Claude Fable 5 vs Opus 4.8: Which Should You Use?

Claude Fable 5 is Anthropic's most capable widely released model, but it costs $10 per million input tokens and $50 per million output tokens, keeps adaptive thinking always on, and runs a refusal classifier that can decline a request; Claude Opus 4.8 is half the price at $5/$25, with a simpler request surface (optional thinking, no refusal classifier, and zero data retention allowed). Pick Opus 4.8 as your default for nearly all production work, and reach for Fable 5 only when a task is hard enough or long enough that the extra capability is worth double the price.

Update — June 23, 2026: Claude Fable 5 is currently suspended and unavailable following a U.S. export-control directive, so Opus 4.8 is the model you can actually run today. The comparison below still matters for when Fable 5 access returns — and the practical answer right now is Opus 4.8.

The quick answer

As of June 2026, Claude Opus 4.8 is the right default for most teams, and Claude Fable 5 is the upgrade you reach for on the hardest jobs. Both are listed as Anthropic's most capable models in their respective tiers, and there are no published benchmark scores to separate them — so the decision comes down to price, request surface, and how demanding your task really is.

If your workload is everyday production work — extraction, summarization, classification, chat, standard coding and agent loops — Opus 4.8 handles it at half the cost and with fewer moving parts. If your workload is a long-horizon autonomous agent run, a sprawling reasoning problem, or a first-shot build of a well-specified system, Fable 5's extra headroom can be worth the premium. Many teams run both: Opus 4.8 as the workhorse, Fable 5 reserved for the small slice of tasks that genuinely need it.

Side-by-side specs

The two models are far more alike than different: same tokenizer, same 1M context window, same 128K max output, and the same shared feature set (effort levels from low through max, task budgets, the memory tool, code execution, and vision). The differences that matter are price, whether thinking is optional, the refusal classifier, and data-retention support.

The price gap and what it buys

Fable 5 costs exactly twice what Opus 4.8 does on both input and output — $10/$50 versus $5/$25 per million tokens. On a high-volume workload that difference compounds fast, so the question is never "which model is better in the abstract" but "is this specific task hard enough to justify paying double."

What the premium buys is capability at the frontier: Fable 5 is positioned as the step up for the most demanding reasoning and the longest-horizon agentic work. On tasks where Opus 4.8 already succeeds, you are paying extra for headroom you do not use. On tasks at the edge of what is possible — runs that span many tool calls and minutes of wall-clock time, or problems where a single wrong turn cascades — the extra capability can be the difference between a finished deliverable and a stalled one.

A practical rule: start on Opus 4.8, measure where it falls short on your own evals, and promote only those routes to Fable 5. Because the two models share a tokenizer, your token counts stay roughly the same across the switch — only the per-token rate changes, which makes the cost delta easy to forecast before you commit.

Behavioral and API differences

Beyond price, Fable 5 has a more opinionated request surface than Opus 4.8, and three quirks are worth knowing before you write code against it.

Always-on thinking. On Fable 5, adaptive thinking is always on — you cannot disable it. Sending thinking: {type: "disabled"} returns a 400, so you simply omit the thinking parameter entirely. The raw chain of thought is never returned; you get a summary or an empty string depending on the display setting. Opus 4.8 makes thinking optional: you can omit it or set an effort level, and it even supports mid-conversation system messages.

Refusals and fallbacks. Fable 5 runs safety classifiers that can decline a request, returning a successful HTTP 200 with stop_reason: "refusal". A request refused before any output is not billed; a mid-stream refusal bills the streamed partial, which you discard. Because false positives can hit benign adjacent work (security tooling, life-sciences tasks), production code should check stop_reason before reading content and opt into a fallback model such as Opus 4.8. Opus 4.8 has no refusal classifier of this kind, so there is nothing extra to handle. You can read Anthropic's Fable 5 introduction for the full refusal and fallback semantics.

Data retention. Fable 5 requires 30-day data retention and is not available under zero data retention — a ZDR organization gets a 400 on every request. Opus 4.8 is available under ZDR, which can be decisive for regulated or privacy-sensitive deployments. Anthropic's models overview lists the current availability details, and the migration guide walks through moving code between the two.

Which to pick by use case

The right choice falls out of how hard the task is and what constraints your organization operates under. Here is how the common cases break down as of June 2026.

Pick Opus 4.8 when: you are doing standard production work — extraction, classification, summarization, Q&A, chat, ordinary coding and agentic loops; you care about cost at volume; you want a simpler request surface with optional thinking and no refusal handling; or your organization requires zero data retention. This covers the large majority of real workloads.

Pick Fable 5 when: the task is at the frontier of difficulty — long-horizon autonomous agents, deep multi-step reasoning, or a one-shot build of a complex, well-specified system; the value of the outcome clearly justifies paying double; and your organization can meet the 30-day retention requirement. Reserve it for the slice of work where Opus 4.8 measurably underperforms.

Limitations and Tradeoffs

Neither model is the universal answer, and choosing the more capable one by default is a common and expensive mistake.

When not to use Fable 5: if cost matters at all and the task is routine, the doubled price is wasted — Opus 4.8 will do the same job for half. Fable 5's refusal classifier can also produce false positives on legitimate but adjacent work like security tooling or bio-adjacent research, which means extra handling code and the occasional retry through a fallback. And if your organization runs under zero data retention, Fable 5 is simply off the table — every request returns a 400.

When Opus 4.8 is enough: for nearly everything. Most production workloads never approach the ceiling where Fable 5's extra capability shows up, so Opus 4.8 delivers the result at lower cost and with a cleaner request surface. The honest tradeoff is that on the genuinely hardest, longest tasks, Opus 4.8 may stall or underperform where Fable 5 would push through — which is exactly the narrow case where paying the premium pays off.

Related Guides

• What Is Claude Fable 5?
• Best Claude Models for OpenClaw
• Sakana Fugu vs Claude Fable 5
• Best Cheap AI Models in 2026

FAQ

Is Claude Fable 5 better than Opus 4.8?

Fable 5 is Anthropic's most capable widely released model and is the step up for the hardest, longest-horizon agentic and reasoning work. But "better" depends on the task: there are no published benchmark scores separating them, and for most production work Opus 4.8 delivers equivalent results at half the price with a simpler request surface. Fable 5 is better only where the task is demanding enough to need its extra headroom.

How much more does Claude Fable 5 cost than Opus 4.8?

Exactly twice as much. Fable 5 is $10 per million input tokens and $50 per million output tokens; Opus 4.8 is $5 input and $25 output. Because both models share a tokenizer, your token counts stay roughly the same when you switch, so the cost difference is the full 2x on the per-token rate.

Should I switch from Opus 4.8 to Fable 5?

Only for the routes where Opus 4.8 measurably falls short on your own evaluations. Keep Opus 4.8 as the default for cost-efficiency and the simpler surface, and promote individual hard or long-horizon tasks to Fable 5 when the outcome justifies double the price. Remember Fable 5 requires 30-day data retention, so a zero-data-retention organization cannot switch at all.

Do Fable 5 and Opus 4.8 share a tokenizer?

Yes. Claude Fable 5 uses the same tokenizer as Claude Opus 4.8, so the same input text produces roughly the same token count on both models. Only the per-token price differs, which makes the cost impact of switching easy to forecast before you commit.