behind one key and one client. It is a project when they do not.
FAQ
What happens when Claude Fable 5 refuses a request?
The Messages API returns a successful HTTP 200 with stop_reason: "refusal", and reports which classifier declined it. It is not an error response, so check the stop reason before reading the content[1].
Am I billed for a refused request?
No, not when the request is refused before any output is generated. If you retry on another model, fallback credit refunds the prompt-cache cost of the switch[1].
What is the difference between Claude Fable 5 and Claude Mythos 5?
Same capabilities, same specs, same price. Mythos 5 ships without the safety classifiers and is available only through Project Glasswing to approved customers. Fable 5 is the generally available version[1].
Can I turn thinking off on Claude Fable 5?
No. Adaptive thinking is always on and thinking: {"type": "disabled"} is unsupported. Control depth with the effort parameter instead[1].
Does Claude Fable 5 support zero data retention?
No. Fable 5 and Mythos 5 are designated Covered Models, carry 30-day retention, and are not available under zero data retention[1].
Why was Claude Fable 5 unavailable in June?
The US government applied export controls on June 12, 2026, requiring restricted access for foreign nationals. With no way to verify nationality in real time, Anthropic suspended access for all users. The controls were lifted June 30 and Fable 5 returned July 1[2].
Should I use Claude Fable 5 or Claude Opus 5?
Start with Opus 5 unless you have measured Fable 5 winning on your own tasks. Anthropic describes Opus 5 as delivering frontier intelligence at half Fable 5's cost, without the refusal layer or the retention requirement[4].
How do I call Claude Fable 5 with the OpenAI SDK?
Point the OpenAI client at https://api.reapi.ai/v1 and set model to claude-fable-5. The endpoint is OpenAI-compatible, and the native Anthropic /v1/messages surface is available too.
Buying the top of the lineup on purpose
Fable 5 is a genuinely different class of model on the row that matters most for agentic work, and the FrontierCode curve is the honest argument for it: on hard problems, extra effort keeps buying accuracy instead of flattening out. If your workload is a repo-scale migration or an overnight loop where a failed run costs more than the tokens, that curve is worth paying for.
Everything else about it is a constraint you have to design around. It declines three domains by contract, it carries mandatory 30-day retention with no zero-retention option, it went fully dark for eighteen days under export controls, and a model at a third of the price now sits close behind it. None of that is disqualifying, and all of it argues the same thing: know how to use Claude Fable 5 with a fallback already wired up, and reach for it deliberately rather than by default.
References
- Anthropic. Introducing Claude Fable 5 and Claude Mythos 5 — capabilities, refusals, fallback, billing, retention, and supported features. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5
- Anthropic. Redeploying Fable 5 — export-control timeline and restoration. Retrieved July 2026 from anthropic.com/news/redeploying-fable-5
- Anthropic. Pricing — token, cache, and batch rates. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/pricing
- Anthropic. What's new in Claude Opus 5. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/models/whats-new-opus-5
- Anthropic. Refusals and fallback — response shapes and retry paths. Retrieved July 2026 from platform.claude.com/docs/en/build-with-claude/refusals-and-fallback
Further reading
- reAPI. How to use Claude Opus 5. reapi.ai/blog/how-to-use-claude-opus-5
- reAPI. How to use Claude Opus 4.8. reapi.ai/blog/how-to-use-claude-opus-4-8
- reAPI. Claude Fable 5 endpoint reference. reapi.ai/docs/claude-fable-5
Autor
Categorías
Más publicaciones
Seedream 5.0 Pro Price: The Pixel Line That Decides It
The Seedream 5.0 Pro price splits at 2.36 million pixels, not at a tier label. What each tier costs, and how to size a 16:9 export to stay in the cheap one.
Best WaveSpeed Alternatives in 2026: 5 Options Compared
Looking for WaveSpeed alternatives in 2026? Compare fal.ai, Replicate, Together AI, RunPod, and reAPI on model range, pricing, speed, and API design.
Best CometAPI Alternatives in 2026: 5 Options Compared
Looking for CometAPI alternatives in 2026? Compare OpenRouter, WaveSpeed, Together AI, Replicate, and reAPI on models, pricing, speed, and API design.