Anthropic now sells a pricier flagship that its own cheaper model beats on many tasks. Claude Fable 5 (June 9, 2026) sits a tier above Opus 4.8 at $10 per million input tokens and $50 per million output. Claude Opus 5 (July 24, 2026) is smaller, cheaper — $5/$25, exactly half — and on several benchmarks it tops Fable 5. So the practical question for anyone building on these models is not "which is better." It's "which fits this job," and the honest answer is a two-line decision rule plus one discipline that outlives both models.
Here is the rule, stated plainly, before the reasoning:
- Reach for Fable 5 when the task is among the genuinely hardest and longest-horizon — multi-day autonomous agentic coding with minimal intervention — where Anthropic itself says Fable 5 still leads, and you can accept its mandatory 30-day data retention.
- Reach for Opus 5 for everything else, which is most work.
Everything below is the justification, the arithmetic, and the one caveat that makes the whole exercise trustworthy.
The prices, done as arithmetic rather than adjectives
Fable 5 is $10/M input and $50/M output — Anthropic describes this as exactly double Opus 4.8. Opus 5 is $5/M input and $25/M output, exactly half of Fable 5. On a task that reads 200K tokens of context and writes 20K tokens of output, that is Fable 5 at (0.2 × $10) + (0.02 × $50) = $3.00, versus Opus 5 at (0.2 × $5) + (0.02 × $25) = $1.50. Same call, half the bill.
The per-token halving still understates the real gap, because Opus 5 also thinks less to get there. One customer report cited by secondary coverage puts Opus 5 at roughly one-seventh the reasoning tokens and half the latency of Opus 4.8. Reasoning tokens are billed as output. If Opus 5 reaches an answer with a fraction of the internal deliberation, its cost per completed task drops below what the sticker price alone implies. That mechanism is why third-party reports can claim Opus 5 surpasses Fable 5's best on OSWorld 2.0 "at just over a third of the cost" even though its list price is half — fewer thinking tokens compound with the lower rate.
Two pricing details matter for how you actually deploy Opus 5. Fast Mode runs it at about 2.5x speed for $10/$50 — identical per-token pricing to Fable 5. Turn it on only when latency is a real product constraint, because it hands back the entire cost advantage. In the other direction, the Batch API runs at $2.50/$12.50 (with 300K max output in beta), so background and offline work should default there. Cache economics reinforce the pattern: cache hits are $0.50/$1, a 5-minute cache write is $6.25/$12.50, a 1-hour write is $10/$20. If your workload re-reads a stable context — a codebase, a policy corpus — caching moves the effective input cost close to the floor.
Where Fable 5 still earns its premium
The premium is not imaginary; it is narrow. Fable 5 was built for long-horizon autonomy — extended agentic coding over "days" with minimal human intervention — and on the hardest agentic coding it leads by a real margin. On SWE-Bench Pro (third-party), Fable 5 scores 80.3% against Opus 4.8's 69.2%, GPT-5.5's 58.6%, and Gemini 3.1 Pro's 54.2%. It posts 95.0% on SWE-bench Verified and 64.5% on Humanity's Last Exam with tools. Anthropic is explicit that Fable 5 still leads on the hardest tasks, and that Opus 5's biology safeguards — while less restrictive than Fable 5's — still limit long-running autonomous research.
So the honest Fable 5 case is specific: a task at the frontier of difficulty and duration, where a multi-day agent must run largely unattended and a few extra percentage points of reliability compound across hundreds of steps. In that regime the $10/$50 rate buys a genuine edge, because the alternative is a human babysitting the run. Outside it, you are paying flagship rates for capability you won't exercise.
There is a non-price cost to Fable 5 that belongs in the decision: mandatory 30-day traffic retention, with no opt-out. Sensitive queries — cyber exploitation, bio/chem weapons, model distillation — are blocked and rerouted to the safer Opus 4.8, expected in under 5% of sessions. If your data governance forbids retention, Fable 5 is off the table regardless of benchmarks.
Where Opus 5 wins, and how much to trust each number
Opus 5's case rests on three legs, and they are not equally solid — attribution is the whole game here.
The vendor's own numbers (treat as claims). On Anthropic's internal Frontier-Bench v0.1, Opus 5 scores 43.3% versus Fable 5's 33.7%, said to more than double Opus 4.8 and surpass all other models. Anthropic also reports SOTA on GDPval-AA v2 (1,861 points, highest tested) and Life Sciences gains over Opus 4.8 (organic chemistry +10.2 points, protein function +7.7). These are Anthropic's benchmarks. A cheaper model topping the pricier flagship on the vendor's own composite is exactly the result that should make you inspect the benchmark, not cheer — internal composites are marketing until someone independent reproduces them.
Third-party numbers (more weight, still early). On CursorBench 3.2, Opus 5 lands within 0.5% of Fable 5's peak at half the cost; on OSWorld 2.0 it surpasses Fable 5's best at just over a third of the cost; on ARC-AGI 3 it reportedly scores about 3x the next-best model; on Zapier AutomationBench it hits a 100% pass rate. These come from secondary coverage rather than Anthropic's own materials, so treat them as directional. But the pattern — parity-or-better with Fable 5 at a fraction of the cost across independent harnesses — is consistent enough to plan around.
The capability differences (concrete, mechanistic). Opus 5 adds adaptive thinking (configurable effort, default high on the API and Claude Code) and, more usefully, self-verification: it is much stronger at checking its own work and iterating until it succeeds — for example, opening pages in a browser at desktop and mobile widths to catch layout bugs. For agentic work, a model that catches and recovers from its own errors changes the economics more than a one-shot accuracy point does, because failed runs are what actually cost you. It also gains mid-conversation tool changes (beta) and Automatic Fallbacks (beta), which route requests to less powerful models when safety triggers fire — a softer failure mode than a hard block.
On safety and governance, Opus 5 is the more deployable model for most teams: an alignment score of 2.3 on overall misaligned behavior (the lowest, best, of recent Claude models), roughly 85% fewer cybersecurity classifier interventions than Fable 5 (fewer false-positive blocks on legitimate security work), and support for zero data retention where Fable 5 requires 30 days. It's the default on Claude Max and the strongest model on Claude Pro, with a May 2026 knowledge cutoff and the same 1M-token context / 128K max output as Fable 5.
The comparison, compact
| Fable 5 | Opus 5 | |
|---|---|---|
| Released | June 9, 2026 | July 24, 2026 (claude-opus-5) |
| Price (input / output per M) | $10 / $50 | $5 / $25 (Fast Mode $10 / $50, ~2.5x speed) |
| Context / max output | 1M / 128K (output third-party reported) | 1M / 128K (300K via Batch beta) |
| Positioned for | Multi-day autonomous agentic coding; hardest, longest tasks | Best cost/performance for the broad middle |
| Still leads on | Hardest long-horizon tasks (per Anthropic) | Many coding, agentic, knowledge tasks at half the price |
| Data retention | Mandatory 30-day | Supports zero retention |
| Self-verification / adaptive thinking | — | Yes |
The meta-rule that outlives both models
Every number above — internal or third-party — describes aggregate behavior across someone else's task mix. None of it describes yours. A vendor benchmark table predicts population averages; your production traffic is a specific, skewed sample of that population. This is the same discipline I argued for reading Kimi K3's benchmarks honestly: the leaderboard tells you where to start looking, never where to land.
So the only decision that counts is a small eval on your own distribution. Take 30–100 real tasks from your actual workload, run them on both models, and score two things the leaderboards mostly ignore: cost-weighted quality (quality per dollar, using your token profile, not the benchmark's) and self-verification / recovery (does the model catch its own mistakes and finish, or does it fail confidently?). One-shot accuracy is the least informative metric for agentic work, because it ignores the retries that dominate real cost. Given Opus 5's self-verification claims, this is precisely where it should either prove itself on your tasks or fail to.
Run that eval and, more often than teams expect, the answer is not one model but two. Route the easy and mid-tier work to Opus 5, reserve Fable 5 for the hardest long-horizon runs, and you get most of the flagship's frontier capability at close to the cheaper model's blended cost. That is not a hedge; it is a concrete case of when to use multiple AI agents — different tools for different task difficulties, chosen by a router, not a preference. Your eval sizes the split: measure what fraction of your traffic genuinely needs the frontier, and price the rest at $5/$25.
Anthropic's release cadence — Opus 4.8 in late May, a Claude 5 wave in June, then Opus 5 in July, which coverage calls its fourth model in two months — guarantees this specific table is stale within a quarter. The rule isn't. A cheaper model can beat a pricier one; a vendor's composite can flatter the wrong winner; the frontier and the sweet spot can be two different products. Pick by your job, verify on your data, and let the leaderboard argue with itself.