Opus 5 ate the default. Fable is waiting.

Anthropic shipped Claude Opus 5 on 24 July 2026 at $5 input and $25 output per million tokens. Claude Fable 5 is still $10 / $50. The launch line is that Opus 5 comes close to Fable’s frontier intelligence at half the price, becomes the default on Claude Max, and is the strongest model on Claude Pro.

The default moved, and the specialty has to justify the extra $30.

Fable was supposed to be the public Mythos-class ceiling. Six weeks later the everyday flagship is an Opus that Anthropic is willing to sell like Opus 4.8.

We have not published a ranked overall score for Opus 5 in the current snapshot. Individual benches have landed in source files. The composite has not. We are not going to invent one to make launch month tidy. Fable 5 is scored. Opus 5 is documented, priced, and still waiting on a ranking-eligible row.

Half the sticker, not half the product

The list price is exact. Fable is 2× Opus on input and output. One million tokens each way is $30 versus $60. Cache and retries stretch that gap. They do not reverse it.

Fast mode buys Fable’s sticker for Opus speed. It does not buy Fable’s job.

Opus 5 Fast mode is the line that people skip. Anthropic says it runs around 2.5× the default speed at twice the Opus base price. That is $10 / $50. You can now buy Fable’s sticker for Opus speed. You cannot buy Fable’s job that way.

The product split underneath the prices is older than this launch. Fable 5, shipped 9 June with Mythos 5, is the public Mythos-class model: higher ceiling, heavier classifiers, $10 / $50. Opus was the production family at $5 / $25. Opus 5 climbs into Fable’s charts without climbing into Fable’s price. The default moved. The specialty has to justify itself again.

What Anthropic measured

Treat the launch post as a first-party claim sheet. The useful rows are the ones that name a bench, an effort setting, and a comparison.

Table 1 Download CSV

Claim (Anthropic, 24 July 2026)	What it is allowed to prove	What it is not
CursorBench 3.2, max effort: within 0.5% of Fable’s peak, half the cost per task	Cost-efficiency on that Cursor harness at max	A universal coding rank
Frontier-Bench v0.1: beats the field, more than doubles Opus 4.8	A large step over the previous Opus on that internal-style run	An independent, ranking-eligible composite
OSWorld 2.0: beats every model at a given cost, beats Fable’s best at about one third the cost	Computer-use efficiency on that protocol	Proof that Fable is obsolete on long computer-use jobs
ARC-AGI 3: about 3× the next-best model	A real jump on a novel-problem test	Transfer to your repo
Behind Mythos 5 on cyber exploit work	The safety split is still real	That Opus 5 is “safer” in the colloquial sense

Anthropic is explicit that Opus 5 does not advance the dual-use frontier the way Mythos does. It comes close to Mythos at finding vulnerabilities and stays well behind at exploiting them. Cyber classifiers are “proportionally less restrictive” than Fable’s. Anthropic expects them to intervene about 85% less often. Flagged Claude.ai / Claude Code / Cowork requests fall back to Opus 4.8. API customers can turn on automatic fallbacks.

That is a product feature. It is also a different risk surface. Fable’s tighter net is a reason to stay, not a footnote.

Customer quotes on the launch page are endorsements. Cursor, Cognition, Zapier, and the rest are describing their evals. They are not our rows.

What we will not flatten

Fable 5’s current mapped snapshot still looks like a ceiling model: SWE-bench Verified 95, SWE-Pro 80, Terminal-Bench 2 84.3, CursorBench 3.2 70.5, OSWorld Verified 85, AA-Briefcase 1574. Those are Fable rows. They are not Opus 5 rows copied across.

Where Opus 5 has appeared in our source snapshots, it is as a named run on a single bench: Terminal-Bench 3, DeepSWE, VoxelBench, EE-Bench, AA evaluation slices. That is how a model enters the building. It is not how it earns an overall rank.

So the honest comparison today is:

• Price: Opus 5 wins, unless you buy Fast mode and land on Fable’s sticker.
• Anthropic’s cost-per-task charts: Opus 5 is the efficiency play they want you to hear.
• Long-horizon / Mythos-class ceiling: still Fable’s job until independent rows say otherwise.
• Cyber and bio posture: different nets. Do not swap them because the coding chart moved.

If your eval is “does this close the ticket,” start on Opus 5. If your eval is “does this survive the weekend on the hard tail,” you do not have permission to fire Fable because a launch chart was close.

Who should still pay for Fable

Pay Fable when you have already measured that Opus 5 drops the cases you care about: long-running refactors, computer-use jobs that sit on OSWorld-class work, or the Mythos-adjacent research Fable was sold to do. Pay Fable when the classifier posture is the feature, not the annoyance.

Pay Opus 5 when you wanted Fable last month and have not measured a failure. The default on Max is a routing hint from the lab that built both models. Take it, then keep a small Fable holdout.

Do not pay Fable prices for Fast Opus unless speed at the Opus quality level is the actual bottleneck. That SKU is a latency product. It is not a ceiling product.

Our Fable 5 vs GPT-5.6 read still holds on the market shape. Capability is not the scarce thing. Access, cost, and which model you are allowed to leave on are. Opus 5 is Anthropic agreeing with itself.

What 5.1 would have to do

Neither Opus 5.1 nor Fable 5.1 has been announced. As of 18 August 2026 the public catalog, pricing page, and release notes still list Opus 5 and Fable 5. Late-July posts, including Andrew Curran’s, claimed a finished Fable increment for August at unchanged $10 / $50, timed against OpenAI. That is a rumor with a plausible cadence behind it. It is not a ship date.

The cadence is real. Opus moved from 4.6 in February to 5.0 on 24 July. Fable and Mythos landed 9 June. Sonnet 5 landed 30 June. Anthropic can produce a point release in the next window. It has not, yet, on this family.

If Fable 5.1 ships, it has one job: reopen a gap worth $10 / $50. The gap has to show up on long-horizon, computer-use, or Mythos-class work that Opus 5 does not win at half price. A 1-point CursorBench bump is not a product. A cheaper Fable would be a different product. The rumor says the price holds. Then the quality delta is the only thing left to sell.

If Opus 5.1 ships, expect the production increment at $5 / $25: fewer tokens per task, a quieter Fast mode, another step on the benches Anthropic already used to justify killing the Fable default. That model would not need to beat Fable. It would need to make the holdout smaller.

The pairing to watch is not “which 5.1 is smarter.” It is whether Anthropic still believes Fable is a family or a luxury SKU it can stop explaining. Opus 5 already ate the default. A silent August is data. It says they are not in a hurry to restore the ceiling, or they are not ready to show it.

Do not wait. Route new work to Opus 5. Keep a measured Fable slice on the jobs that already fail. If a 5.1 appears with a model ID, a price, and independent rows, re-run that slice. Until then, 5.1 is a rumor standing on a family that just lost its everyday reason to exist.