Claude Opus 5 vs Fable 5: Benchmarks, Price, Verdict

When Anthropic shipped Claude Opus 5 on July 24, 2026, the pitch was unusually blunt for a frontier lab: a model that "comes close to the frontier intelligence of Claude Fable 5 at half the price." That is a claim you can check against numbers — and six weeks later, with Anthropic's own head-to-head evals published, it holds up better than the company's own phrasing suggested.

This is the buyer's-eye comparison. Not a spec dump — a decision. When do you reach for the cheaper Opus 5, and when is Claude Fable 5 still worth double the money?

What are Claude Opus 5 and Claude Fable 5?

They sit in different tiers of Anthropic's lineup, and the tiers are real rather than marketing.

Claude Opus 5 (claude-opus-5, released July 24, 2026) is Anthropic's model for complex agentic coding and enterprise work. It ships with a 1M-token context window as both default and maximum, 128K max output, adaptive thinking on by default, and a five-level effort toggle (low through max) that trades reasoning depth against latency and cost per request. On the Message Batches API it goes up to 300K output tokens with a beta header.

Claude Fable 5 (claude-fable-5, released June 9, 2026) is the frontier tier above Opus, built for demanding reasoning and long-horizon agentic work — agents that run for hours. Same 1M context and 128K output, but thinking is always on and cannot be disabled.

One structural fact changes how you should read everything below: Fable 5 is now a legacy model. Claude Fable 5.1 shipped on September 1, 2026 and took over as the current Fable release. Fable 5 remains fully supported — Anthropic commits to retiring it no sooner than June 9, 2027 — but it is no longer the top of the ladder. Above Opus 5 today sits Fable 5.1, not Fable 5.

Is Claude Opus 5 really half the price of Claude Fable 5?

Yes, on both input and output, with no asterisk on the headline rates.

Every line halves. Opus 5 inherited Claude Opus 4.8's pricing unchanged, and Fable 5 sits at exactly 2× that across the board. Rates are from Anthropic's pricing page.

Two wrinkles worth knowing before you model your bill:

• Opus 5 has a fast mode — a research preview on the Claude API only, priced at $10 / $50 per million, i.e. exactly Fable 5's rate, for around 2.5× the default output speed. It is not available on Amazon Bedrock, Google Cloud, or Microsoft Foundry, and not on the Batch API. So at Fable 5's price point you have a real choice: Fable 5, or Opus 5 running considerably faster.
• Fable 5.1 changes the cache maths, not the headline. Fable 5.1 keeps $10 / $50 but prices cache reads at $0.25 per million rather than $1.00 — 0.025× base input instead of the usual 0.1×. That is the one place a Fable model undercuts Opus 5 ($0.25 vs $0.50 per cached million), and on a long agent loop re-reading a big cached prefix it matters. Against Fable 5 specifically, though, Opus 5 is cheaper on every single line.

For a deeper look at using the effort toggle to cut spend within a single model, see our guide to Claude Opus 5 pricing and the effort toggle.

How do Claude Opus 5 and Claude Fable 5 compare on benchmarks?

The cleanest source is the eval set Anthropic published with the Fable 5.1 announcement, which ran the same benchmarks across Fable 5.1, Fable 5, Opus 5, and GPT-5.6 Sol. Same harness, same day, no cross-vendor cherry-picking.

Source: Anthropic's Claude Fable 5.1 and Mythos 5.1 announcement.

Opus 5 takes six of nine, and its wins are the wider ones. Terminal-Bench 4.0 — agentic coding, the closest proxy to real engineering work — is a 10-point gap in Opus 5's favour. AutomationBench, which measures end-to-end business workflows, is nearly 10 points. GDPval-AA v2, the human-graded knowledge-work Elo, is 101 points.

Fable 5's three wins are all inside two points, and two of them are inside half a point. On CursorBench and Humanity's Last Exam the two models are, for practical purposes, tied — at double the price.

Anthropic's own Opus 5 announcement adds a few figures the table doesn't carry: Opus 5 "surpasses all other models" on Frontier-Bench v0.1 and more than doubles Opus 4.8 at a lower cost per task; on CursorBench at max effort it lands within 0.5% of Fable 5's peak score at half the cost per task; on ARC-AGI-3 it scores roughly three times the next-best model; and on Zapier's AutomationBench its pass rate is around 1.5× the next-best model at the same cost per task.

The framing to take away is not "Opus 5 is nearly as good." On evals Anthropic ran itself, Opus 5 is better than Fable 5 more often than not, and costs half as much. If you are evaluating models for a coding-agent stack specifically, our AI coding agents guide covers how these scores translate into day-to-day agent reliability, and Claude Opus 5 benchmarks explained unpacks the individual evals.

Where does Claude Fable 5 still beat Claude Opus 5?

Honesty matters here, because "half the price and better numbers" reads as "never buy the expensive one." Three genuine carve-outs remain.

• Long-horizon autonomous agents. This is the explicit positioning. Anthropic's model-selection guidance routes "agent sessions that run for hours, multistep deep research, analysis carried through to a finished document, spreadsheet, or deck" to the Fable tier. Benchmarks are turn-limited by construction and under-measure exactly this. If your product is a multi-day autonomous worker rather than a request/response assistant, test rather than assume.
• The narrow reasoning edge. Fable 5's leads on CursorBench and Humanity's Last Exam are small, but they are real and they sit on the hardest problems. If your workload lives at the top of that difficulty band, a sub-point margin can compound across a long trajectory.
• Cybersecurity work, with a caveat. Opus 5 remains behind Claude Mythos 5 on cybersecurity tasks — but note that Mythos 5 is Fable 5 without the safety classifiers, invitation-only through Project Glasswing, not a higher tier. Plain Fable 5 carries the same classifiers that make this awkward.

For everything else — most coding, most agentic workflows, most knowledge work, computer use, business automation — Opus 5 is the rational default at half the cost.

How do the two API surfaces differ?

This is where a "just change the model string" migration goes wrong. Four differences bite.

Data retention is the one that decides procurement conversations. Fable 5 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorises it — an organisation configured for ZDR gets a 400 invalid_request_error on every Fable 5 request no matter how valid the payload. Opus 5 has no such requirement for general access. For teams with client contracts restricting where prompt data lingers, that is a concrete reason to prefer Opus 5 independent of price or benchmarks.

Both models can refuse. A declined request returns a successful HTTP 200 carrying stop_reason: "refusal" and a stop_details object naming the policy area, not an error — so code that indexes response.content[0] without checking stop_reason will break on either model. Anthropic says Fable 5's classifiers (cybersecurity exploitation, bio/chem, distillation) trigger in under 5% of sessions. Ship the fallbacks parameter by default on both: fallbacks: "default" retries server-side on the model Anthropic recommends for that refusal category, and fallback credit refunds the prompt-cache cost of the switch. You are not billed for a refusal that arrives before any output.

Does Claude Fable 5.1 change this comparison?

It changes which Fable you should be comparing against, and it narrows Opus 5's advantage without erasing it.

Fable 5.1 keeps $10 / $50 but improves on Fable 5 across every published benchmark, and the gains land squarely on the evals Opus 5 was winning: Terminal-Bench 4.0 goes 42.0% → 55.8% (past Opus 5's 52.3%), GDPval-AA v2 goes 1,723 → 1,853 (past 1,824), AutomationBench 17.1% → 31.4% (past 26.9%), Terminal-Bench-Science 24.7% → 52.6% (past 29.0%). It also cuts cache reads to $0.25 per million — below Opus 5's $0.50 — which is worth roughly 25% off a typical workload and ~45% off a heavily agentic one versus Fable 5.

So the "Opus 5 wins on the numbers" result in this article is specific to Fable 5, and it is largely a consequence of Fable 5 predating Opus 5 by six weeks. Against Fable 5.1 the ranking flips back and the decision returns to a straight price-versus-capability call. That comparison is its own piece: Claude Fable 5.1 vs Claude Opus 5.

Which model should you actually pick?

A short decision rule:

• Default to Claude Opus 5 for coding assistants, request/response agents, knowledge work, computer use, business automation, and any high-volume workload where a halved rate compounds. Against Fable 5 you are paying half as much for better scores on six of nine published evals, plus no data-retention requirement.
• Tune effort before you switch models. Anthropic's own guidance is that the effort parameter is usually a better lever than changing model. Start at the high default and move up or down against your evals.
• Use Opus 5 fast mode when you were prepared to pay Fable 5's rate anyway. Same $10 / $50, roughly 2.5× the output speed, still ahead on most benchmarks. Claude API only.
• Move up to Claude Fable 5.1, not Fable 5, when your Opus 5 evals at xhigh or max effort still fall short on demanding reasoning or genuinely long-horizon autonomous work. 5.1 is the current Fable release, scores higher than Fable 5 everywhere, and costs the same at the headline with cheaper cache reads.
• Stay on Claude Fable 5 only for a specific reason — a pinned eval baseline, a compliance review tied to that snapshot, or an integration you can't yet adapt to 5.1's thinking-block binding rules.

The honest summary is that Fable 5 has been squeezed from both sides. Opus 5 undercuts it on price and beats it on most benchmarks; Fable 5.1 matches its price and beats it on all of them. It remains a solid, fully supported model with a year of committed life ahead of it — it is just no longer the answer to "which one should I start with?" for almost anybody.

If you are also weighing Anthropic against OpenAI, our GPT-5.6 vs Claude Fable 5 comparison sits alongside this one — note that on the same eval set Opus 5 beats GPT-5.6 Sol on Terminal-Bench 4.0 (52.3% vs 37.3%), GDPval-AA v2 (1,824 vs 1,711), CursorBench (70.0% vs 67.2%), and AutomationBench (26.9% vs 19.6%).

FAQ

Is Claude Opus 5 really half the price of Claude Fable 5?

Yes, on every line. Opus 5 is $5 per million input tokens and $25 per million output; Fable 5 is $10 and $50. Cache writes, cache reads, and batch rates all halve too. Opus 5's optional fast mode is priced at $10 / $50 — the same as Fable 5 — for roughly 2.5× the output speed.

Is Claude Opus 5 actually as good as Claude Fable 5?

On Anthropic's own published head-to-head it is better on most evals. Opus 5 leads Terminal-Bench 4.0 (52.3% vs 42.0%), GDPval-AA v2 (1,824 vs 1,723), AutomationBench (26.9% vs 17.1%), OSWorld 2.0, and Terminal-Bench-Science. Fable 5 holds sub-two-point leads on CursorBench and Humanity's Last Exam.

When should I still use Claude Fable 5 instead of Opus 5?

Rarely, and usually you should be considering Claude Fable 5.1 instead. The genuine cases for the Fable tier are long-horizon autonomous agents running for hours, multistep deep research, and workloads at the very top of the reasoning-difficulty band where Opus 5 at max effort still falls short on your evals.

Does Claude Opus 5 have the same data retention policy as Claude Fable 5?

No. Claude Fable 5 requires 30-day data retention and is not available under zero data retention without express authorisation from Anthropic — a ZDR-configured organisation gets a 400 error on every request. Claude Opus 5 has no data retention requirement for general access.

What is Claude Opus 5's effort toggle?

A per-request parameter (low, medium, high, xhigh, max) inside output_config that controls how much reasoning compute the model spends. It defaults to high. Anthropic recommends tuning effort before switching models, since it trades intelligence against latency and cost within one model.

Is Claude Fable 5 still available in 2026?

Yes. Fable 5 is listed as Active (legacy) after Claude Fable 5.1 replaced it on September 1, 2026. It works normally on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry, and Anthropic commits to retiring it no sooner than June 9, 2027.

Can Claude Opus 5 and Claude Fable 5 refuse requests?

Both can. A refusal returns HTTP 200 with stop_reason: "refusal" and a stop_details category, not an error, so check stop_reason before reading content. Fable 5's classifiers cover cyber exploitation, bio/chem, and distillation, and fire in under 5% of sessions. Use the fallbacks parameter to retry server-side.

What is the context window on Claude Opus 5 and Claude Fable 5?

Both offer 1,000,000 tokens as the default and maximum, billed at standard per-token rates across the full window, with 128,000 max output tokens per request. Opus 5 additionally supports up to 300,000 output tokens on the Message Batches API with a beta header. Both have a 512-token minimum cacheable prompt.