选型指南 · 2026-06-13 · by @zayuerweb-dev

Claude Fable 5 Is Out: Price Doubles to $10/$50, How It Differs from Mythos 5, Who Should Upgrade

Q: How much better is Fable 5 than Opus 4.8?

Anthropic claims state-of-the-art on nearly all benchmarks but publishes no specific numbers, so there is no reliable quantitative comparison yet. Use the Opus 4.7 baseline (SWE-bench Verified 87.6%) as the anchor; this page will update when data lands.

Anthropic shipped Claude Fable 5 and Claude Mythos 5 on 2026-06-09 — a new tier that sits above Opus, which Anthropic calls "Mythos-class." Two things to get straight first: pricing is $10 input / $50 output per million tokens, exactly double Opus 4.8 ($5/$25); and Fable 5 and Mythos 5 are the same model with the same specs and price — the only difference is the safety guardrails. The one you and almost every developer can use is Fable 5; Mythos 5 is the restricted version. This piece covers the difference, whether the price is worth it, and whether to move up from Opus 4.8. Full benchmarks aren't published yet — more on that at the end. Prices are a June 2026 snapshot; check the official site before you build.

30-second answer

New tier: Fable 5 is Anthropic's most capable widely released model, opening a "Mythos-class" tier that sits above the Opus line.
Price doubled: $10 / $50 per 1M tokens (input/output) — double Opus 4.8 ($5/$25). 1M context, up to 128k output per request. Batch is $5/$25; cached input gets a 90% discount.
Fable 5 vs Mythos 5 = guardrails only: same model, same capabilities and price. Fable 5 (public) ships safety classifiers that decline or reroute sensitive requests (cybersecurity, bio/chem, distillation) to Opus 4.8; Mythos 5 (restricted) drops those classifiers and is offered only through Project Glasswing to approved organizations. What you and nearly every developer touch is Fable 5.
Who it's for: new projects needing top-end reasoning and long-horizon agents can use Fable 5; for everyday, high-volume, cost-sensitive work, Opus 4.8 ($5/$25) or even Sonnet 4.6 (~$2.6/$15) is the smarter buy. Free on subscription plans June 9–22 — worth a test run.
Benchmarks: Anthropic only claims "state-of-the-art on nearly all benchmarks" and publishes no specific numbers — we won't invent them here.

See the full Claude Fable 5 price table and capabilities →

What's confirmed

As of June 2026, here's what's certain about Fable 5 / Mythos 5:

Release: 2026-06-09.
Price: $10 / 1M input, $50 / 1M output — identical for both models. Anthropic describes it as "less than half the price of Mythos Preview." Batch (async) is $5/$25; prompt-cache hits get roughly a 90% discount.
Specs: 1M-token context, up to 128k output tokens per request.
API IDs: claude-fable-5 and claude-mythos-5.
Availability: Fable 5 is generally available on the Claude API, AWS (including Amazon Bedrock), Google Vertex AI, and Microsoft Foundry; Mythos 5 is not sold publicly — only to approved customers through Project Glasswing.
Data retention: both require 30-day data retention (they're Covered Models) and are not available under zero data retention.
Thinking mode: adaptive thinking is the only mode and is always on (can't be disabled); raw chain-of-thought is never returned — you get a summary or an empty block.

Fable 5 vs Mythos 5

This is where the names mislead. The two models share the same architecture, capabilities, context, and price. There's exactly one difference: safety guardrails.

Claude Fable 5 (public): ships safety classifiers. When a request lands in a sensitive area — cybersecurity, biology/chemistry, model distillation — it may decline, or hand the request to the next-most-capable model, Claude Opus 4.8, to answer. This is the version for everyone.
Claude Mythos 5 (restricted): the same model with those classifiers removed. It's offered in limited release through Project Glasswing (Anthropic's collaboration with the US government), initially to existing Mythos Preview users — cyberdefenders, infrastructure providers — with a planned program for biology researchers that keeps the cyber guardrails on. Access requires going through your Anthropic, AWS, or Google Cloud account team.

In one line: Mythos 5 isn't "a stronger Fable 5" — it's "Fable 5 with some safety limits lifted, for specific organizations." For an ordinary developer the capability is identical; the one you integrate is Fable 5.

What $10/$50 means

It's clearest set against Anthropic's own lineup (per 1M tokens):

Model	Input	Output	Context
Fable 5 / Mythos 5	$10	$50	1M
Opus 4.8	$5	$25	1M
Sonnet 4.6	~$2.6	~$15	~1M

Fable 5's output is 2× Opus 4.8 and more than 3× Sonnet 4.6. That means:

Worth it for: the hardest reasoning, long agent chains, automations that run for a long stretch — places where getting it right once saves rework worth far more than the token delta.
Skip it for: everyday Q&A, writing, high-volume batch work. Opus 4.8 has the same 1M context at half the price; Sonnet 4.6 is cheaper still. Running Fable 5 for everything will blow up the bill fast.
Pragmatic move: keep Sonnet/Opus as the workhorse and route only the steps that genuinely stall to Fable 5. For how to split by task, see which model to use for AI agents.

Three things developers must handle

If you wire up Fable 5 (not Mythos 5) in code, three things differ from earlier Claude models — plan for them before you ship:

1. Refusals: when Fable 5's classifiers decline a request, the Messages API returns stop_reason: "refusal" as a normal 200 response, not an error, and tells you which classifier declined it. Your code should treat this as a normal branch, not an exception.
2. Fallback: a request Fable 5 refuses can usually be served by another Claude model. Anthropic offers three ways: server-side (pass the fallbacks parameter and let the API retry — currently beta), client-side (SDK middleware in several languages), or roll your own retry.
3. Billing: a request refused before any output is generated is not billed; when you retry on another model, "fallback credit" refunds the prompt-cache cost of switching, so you don't pay for that cache twice.

At launch it also supports effort (control thinking depth), task budgets (beta), the memory tool, code execution, programmatic tool calling, context editing, compaction, and vision.

Should you move up from Opus 4.8

Unlike the Opus 4.7→4.8 "same-price upgrade," this one doubles the price, so the question is "does my task justify the price?" By situation:

New project with genuinely hard tasks (complex reasoning, long-horizon agents): go ahead and try Fable 5 — use the free window (June 9–22) to load-test it on your real workload and see how much it actually beats Opus 4.8 before committing.
Production / agent flows already stable on Opus 4.8: don't rush. Price doubles, and Fable 5 adds the refusal/fallback behavior, so switching blindly disturbs a chain you've tuned. Wait until you can quantify "2× the cost buys how much."
Everyday Q&A, writing, high volume: mostly unnecessary. Opus 4.8 or even Sonnet 4.6 is plenty, and the savings are real.

In one line: Fable 5 is a heavy tool you call on demand, not a default workhorse. Whether it's worth it comes down to whether you have tasks that truly justify $10/$50.

About the benchmarks

At the time of writing, Anthropic has not published specific benchmark numbers for Fable 5 (no scores for SWE-bench Verified, Terminal-Bench, GPQA, etc.). The official line is "state-of-the-art on nearly all tested benchmarks," and it names two third-party evals: Cognition's FrontierCode (says it scores highest among frontier models) and the Hebbia finance benchmark (says it's the highest of any model) — but without comparable numbers.

So this piece gives it no scores. The anchor you can use is the last flagship, Opus 4.7: SWE-bench Verified 87.6%, SWE-bench Pro 64.3%, Terminal-Bench 69.4% (full breakdown here). Once Fable 5's official data lands, this page will add its head-to-head against Opus 4.8 and GPT-5.5. To see live prices and capability dimensions any time, use the comparison tool.

FAQ

How much does Claude Fable 5 cost? Officially $10 / 1M input tokens and $50 / 1M output tokens — exactly double Opus 4.8 ($5/$25). Batch (async) is $5/$25; prompt-cache hits get roughly a 90% discount.

What's the difference between Fable 5 and Mythos 5? Capabilities, specs, and price are identical; the only difference is safety guardrails. Fable 5 (public) ships safety classifiers that decline or reroute sensitive requests to Opus 4.8; Mythos 5 (restricted) drops them and is offered only through Project Glasswing to approved organizations. The one ordinary people and developers use is Fable 5.

When was Fable 5 released and where can I use it? 2026-06-09. Fable 5 is available on the Claude API, AWS Bedrock, Google Vertex AI, and Microsoft Foundry, with a 1M context window and up to 128k output per request.

How much better is Fable 5 than Opus 4.8? Anthropic claims "state-of-the-art on nearly all benchmarks" but publishes no specific numbers, so there's no reliable quantitative comparison yet. Use the Opus 4.7 baseline (SWE-bench Verified 87.6%) as the anchor; this page will update when data lands.

Should I upgrade from Opus 4.8 to Fable 5? The price doubles, so check whether your task justifies it. Try the hard tasks during the free window (June 9–22); don't rush a production flow that's stable on Opus 4.8; and for everyday or high-volume work, Opus 4.8 or Sonnet 4.6 is the better value.

Where does this data come from? Anthropic's official announcement and Claude documentation; prices are a June 2026 snapshot. Verify on the official site before building. About the data.