Fable 5 vs Opus 5: Why I Demoted the Model I Just Wrote a Whole Post About

A month ago I switched my agent harness’s default model to Fable 5. On July 24, I switched it again, this time to Opus 5, at half the price, and I have not switched back. That is the whole story compressed into two sentences. Here is the routing logic behind it, and the two benchmarks where I did not switch.

TL;DR

Opus 5 became my harness’s main-loop default on 2026-07-24, superseding a Fable 5 default that had held the seat since July 8. My own routing notes put Opus 5 ahead or tied with Fable 5 on 9 of 13 shared benchmarks, at $5 input / $25 output per million tokens against Fable’s $10 / $50, exactly half. Fable 5 kept its seat on exactly two benchmarks in my doctrine: SWE-bench Pro and DeepSWE, both tied to single hard code diffs. Everywhere else, Opus 5 matches or beats it for less, so Fable 5 is now a narrow reservation, not the ceiling.

Which model should you route to, Fable 5 or Opus 5?

Route long-horizon agentic work, computer use, browsing, and multi-tool chains to Opus 5 by default. Reserve Fable 5 for the specific case my routing doctrine still shows it winning: a single hard code diff where the alternative model gets it wrong on the first pass and a retry is expensive.

Workload	My default	Why
Interactive main-loop work, architecture, review, synthesis	Opus 5	Took the main-loop seat 2026-07-24; ahead or tied on 9 of 13 shared benchmarks at half Fable’s price
A single hard code diff, novel and expensive if wrong	Fable 5, gated	The two benchmarks in my doctrine where Fable still leads (SWE-bench Pro, DeepSWE)
Spawned subagent work, bulk builds	Sonnet 5	Cheaper, pinned as the harness’s subagent default; Opus 5’s over-delegation habit makes it a worse spawn target for routine work
Judgment stuck between models	Opus 5 first	Tiebreaker in my doctrine is intelligence, then taste, then cost; Opus 5 and Fable 5 tie on the first two, so cost decides

Why did Opus 5 replace Fable 5 as the default, not the other way around?

Because the axes that are supposed to separate a flagship model from its predecessor, intelligence and taste, came back tied in my own routing doctrine, and the axis that actually differed, cost, favored Opus 5 by exactly half. When two models tie on the axes that should decide the call, the tiebreaker in my doctrine is intelligence first, then taste, then cost, in that order, and cost is what actually moved.

I want to be precise about what kind of evidence that is. This is not a benchmark reproduction. It is the internal routing table I maintain for my own agent harness, scored against how I actually use these models, not against a public leaderboard. My Fable 5 vs Opus 4.8 post carries the system-card benchmark numbers, sourced directly to Anthropic’s published figures, for the prior model generation. This post carries something different: first-party routing observations from running both current models daily in production, framed as one operator’s harness doctrine, not a benchmark claim.

My doctrine scores both models at the top on the two axes that matter most:

Model	Intelligence	Taste	Cost-here	Seat
Opus 5	5	5	Half of Fable, matches Opus 4.8 pricing	Main-loop default (2026-07-24)
Fable 5	5	5	Metered, roughly 2x Opus	Narrow reservation, gated

Cost-here is not list price. It is what a model actually costs to run against my usage pattern, and on that axis Fable 5 lost the tie. Opus 5 lists at $5 input and $25 output per million tokens, identical to Opus 4.8’s pricing and exactly half of Fable 5’s $10 and $50. Fable 5 also runs metered in my harness, which compounds the sticker gap into a real operating cost, not just a rate-card one.

Where does Fable 5 still win?

On exactly two benchmarks in my own routing doctrine, and both point at the same kind of task: a single hard code diff. SWE-bench Pro has Fable 5 at 80.0 against Opus 5 at 79.2. DeepSWE has Fable 5 at 69.7 against Opus 5 at 68.8. Both gaps are inside a point and a half. That is not the wide, easy-to-defend lead Fable 5 held over Opus 4.8 in the prior generation, where it beat Opus 4.8’s best tier at its own lowest effort setting. Against Opus 5, Fable’s edge survives only where the work is novel, expensive to get wrong, judgment-heavy rather than volume-heavy, and shaped like one hard diff rather than a long agentic run. My harness gates Fable spawns behind exactly that four-condition test, and the gate exists because the margin no longer covers casual use.

The honest reading: Opus 5 closed most of the gap that justified Fable 5’s premium in the first place. What is left is a reservation, not a ceiling.

What actually changed in Opus 5’s behavior, not just its benchmarks?

Opus 5 reverses several habits that Fable 5 and Opus 4.8 trained me to expect, and every one of them is a direction reversal, not a magnitude change. Sourcing caveat first: these deltas come from Anthropic’s release notes, and my own harness telemetry is not yet large enough to grade them independently (a fair grade needs a few thousand turns, and I abandoned an underpowered attempt at n=115). What I can verify first-hand is that my guardrails had to flip direction. If your prompting scaffolding was tuned for the prior generation, it is now pointed the wrong way on Opus 5.

• It self-verifies without being asked. Verification instructions I wrote to compensate for Opus 4.8’s habit of skipping checks are now redundant on Opus 5. I kept the external evidence gates, the artifact-or-it-did-not-happen rule, because those catch a different failure than prompt scaffolding does. But the reminder-to-double-check line in my prompts is now dead weight.
• It over-delegates. Opus 4.8 under-reached on spawning subagents when a task called for it. Opus 5 goes the other way, reaching for a subagent spawn more often than the task needs. My routing rule now includes an explicit do-not-spawn check for anything three files or fewer, anything where a spawn’s coordination overhead exceeds just doing the edit inline, and anything that is the judgment call itself rather than delegatable execution.
• Its responses are longer, and effort does not fix that. Lowering the effort setting does not reliably shorten Opus 5’s visible output the way it does on other models. Only an explicit conciseness instruction cuts it, by roughly 20% per the release notes. If your interface is getting verbose Opus 5 replies, the effort knob is not the lever, the prompt is.
• It expands task scope. Left unconstrained, Opus 5 tends to do more than was asked, not less. The counterweight in my harness is an explicit rule that the requested scope is the deliverable, not a floor to build past.
• Its API defaults changed underneath prior code. Thinking now defaults on when the thinking parameter is omitted from a call, a silent reversal from 4.8. Both budget_tokens and an explicit thinking: {"type": "disabled"} return 400 errors at effort xhigh or max. Code written against 4.8’s defaults will not port cleanly.

None of these are reasons to avoid Opus 5. They are reasons to point your guardrails in the direction the model actually drifts, instead of the direction the previous model drifted.

How should you set effort on Opus 5, low through max?

Start at the official default effort tier and move up only for judgment-heavy turns, the same discipline that already applied to Fable 5. Fable 5’s own workflow guidance, distilled from builders running it against real workloads, makes the underlying mechanism explicit: reasoning effort is spent per tool call and per change, not against total run length. A long, multi-step task does not get more steps out of a higher effort setting, it gets more thinking spent on each individual step, and most steps do not need it. The people burning through usage fastest were consistently the ones running at xhigh or max as a habit rather than a deliberate escalation.

That mechanism is a property of how the effort dial works, not something specific to one model generation, so I apply the same instinct to Opus 5: default effort, escalate deliberately, and treat max as a fan-out tool rather than a default setting. I do not yet have Opus-5-specific effort-tier benchmark data the way the prior post had system-card figures for Fable 5 versus Opus 4.8; what I have is the operating discipline, and it has held across both model generations so far.

What changed in my own agent stack on July 24

I re-routed my harness’s main-loop default from Fable 5 to Opus 5 the day it shipped, and I did not touch the spawned-subagent tier, which stays pinned to Sonnet 5 for cost reasons unrelated to this comparison. Fable 5 moved from the seat it held for sixteen days to a narrow reservation gated behind a four-condition test: the work has to be novel, expensive if wrong, judgment-heavy rather than volume-heavy, and long-horizon or tightly targeting-bound. Most turns do not clear that bar, so most turns run on Opus 5 now.

What I did not do: I did not treat this as a verdict against Fable 5. The prior post’s numbers still stand for the prior generation, Fable 5 at low effort genuinely did beat Opus 4.8 at its highest tested tier on SWE-bench Pro. What changed is the comparison target. Opus 5 closed most of that gap at half the price, and the two benchmarks where Fable 5 still leads are narrow enough that keeping it as a default would now be the wrong call, not a cautious one.

For a setup-focused walkthrough of the new model, how to use Claude Opus 5 covers the migration mechanics. For the operating system this routing logic sits inside, see the broader Claude Code production workflow. For the system-card-sourced benchmark comparison this post’s routing doctrine descends from, the Fable 5 vs Opus 4.8 post has the page-cited numbers from the prior model generation.

What to do next

Run the same one-week test I ran before trusting either model’s reputation: pick a real task from your backlog, write one complete brief, and hand it to both models at their default effort setting. Track cost per completed task and how many times each needed a rescue, not cost per token. If your harness has a routing doctrine of its own, revisit it now. Mine flipped in sixteen days. Yours might need to.

Sources

• Introducing Claude Opus 5 (Anthropic launch announcement, July 24, 2026)
• Introducing Claude Fable 5 and Mythos 5 (Anthropic launch announcement, June 9, 2026)
• My own model-routing doctrine, maintained as first-party harness documentation and updated as behavioral evidence accrues; the benchmark figures cited above (SWE-bench Pro, DeepSWE, the 9-of-13 shared-benchmark count) are drawn from that internal reference, not reproduced from a public leaderboard