Claude Fable 5: What Developers Need to Know Before You Ship It (Benchmarks, Cost Per Task & Fallback Behavior)

Claude Fable 5 is Anthropic's first generally available Mythos-class model, released on June 9, 2026, priced at $10 per million input tokens and $50 per million output tokens — exactly double Claude Opus 4.8. The key thing to know: it leads every public benchmark Anthropic published (95.0% on SWE-bench Verified, 80.3% on SWE-Bench Pro), but it costs 2x Opus 4.8 and ships with safety classifiers that quietly reroute some requests to Opus 4.8 mid-call.

For developers deciding whether to wire it into a product, three facts matter more than the headline scores:

• It is a premium tier, not a free upgrade. At $10/$50 per million tokens, Fable 5 is the most expensive major model generally available — but a 90% prompt-caching discount and lower token-per-task counts narrow the real gap on the right workloads.
• You will not always get Fable 5. Requests flagged as cybersecurity, biology, chemistry, or distillation fall back to Opus 4.8 (under 5% of sessions), and you are not billed Fable rates for those.
• Using it requires 30-day data retention. All Mythos-class traffic is retained for 30 days for safety monitoring — not for training — which has real implications for your privacy policy.

This guide covers what Fable 5 actually is, the cost-per-task math, the fallback behavior you have to handle in code, and a direct framework for when it is worth the 2x price.

What is Claude Fable 5?

Claude Fable 5 is the publicly available version of a Mythos-class model — a Claude tier that Anthropic positions above its Opus class in capability. It uses the same underlying weights as Claude Mythos 5, the restricted variant available only to Project Glasswing cyber-defense partners and the US government. The single difference is the safeguard layer: Fable 5 ships with production safety classifiers, while Mythos 5 has some of those classifiers lifted for vetted users.

The launch-day facts every developer should have on hand:

• API model ID: claude-fable-5. The restricted sibling is claude-mythos-5 (Glasswing-only).
• Context window: 1,000,000 tokens by default, with up to 128K output tokens per request.
• Inputs: text, image, and file inputs; text output.
• Thinking mode: one mode only — adaptive.
• Pricing: $10 per million input tokens, $50 per million output tokens; batch pricing is $5/$25.
• Availability: generally available through the Claude API and major cloud marketplaces (AWS, Google Cloud Vertex AI, Microsoft Foundry) as of June 9, 2026.

Anthropic describes Fable 5 as state-of-the-art at coding, knowledge work, vision, and computer use, with the lead over older Claude models growing as tasks get longer and more complex. In plain terms: it is built for long-horizon, autonomous work — large migrations, multi-stage agent runs, and deep analysis — rather than short, high-volume calls.

The benchmarks: where Fable 5 actually wins

Fable 5 leads the public benchmark board, but the scores split into two categories you must read carefully.

On the capabilities most developers actually build on, Fable 5 posts genuine state-of-the-art numbers: 95.0% on SWE-bench Verified, 80.3% on SWE-Bench Pro, and 72.9% on CursorBench at max effort. For comparison, on SWE-Bench Pro, Fable 5's 80.3% beats GPT-5.5's 58.6% and Gemini 3.1 Pro's 54.2% — and Anthropic reports the lead widens as tasks get longer. Stripe, in early testing, reported that Fable 5 compressed a 50-million-line Ruby codebase migration that would have taken a team over two months into a single day.

The second category is the trap: the dazzling cybersecurity and biology benchmark scores belong to Mythos 5, not the Fable 5 you can deploy. Anthropic's table shows the higher of the two models' scores, and the cyber/bio rows are starred to indicate they reflect Mythos 5. The clearest example: on offensive-cyber exploitation tasks, the restricted model scores in the high 70s, but Fable 5 in blocking mode made 0% progress — its safeguards route those queries to Opus 4.8. The rule for developers: never benchmark-shop on Mythos numbers for a Fable deployment.

What Claude Fable 5 actually costs per task

The sticker price ($10/$50 per million tokens) is double Opus 4.8 ($5/$25), but the per-token rate is not the number that determines your bill. Two factors reshape the real cost: prompt caching and token efficiency.

The cost formula is simple: cost = (input_tokens / 1,000,000 × $10) + (output_tokens / 1,000,000 × $50).

Here is a worked example for a single agentic task using 200K input and 50K output tokens:

• Uncached: (0.2 × $10) + (0.05 × $50) = $4.50 per task.
• With prompt caching: Fable 5's 90% prompt-caching discount bills cached input at $1/M instead of $10/M. The same task becomes (0.2 × $1) + (0.05 × $50) = $2.70 per task — a 40% reduction on a workload that resends a large, stable context.

Prompt caching is the single most important lever. A 100K-token system prompt cached for five minutes costs about $1.25 to write and $0.10 to read on each subsequent call, cutting repeat-context cost by 90%. To maximize cache hits, keep the cached prefix stable across turns and push volatile, per-step detail to the end of the prompt.

The second factor that narrows the gap: Anthropic and early customers report Fable 5 finishing tasks in fewer turns and fewer total tokens than Opus-tier models. GitHub's Copilot team reported Fable 5 completing equivalent autonomous coding work with fewer tool calls and lower token consumption. So a job at 2x the per-token rate can land much closer to Opus 4.8 on total cost — but only on hard, long-horizon tasks where that efficiency shows up.

Two pricing footnotes worth knowing: US-only inference (for data-residency needs) adds a 1.1x multiplier, and on Claude.ai subscription plans Fable 5 counts as 2x usage.

The fallback behavior you have to handle in code

This is the part of Fable 5 that catches developers off guard. Fable 5's safety classifiers can refuse a request and reroute it to Claude Opus 4.8 mid-flow — and the behavior differs between the consumer apps and the API.

In the Claude apps, the reroute is automatic and the user simply sees a notice that another model handled the answer. On the API, it is not automatic. When a classifier fires, the API returns stop_reason: "refusal" as an HTTP 200 response, with a field identifying which classifier triggered (cybersecurity, biology/chemistry, or distillation). Your integration has to detect this and decide what to do next.

Anthropic provides a fallbacks parameter so you can retry automatically on Opus 4.8, and you are not charged Fable 5 rates for rerouted requests — fallback credit offsets the prompt-cache cost of the switch. In Claude Code, you enable this by setting ANTHROPIC_DEFAULT_FABLE_MODEL and ANTHROPIC_DEFAULT_OPUS_MODEL; if either model can't be resolved, Claude Code does not switch and the flagged request ends in a refusal you handle with /model.

How often does this happen? Fewer than 5% of real-world sessions trigger a fallback, and Anthropic reports that more than 95% of Fable sessions involve no fallback at all. But the rate is highly domain-dependent: offensive-security work (penetration testing, CTF exercises) and biology-adjacent codebases trigger fallback frequently, often on the very first request. For substantive biology work, expect nearly all requests to reroute. This is expected routing for those domains, not an account flag.

The practical takeaway: if your product touches security research, bio/chem, or anything a classifier might read as distillation, build refusal handling and fallback logic before launch, log which requests fell back, and route safeguarded-domain work to Opus 4.8 directly so you don't pay the Fable rate for an Opus answer.

The 30-day data retention rule

Using any Mythos-class model — Fable 5 included — requires 30-day data retention on all traffic, for both first-party and third-party (cloud marketplace) surfaces. Anthropic uses this data only for safety monitoring (detecting novel jailbreaks and cross-request attacks) and for reducing false positives, not for training new models, and has added new privacy protections including logging all human access and deleting the data after 30 days in almost all cases.

For developers, this matters in two places. First, your privacy policy and data processing agreements need to reflect that prompts sent to Fable 5 are retained for 30 days. Second, if you serve users in regulated regions or industries (healthcare, finance, EU data subjects), confirm that 30-day retention is acceptable for the data your product sends before you ship. On GitHub Copilot, for example, the Fable 5 model policy is off by default for Business and Enterprise and requires explicitly acknowledging the retention term.

Claude Fable 5 vs Opus 4.8: which should you use?

The core difference: Fable 5 is meaningfully more capable on long, complex, autonomous tasks and costs 2x; Opus 4.8 is the better default for everyday work on cost and latency. They are closely related — Fable 5's safeguards literally fall back to Opus 4.8 — so the smart move is to treat them as a pair and route by task, not pick one for everything.

Use Claude Fable 5 when:

• The task is long-running or multi-stage — large codebase migrations, multi-day agent runs, deep research.
• Quality on hard problems matters more than per-token cost.
• You're doing complex analysis or high-fidelity coding where Opus has plateaued.
• The cost of a wrong answer is higher than the token cost of getting it right.

Use Claude Opus 4.8 (or a cheaper model) when:

• The task is well-scoped and routine.
• Latency or cost per request is the priority.
• You're running high volume, where 2x pricing compounds fast.

The pattern most teams settle on: route by task complexity — send hard, high-value jobs to Fable 5 and default everything else to Opus 4.8 or a smaller model. A common cost-efficient agent architecture uses Fable 5 as the orchestrator and hard-reasoning model while delegating mechanical sub-tasks (file edits, test runs, search) to a cheaper model.

If your goal is maximum capability on a genuinely difficult, long-horizon task, choose Fable 5. If your goal is predictable cost on routine, high-volume work, choose Opus 4.8.

A simple decision framework

Three rules to decide whether Fable 5 belongs in your stack:

1. Validate on one real task during the free window. Fable 5 is free on Pro, Max, Team, and seat-based Enterprise plans from June 9 through June 22, 2026; usage credits apply from June 23. Spend that window running one genuinely hard, long-horizon task and measure cost per completed task — not per token, and not vibes — against Opus 4.8.
2. Score outcomes, not edge cases. Track task completion, factual errors, edits required, number of turns, tool calls, latency, and cost. Don't let Fable 5 win just because it's stronger on a rare edge case; decide which tasks actually need it.
3. Architect for cost from day one. Route by difficulty, exploit the 90% prompt-caching discount on stable context, instrument the safeguard fallback so you're not paying Fable rates for Opus answers, and set hard per-task and per-day token budgets so a runaway agent loop can't surprise you.

Frequently asked questions

What is Claude Fable 5?

Claude Fable 5 is Anthropic's first generally available Mythos-class model, released June 9, 2026. It uses the same underlying model as the restricted Claude Mythos 5 but adds production safeguards. It is available via the Claude API as claude-fable-5 at $10 per million input tokens and $50 per million output tokens.

How much does Claude Fable 5 cost?

Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens — double Claude Opus 4.8's $5/$25, but less than half the price of the earlier Mythos Preview. A 90% prompt-caching discount applies to repeated input tokens, and batch pricing is $5/$25.

Is Claude Fable 5 better than Opus 4.8?

On the benchmarks Anthropic published, yes — notably +11 points on SWE-Bench Pro (80.3% vs Opus's lower score) and roughly double on FrontierCode, with the biggest gains on long, complex tasks. On short, routine work the gap narrows and Opus 4.8 is the better value at half the price.

Why did my Fable 5 request return an Opus 4.8 answer?

Fable 5's safety classifiers route flagged cybersecurity, biology, chemistry, or distillation queries to Opus 4.8. This happens in under 5% of sessions, you are not billed Fable rates for the rerouted request, and on the API it returns stop_reason: "refusal" that your code must handle.

What is the difference between Claude Fable 5 and Claude Mythos 5?

They are the same underlying model. Claude Fable 5 is the general-access version with production safeguards; Claude Mythos 5 lifts certain safeguards (mainly cybersecurity) and is restricted to vetted partners through Project Glasswing. Mythos 5 performs slightly stronger on safeguarded domains because Fable 5 falls back to Opus 4.8 there.

Does using Claude Fable 5 require data retention?

Yes. All Mythos-class traffic, including Fable 5, requires 30-day data retention for safety monitoring on both first- and third-party surfaces. The data is not used for training and is deleted after 30 days in almost all cases, but your privacy policy and compliance review should account for it.

When should I use Claude Fable 5 instead of a cheaper model?

Use Fable 5 for genuinely hard, long-horizon tasks — large migrations, multi-stage agent runs, deep research, and high-fidelity coding — where capability changes the outcome and the cost of failure exceeds the token cost. For routine, latency-sensitive, or high-volume work, Opus 4.8 or a smaller model is the smarter economic choice.

Data sourced from Anthropic's official Claude Fable 5 and Mythos 5 launch announcement (June 9, 2026), Anthropic's Claude API and Claude Code documentation, GitHub Copilot's changelog, and published cost-modeling figures. Benchmark scores are Anthropic-reported; independent replications were still pending at the time of writing. Pricing and availability may change — always verify against Anthropic's current pricing page before committing.