Claude Opus 5 vs Fable 5: the cheaper model wins most of the benchmarks — so what is Fable still for?

Anthropic shipped Claude Opus 5 on July 24, 2026, and described it in one sentence: “a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.”

Read Anthropic’s own documentation the same week and you get a different message. The models overview still calls Claude Fable 5 “Anthropic’s most capable widely released model” and tells you to reach for it “for workloads that need the highest available capability,” while pointing you to Opus 5 for “complex agentic coding and enterprise work.”

Both statements can’t be the practical default. So we did what we did for the Fable 5 launch chart: went to the numbers, matched each one to the model it actually belongs to, and separated the differences that show up on a benchmark from the ones that decide whether you can use a model at all. Sources are at the end.

The 30-second version

1. On the headline benchmarks, the cheaper model wins. Frontier-Bench v0.1: Opus 5 43.3% vs Fable 5 33.7%. SWE-bench Verified: 96.0% vs 95.0%. On OSWorld 2.0, Anthropic reports Opus 5 surpassing Fable’s best result at just over one-third of the cost.
2. Fable’s remaining wins are narrow. SWE-bench Pro 80.0% vs 79.2%, and CursorBench 3.2, where Anthropic itself says Opus 5 at max effort lands within 0.5% of Fable’s peak. These are ties in everything but the ranking.
3. Half the price is not half the bill. One hands-on comparison measured Opus 5 spending ~698K tokens per task vs Fable’s ~537K — about 30% more — with more tool calls and longer runtimes. Final cost on a two-task coding eval: $106.69 vs $118.46. Cheaper, by about 10%, not 50%.
4. The decisive differences aren’t on any chart. Fable 5 forces 30-day data retention and is rejected outright for zero-retention organizations; Opus 5 is not. Opus 5’s knowledge runs four months newer (May 2026 vs January 2026). Opus 5 lets you turn thinking off; Fable doesn’t. Priority Tier works on Fable, not on Opus 5.
5. Fable still has more safety classifiers — a feature if you want the refusals, a tax if your work is security-adjacent. Fable’s own system card recorded safety refusals in 20.9% of Terminal-Bench runs.
6. The default flipped. Opus 5 is now the default model on Claude Max and the strongest on Pro. Fable 5 is no longer the obvious choice for hard work — it’s a specific choice for specific constraints.

The benchmark table

Anthropic published the Opus 5 comparisons as chart images rather than readable tables, which is why most coverage repeated the adjectives instead of the numbers. Here are the numbers, with the source of each marked:

Opus 5 and Fable 5 figures originate in Anthropic’s launch materials and system cards (July 24 and June 9, 2026); the Frontier-Bench, CursorBench and ARC-AGI values are as compiled by third-party analysts reading Anthropic’s chart images, and ARC-AGI 3 by the ARC Prize Foundation. GDPval-AA is deliberately left out: Opus 5 reports on v2 of that evaluation and Fable 5 on the earlier version, so the two Elo figures are not comparable and printing them side by side would invent a result.

Two things stand out. The first is where Opus 5’s gains land: not on the saturated benchmarks — SWE-bench Verified moved a single point, because there is barely a point left to move — but on Frontier-Bench, where it scores 28% higher than the model that costs twice as much, and on ARC-AGI 3, where Anthropic’s claim of “three times the next-best model” is confirmed by the ARC Prize Foundation’s own figures. That’s the pattern of a genuinely stronger model, not of leaderboard tuning, and it’s the same test we applied when Fable itself launched.

The second is what’s missing. Fable 5 has no published ARC-AGI 3 result, so the largest single gap in the table is a gap in the data, not a measured defeat. Per our benchmark policy, we’ll update if that number appears.

Half the sticker price is not half the bill

The pricing page comparison is simple and, on its own, misleading:

Cost per task is price per token multiplied by tokens spent, and the second term isn’t equal. A published hands-on comparison ran both models through a tool-calling evaluation and a two-part coding build, and measured:

Third-party hands-on evaluation, July 2026 — a small sample on one workload, not a benchmark. Directionally consistent with Anthropic’s own latency ratings, which list Fable as “Slower” and Opus 5 as “Moderate” per response.

Opus 5 still came out cheaper, but by about 10%, not the 50% the price list implies — because it worked harder to get there, spending about 30% more tokens across more tool calls. On a single-shot API call the pricing page is the answer. On a long agentic run, where token spend compounds across dozens of tool calls, you should measure your own traces before assuming the discount survives contact with your workload. This is the same trap as Claude Code’s token overhead: the per-token number is real and it is not the number you get billed.

One more line of that table deserves attention on its own. Opus 5’s fast mode costs exactly what Fable 5 costs — $10/$50, for roughly 2.5x the default speed. That’s the same trade the Fable 5 launch presented against Opus 4.8, except the capability argument has now moved to the other side. For the price of Fable, you can have a model that matches or beats it on most evaluations, running two and a half times faster.

The differences that decide it, and they aren’t capability

If you only compare scores you’ll pick the wrong model, because the constraints below rule one out entirely in common situations:

Data retention is the one that ends arguments. Fable 5 is a Covered Model: 30-day minimum retention, and it is not available under zero-retention arrangements at all. An organization with a ZDR agreement that calls claude-fable-5 gets a 400 invalid_request_error — not degraded service, a rejection. Opus 5 carries no model-specific retention requirement. If you work under a ZDR arrangement, this comparison was never close; Fable has never been available to you and Opus 5 is.

The knowledge cutoff is the quiet one. Opus 5’s training data runs to May 2026 against Fable’s January 2026 — the “lesser” model has four months more knowledge of the ecosystem you’re coding against. For an agent working with libraries, framework releases, and APIs that shipped this spring, that gap can matter more than a point of SWE-bench.

Thinking control is a real API divergence. Both models run adaptive thinking by default and neither returns raw chain of thought. But Opus 5 will accept thinking: {type: "disabled"} at effort high or below, which is how you keep latency down on simple, high-volume calls. Fable rejects it — disabled thinking and the older budget_tokens form both return a 400. There is no way to buy a fast, shallow Fable response.

Priority Tier is Fable’s real remaining advantage, and it’s a shrinking one. Priority Tier works on Fable 5 and explicitly does not work on Opus 5 or Sonnet 5. But Anthropic has stopped selling new Priority Tier commitments — existing contracts run to their end date. So this decides the question only for organizations that already bought capacity, and it stops mattering entirely when those contracts expire.

The refusal question cuts both ways

Both models ship safety classifiers that can decline a request. The mechanics are identical: stop_reason: "refusal" returned as an HTTP 200, with a stop_details.category such as cyber, bio, or reasoning_extraction. Input tokens are free if the refusal fires before any output; a mid-stream refusal bills what was streamed. If you’ve built against this already, the handling carries over unchanged — see our Fable 5 refusal handling guide.

The difference is how many classifiers are watching. Fable 5 carries additional ones beyond Opus 5’s cyber category, and that isn’t theoretical: Fable’s own system card recorded safety refusals in 20.9% of Terminal-Bench 2.1 runs, forcing mid-task fallbacks that visibly cost it points. If your work is security-adjacent — reading exploit code, writing detection rules, auditing auth flows — Fable will interrupt you more often, and the extra money buys you the interruptions.

Opus 5 adds one thing Fable’s launch didn’t have: server-side fallback is now opt-in with a single parameter, fallbacks: "default", behind a beta header, which reroutes a cyber-category refusal to another model inside the same call. It doesn’t work on the Batch API or on Bedrock, Google Cloud, or Microsoft Foundry, so self-hosted retry logic still earns its keep there.

For a risk-averse deployment — a product exposed to untrusted end users — Fable’s extra classifiers are a feature you may want to pay for. For an engineer doing defensive security work, they’re a tax. Same mechanism, opposite sign, depending entirely on who’s typing.

So what is Fable 5 still for?

An honest list, and it’s shorter than it was in June:

• Long-running orchestration where efficiency compounds. The measured token and tool-call gaps favor Fable, and on a multi-hour agent run a 30% token difference outweighs a 50% price difference. This is the strongest remaining case, and it’s the one worth testing on your own workload rather than taking on trust.
• An existing Priority Tier commitment. Opus 5 can’t use it. If you’re paying for guaranteed capacity, that capacity only applies to Fable.
• Deployments that want maximum safety filtering. More classifiers, more refusals, fewer ways for an end user to steer the model somewhere you don’t want it.
• Anthropic’s own label. The docs still name Fable the most capable widely released model. If your procurement process cares about “the top model,” that sentence is the reason, not the scoreboard.

Notice what’s not on the list: raw coding capability. That was Fable’s entire pitch in June, and it’s now roughly a tie at double the price. The June verdict said Fable was the first model worth paying double for when one failed attempt costs you an afternoon. Six weeks later that reasoning points at Opus 5 instead — same argument, cheaper answer.

What to actually do

Default to Opus 5. It’s already the default on Claude Max and the strongest model on Pro, it reached Claude Code on launch day, it matches or beats Fable on nearly everything published, it costs half as much per token, it knows four months more, and it works under ZDR. That’s not a close call for general coding and agent work.

Switch to Fable deliberately, for a stated reason — an orchestration workload where you’ve measured the token efficiency paying off, an existing Priority Tier contract, or a deployment that wants the extra classifiers. “It’s the top model” is not a reason; it hasn’t been the top scorer since July 24.

Re-tune before you migrate. Two changes bite on the way from Opus 4.8: requests with no thinking field now run with adaptive thinking, so revisit max_tokens — it’s still a hard cap on thinking plus response text combined. And thinking: {type: "disabled"} combined with xhigh or max effort is now a 400; either re-enable thinking or drop to high. Anthropic’s guidance is to run a fresh effort sweep rather than carry over your Opus 4.8 settings, and to start at 64K max_tokens if you run at xhigh or max.

Measure cost on your own traces. The 2x price gap is real per token and about 10% in one measured real task. Whichever direction your workload lands, the pricing page will not tell you.

Companion reading

• Claude Fable 5 vs Opus 4.8: the June verdict — the launch chart, split by which model each number belongs to
• Fable 5 refusal handling — the stop_reason: "refusal" path and how to build against it
• Fable 5, suspended and restored — the 18-day export-control outage, and what it says about single-vendor risk
• AI agent benchmarks: real signal vs marketing — why SWE-bench moving one point isn’t the story
• Claude Code’s token overhead — the other place the per-token price stops predicting the bill

Sources

1. Introducing Claude Opus 5 — Anthropic
2. Claude Opus 5 System Card (July 24, 2026) — Anthropic
3. Models overview: Fable 5 and Opus 5 specs, cutoffs, latency — Anthropic docs
4. Migrating from Claude Opus 4.8 to Claude Opus 5 — Anthropic docs
5. Service tiers: Priority Tier supported models — Anthropic docs
6. Claude Opus 5 benchmarks compiled from Anthropic’s chart images — Vellum
7. Claude Opus 5 vs Fable 5: coding and tool-call hands-on comparison — Composio
8. Anthropic launches Claude Opus 5 — TechCrunch

FAQ

Is Claude Opus 5 better than Claude Fable 5? On most published benchmarks, yes — at half the sticker price. Opus 5 takes Frontier-Bench v0.1 (43.3% vs 33.7%), edges SWE-bench Verified (96.0% vs 95.0%), and Anthropic reports it passing Fable’s best OSWorld 2.0 result at just over a third of the cost. Fable holds SWE-bench Pro (80.0% vs 79.2%) and CursorBench 3.2, where Opus 5 at max effort lands within 0.5%. Capability is close, Opus 5 wins the hardest new evaluations, Fable is more efficient per task — and Anthropic’s docs still label Fable the most capable model it sells, so the label and the scoreboard now disagree.

Is Opus 5 actually half the price of Fable 5? Half the per-token price, not half the bill. $5/$25 vs $10/$50 per million tokens, but a published hands-on test measured Opus 5 spending ~698K tokens per task against Fable’s ~537K, with more tool calls and longer runtimes. Combined cost across two coding tasks came out at $106.69 vs $118.46 — cheaper by about 10%, not 50%. Budget from your own traces.

Can I use Fable 5 with zero data retention? No. Fable 5 is a Covered Model with a 30-day minimum retention requirement and no ZDR availability; a ZDR organization calling claude-fable-5 gets a 400 invalid_request_error. Opus 5 has no model-specific retention requirement and inherits your org default, so it works under ZDR. In regulated environments this decides the comparison before any benchmark does.

What are the API differences? Four. Thinking: both default to adaptive, but Opus 5 accepts thinking: {type: "disabled"} at effort high or below while Fable rejects it with a 400. Retention: Fable requires 30 days and blocks ZDR; Opus 5 doesn’t. Priority Tier: supported on Fable, not on Opus 5 — though new commitments are no longer sold, so it only affects existing contracts. Knowledge: Opus 5’s training data runs to May 2026 vs Fable’s January 2026. Both share a 1M context window, 128K max output, and identical refusal semantics.

Which should I use in Claude Code? Opus 5 by default — it’s the default on Max, the strongest on Pro, and it landed in Claude Code on day one. Keep Fable for long-running orchestration where its lower token spend compounds, for an existing Priority Tier commitment, or where you want its extra safety classifiers. And note the price math before you switch: Opus 5’s fast mode is $10/$50, exactly Fable’s standard rate, so the same money buys either Fable’s capability or about 2.5x the speed on Opus 5.