Claude Fable 5 vs Claude Opus 5: Capability, Safeguards, Price, and Practical Differences
Anthropic shipped a cheaper model that beats its own more expensive one on most published benchmarks — which inverts the usual logic of model tiers and makes the Fable 5 premium hard to justify for mainstream work.
........
Fable 5 launched June 9, 2026 at $10 per million input tokens and $50 output; Opus 5 followed on July 24, 2026 at exactly half that.
Opus 5 wins five of nine directly comparable rows in Anthropic's own launch tables, including Frontier-Bench, GDPval-AA, BrowseComp, OSWorld 2.0, and AutomationBench.
Fable 5 retains published leads on DeepSWE v1.1 and the Legal Agent Benchmark.
Fable 5's cyber classifiers intervene roughly 85% more often than Opus 5's, and blocked requests fall back to an older model.
Opus 5 is now the default model on Claude Max, and posted Anthropic's best score on its automated misaligned-behaviour audit.
··········
WHAT EACH MODEL ACTUALLY IS.
These are not two sibling point-releases. They come from different places in Anthropic's lineup and were built to different design briefs.
Claude Fable 5 is the generally available, production-safeguarded deployment of the same underlying model as Claude Mythos 5 — the restricted tier Anthropic positions above Opus entirely. Fable is what the public gets: identical capability, wrapped in safety classifiers that Mythos does not carry.
Claude Opus 5 sits one tier below and is the everyday flagship. Anthropic released it on July 24, 2026 and positioned it explicitly as the default rather than the ceiling.
So the comparison is a workhorse model against a deliberately gated frontier model — and, critically, the frontier model degrades to workhorse behaviour precisely in the domains Anthropic treats as sensitive.
That relationship, rather than any single benchmark row, is what makes the choice interesting.
··········
PRICING.
The price gap is clean and unambiguous: Fable 5 costs exactly twice what Opus 5 costs, on both input and output.
Fable 5 lists at $10 per million input tokens and $50 per million output tokens.
Opus 5 lists at $5 per million input tokens and $25 per million output tokens — the same price the previous Opus generation carried.
Both models list the same 1 million token context window, which for Opus 5 is both the default and the ceiling.
Worked examples from independent cost modelling put the practical difference in perspective: for a standard chat preset, $0.035 on Fable 5 against $0.0175 on Opus 5; a repository review at $0.65 against $0.325; a cache-heavy agent loop at $0.90 against $0.45. The ratio holds at exactly 2:1 across workload shapes.
··········
Pricing and specs
Claude Fable 5 | Claude Opus 5 | |
|---|---|---|
Released | June 9, 2026 | July 24, 2026 |
Input per 1M tokens | $10.00 | $5.00 |
Output per 1M tokens | $50.00 | $25.00 |
Context window | 1M tokens | 1M tokens |
Tier | Mythos-class (safeguarded) | Opus-class |
Default on Claude Max | No | Yes |
··········
THE BENCHMARK PICTURE.
On Anthropic's own launch tables, the cheaper model takes most of the directly comparable rows — and its margins are larger than Fable's where it wins.
Opus 5 leads on Frontier-Bench, GDPval-AA, BrowseComp, OSWorld 2.0, and AutomationBench. On Frontier-Bench v0.1 specifically, Opus 5 surpasses other models and more than doubles Opus 4.8's score at a lower cost per task.
On CursorBench 3.2 at maximum effort, Opus 5 comes within 0.5% of Fable 5's peak score — at half the cost per task.
Fable 5 holds published leads on DeepSWE v1.1 and the Legal Agent Benchmark, and some of the hardest SWE-bench Pro rows still favour Fable or Mythos-tier models, though by small margins.
Independent aggregation lands them almost on top of each other: one composite puts Fable 5 at 83.32 against Opus 5's 83.24, with overlapping 90% score intervals — a lead, not a settled result.
··········
A CAVEAT THAT CHANGES HOW TO READ THE TABLES.
Several headline figures published beside Fable 5 are explicitly Mythos 5 numbers rather than measurements of the safeguarded public model.
Because Fable and Mythos share weights, it is tempting to treat their results as interchangeable. They are not, for the purposes of anyone actually buying Fable 5.
The safeguards are the difference between the two products, and safeguard interventions change what the model does on a real request. A Mythos benchmark score is a measurement of the unrestricted configuration.
Anyone comparing published numbers should check which configuration each row was measured on before treating it as a Fable 5 result.
··········
SAFEGUARDS: THE REAL DIVIDING LINE.
The safeguard posture is where the two models genuinely diverge, and it cuts in a direction that surprises people.
Fable 5's safeguards route requests in cybersecurity, biology, chemistry, and model distillation to a fallback on an older, less capable model. In those domains, a Fable 5 customer is paying the premium rate for an answer produced by a cheaper model.
Opus 5's cyber classifiers intervene roughly 85% less often than Fable 5's. When they do fire — inside Claude AI, Claude Code, and Claude Cowork — requests fall back to the previous Opus generation.
Opus 5 also carries no mandatory retention requirement for general API access, consistent with prior Opus models.
There is a design choice worth knowing behind the cyber behaviour: Opus 5 is nearly as capable as Anthropic's top tier at finding software vulnerabilities, but was deliberately not trained to exploit them, so it lags well behind on converting findings into working attacks. For legitimate defensive security work, that is a meaningful distinction — vulnerability detection stays available, exploitation does not.
··········
ALIGNMENT SCORES.
Opus 5 posted Anthropic's strongest result to date on its automated misaligned-behaviour audit.
Launch materials put Opus 5 at 2.3 overall on that audit, which is lower — meaning better — than Opus 4.8, Sonnet 5, or Fable 5.
That is an unusual result for a model positioned as the cheaper, more permissive option, and it undercuts the intuition that heavier safeguards necessarily mean a better-behaved model.
The safeguards on Fable 5 govern which requests get answered. The alignment audit measures how the model behaves on the requests it does answer. Those are separate properties, and Fable's stricter posture on the first does not translate into a better result on the second.
··········
SPEED, AND A DISAGREEMENT ABOUT IT.
Latency is the one dimension where vendor labels and independent measurement point in opposite directions.
Anthropic lists Fable 5's comparative latency as slower and Opus 5's as moderate.
Artificial Analysis measured the reverse on throughput, clocking Fable 5 with fallback at 73 output tokens per second against 52.6 for Opus 5.
Opus 5's clearer speed advantage is its Fast mode, which Fable 5 does not offer at all.
There is also a token-consumption wrinkle: some reports note Opus 5's token usage can run higher than expected on certain tasks, which can narrow the practical cost gap below the clean 2:1 the rate card suggests. That is a reason to measure cost per completed task on real workloads rather than trusting the ratio.
··········
WHERE FABLE 5 STILL MAKES SENSE.
Fable 5 is not a bad model. It is a specialist one, and the specialisms are narrow but real.
The clearest case is the longest autonomous runs — multi-day work in Claude Code or managed agent setups that plan, verify, and self-correct across days without human steering. Fable retains an edge at the frontier of that kind of long-horizon operation.
Where its published leads on DeepSWE v1.1 or the Legal Agent Benchmark map directly onto your workload, that lead is a legitimate reason to pay double.
The advisor pattern is another: Fable 5 acting as a strategic planner over a cheaper executor model such as Sonnet 5, where the expensive model produces a small number of high-leverage tokens and the cheap one does the volume.
For biology and cyber frontiers, Mythos 5 remains ahead on dual-use capability, and Fable 5 is the top generally available tier below it — but in exactly those domains the safeguard fallback is most likely to fire.
··········
THE ROUTING RECOMMENDATION.
For most production teams the sensible policy is Opus 5 as the premium default and Fable 5 as a measured exception.
Opus 5 matches or beats Fable on coding, wins the knowledge-work Elo benchmark, does it for half the price, and carries the better alignment score, all while being the default on Claude Max.
Outside the narrow slice where Fable's published leads apply, paying double for Fable means paying double to score slightly lower on the benchmarks that matter most for coding and analysis.
The stronger design is not picking one model permanently but writing a versioned routing policy: defined task classes, acceptance checks, escalation limits, checkpoints, and a rollback path.
That policy is worth revisiting when independent evaluations land, when Anthropic changes Fable's retention terms or safeguard behaviour, or when your own workload mix shifts.
What Fable 5's price is really buying, on the evidence available, is its safeguard posture and its Mythos lineage — not additional capability on the work most teams actually do.
··········
·····
FOLLOW US FOR MORE.
·····
·····
DATA STUDIOS
·····
