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

Anthropic now ships two genuinely top-tier models, and clients keep asking us the same question: should we move our hardest work to Claude Fable 5, or stay on Claude Opus 4.8? Fable 5 wins the benchmark headlines, but it costs roughly twice as much and arrived with a refusal problem. Opus 4.8 is the autonomous default Anthropic points most teams toward. Here is the honest, pragmatic head-to-head we give companies across the USA, UK, Canada, Europe and Australia when we design their AI integration stack.

The two models at a glance

Claude Fable 5 is Anthropic's most capable widely released model, built for demanding reasoning and long-horizon agentic work. It carries a 1M-token context window and up to 128K tokens of output, with thinking always on, adaptive and effort-controlled. It went generally available on June 9, 2026 across the Claude API, AWS Bedrock, Vertex and Foundry. Pricing is $10 input and $50 output per million tokens, with up to a 90% prompt-caching discount.

Claude Opus 4.8 is the most capable Opus-tier model: highly autonomous, with the same 1M-token context and 128K output ceiling. It is priced at $5 input and $25 output per million tokens — roughly half of Fable 5. Anthropic positions Opus 4.8 as the default for most complex work and reserves Fable 5 for the highest-capability needs.

That framing matters. This is not a "newer is always better" story. Anthropic itself is telling teams to start on Opus 4.8 and graduate to Fable 5 only where the work demands it.

Benchmarks: where Fable 5 pulls ahead

On published third-party benchmarks citing Anthropic's own materials, Fable 5 is clearly the stronger model on hard coding and agentic tasks. SWE-bench Verified lands around 95% for Fable 5 versus roughly 88.6% for Opus 4.8. On the harder SWE-bench Pro, Fable 5 posts a top score of 80.3% against Opus 4.8's 69.2%. FrontierCode Diamond — a deliberately brutal coding evaluation — shows the widest gap: 29.3% for Fable 5 versus 13.4% for Opus 4.8. And on GDPval agentic Elo, Fable 5 reaches 1932 to Opus 4.8's 1890.

Those are meaningful gaps on the tasks where models are pushed to their limit. The most striking field anecdote: Stripe reportedly used Fable 5 to complete a 50-million-line Ruby migration in a single day. If your roadmap includes work at that scale and difficulty — sprawling refactors, multi-day autonomous agents, frontier-level reasoning — Fable 5's lead is exactly what you are paying for.

The caveat is that benchmark gaps shrink fast as tasks get easier. On routine production work — summarisation, extraction, customer-facing assistants, standard code generation — both models clear the bar comfortably, and the 2x price difference buys you very little.

Price, latency and the cost of "always thinking"

The price story is simple: Fable 5 is about twice Opus 4.8 per token. At $10/$50 versus $5/$25, that difference compounds quickly across a high-volume product. The 90% caching discount helps when you reuse large, stable context, but it does not change the underlying ratio for fresh tokens.

Fable 5 also keeps its thinking always on. That adaptive, effort-controlled reasoning is part of why it scores so well, but it tends to spend more tokens and add latency on requests that did not need deep deliberation. For a chatbot answering thousands of simple questions an hour across European and Australian timezones, that is cost and slowness you feel. Opus 4.8 stays the cheaper, leaner choice for the broad middle of a workload. If you want the wider tiering picture, our Opus vs Sonnet vs Haiku guide covers how these fit alongside the rest of the family.

Refusals: the launch problem you should plan around

Fable 5 arrived with a real wrinkle. At launch its safety classifiers over-refused benign prompts — reports from The Register and others cited mundane requests like resume editing and ordinary biology or "cancer" questions getting blocked. Commentators including Nathan Lambert raised a separate "silent degradation" concern, where the model quietly under-delivers rather than refusing outright. Anthropic reportedly acknowledged it "made the wrong tradeoff" and committed to visible fallback behaviour and clearer refusal reasons.

For a business, refusals are not an abstract safety debate — they are failed user interactions and support tickets. Opus 4.8 currently triggers fewer of these false refusals, which is one more reason it is the safer default for customer-facing flows. If you do deploy Fable 5, build for it: evaluate on your real prompts, log refusals, and keep a fallback path. This is the kind of guardrail we bake into every AI agent build. Our writeup on Fable 5's pros and cons for business goes deeper on the tradeoffs.

The cyber, bio and chem catch

Here is a detail that quietly settles a lot of decisions. Cyber, bio, chem and distillation-style requests route or fall back to Opus 4.8 even when you call Fable 5. In those domains, Fable 5 effectively behaves like Opus 4.8 — same capability, but you would be paying the Fable 5 rate for it. If your application is concentrated in security tooling, life sciences or chemistry workflows, the premium buys you essentially nothing over Opus 4.8 for the requests that route away. For teams in regulated sectors across the UK, Canada and Europe, that is a clean reason to default to Opus 4.8 and skip the upcharge. If you are still getting your bearings on the model itself, our explainer on what Claude Fable 5 is is a good primer.

When each one wins

Choose Opus 4.8 when: you want a strong, autonomous default for most complex work; cost and latency matter at volume; you are running customer-facing flows where refusals hurt; or your workload lives in cyber, bio or chem domains where Fable 5 routes to Opus 4.8 anyway. For the large majority of business builds, this is the smarter pick.

Choose Fable 5 when: the task genuinely stretches the frontier — large-scale code migrations, multi-day autonomous agents, the hardest reasoning problems where the SWE-bench and FrontierCode gaps translate into real outcomes. When a single correct result is worth far more than the token premium, and you have evaluated that Fable 5 actually outperforms on your specific work, the extra capability earns its price. One more operational note: Fable 5 requires 30-day data retention, which can matter for compliance-sensitive teams.

Our honest, augmentation-first advice: start on Opus 4.8, instrument your pipeline, and promote only the tasks where evaluation proves Fable 5 is worth 2x. Picking between siblings within a single product is its own exercise — our Fable 5 vs Sonnet and Haiku selection guide walks through routing the rest of the workload. We design exactly this kind of model routing as part of every enterprise AI engagement, so you get frontier capability where it counts without paying frontier prices everywhere.

Frequently Asked Questions

Should a business use Claude Fable 5 or Opus 4.8?

For most complex work, Claude Opus 4.8 is the smarter default — it is highly autonomous, costs roughly half as much as Fable 5, and ships fewer refusals on benign prompts. Reach for Fable 5 only when a task genuinely needs the highest-capability reasoning or long-horizon agentic work, where its published benchmark lead justifies the premium.

How much do Claude Fable 5 and Opus 4.8 cost?

Claude Fable 5 is priced at $10 input and $50 output per million tokens, with up to a 90% prompt-caching discount. Claude Opus 4.8 is $5 input and $25 output per million tokens. Fable 5 is roughly twice the price of Opus 4.8, so the extra capability needs to pay for itself.

Is Fable 5 actually better than Opus 4.8 on benchmarks?

On published third-party benchmarks citing Anthropic materials, Fable 5 leads: SWE-bench Verified around 95% versus 88.6%, SWE-bench Pro 80.3% versus 69.2%, FrontierCode Diamond 29.3% versus 13.4%, and a GDPval agentic Elo of 1932 versus 1890. The gaps are real but vary by task, so they only matter where your workload is genuinely that hard.

Why does Fable 5 refuse some safe requests?

At launch Fable 5's safety classifiers over-refused some benign prompts — reports from The Register and others cited resume editing and biology questions. Anthropic reportedly acknowledged it made the wrong tradeoff and committed to visible fallback and clearer refusal reasons. Opus 4.8 currently triggers fewer of these false refusals.

Do cyber, bio or chem requests behave differently on Fable 5?

Yes. Cyber, bio, chem and distillation-style requests route or fall back to Opus 4.8, so in those domains Fable 5 effectively behaves like Opus 4.8. If your workload is concentrated there, paying the Fable 5 premium buys you little additional capability.

What context and output limits do Fable 5 and Opus 4.8 have?

Both Claude Fable 5 and Claude Opus 4.8 offer a 1M-token context window and up to 128K tokens of output. Fable 5's thinking is always on and effort-controlled, and it requires 30-day data retention. Both are available across major clouds including AWS Bedrock, Vertex and Foundry.