Claude Opus 5 lands: near-Fable-5 quality at half the price, fewer restrictions, and Anthropic's lowest-ever misalignment score

What launched

On July 24, 2026, Anthropic introduced Claude Opus 5, and the framing is deliberately about value, not just a new ceiling. Per Bloomberg, TechCrunch, and CNBC:

• Smaller than Fable 5, but beats it on several benchmarks. Opus 5 is a more efficient model that, on Anthropic’s own table, matches or exceeds the Mythos-class Fable 5 in many categories — remarkable given Fable 5 is Anthropic’s most capable widely-available model.
• Half the price of Fable 5. Pricing is $5 input / $25 output per million tokens — identical to Opus 4.8, and exactly half what Fable 5 costs.
• New lineup position. It’s the new default model on Claude Max and the strongest model on Claude Pro. Free and Pro users default to Sonnet 5.
• Benchmarks (Anthropic’s own): Frontier-Bench 43.3%, ARC-AGI-3 30.2%, and more than double Opus 4.8 on agentic terminal coding. Anthropic also highlights stronger self-verification — an example where Opus 5 “wrote its own computer vision pipeline in response to an incomplete prompt.”
• Fewer restrictions. Unlike Fable 5 and Mythos, Opus 5 is free from many restrictions and not subject to the 30-day data-retention policy; cyber safeguards remain (it can analyze source code but not scan binaries for vulnerabilities), and safety classifiers engage 85% less often than for Fable 5.
• New beta feature — “Automatic Fallbacks”: blocked requests route to a less powerful model instead of erroring.

Two months after Opus 4.8, Anthropic has effectively obsoleted its own mid-2026 lineup.

Why this matters

1. This is the “frontier at half price” move, and it resets the Claude value equation. The headline isn’t that Opus 5 is the most powerful Claude — Fable 5 arguably still is at the extreme. It’s that near-top-tier quality now costs $5/$25 instead of Fable’s $10/$50. For anyone running Claude at volume — agents, coding, document work — halving the price of frontier-class output is a bigger deal than a few benchmark points. It’s the same “efficiency over raw ceiling” logic behind GPT-5.6’s Terra tier, applied at the top of Anthropic’s range rather than the middle. The buyer takeaway is direct: your Claude bill for the same work likely just dropped.

2. It simplifies a lineup that had gotten confusing. Anthropic’s mid-2026 range was a tangle: Opus 4.8, Sonnet 5, Fable 5 (Mythos-class, export-controlled), Mythos 5 (partner-gated). Opus 5 collapses much of that: it’s the new everyday frontier default on Max, cheaper than Fable, less restricted, and stronger than Opus 4.8. For most users the decision is now clean — Opus 5 for serious work, Sonnet 5 for cheap/fast, Fable 5 only when you need the absolute ceiling. That clarity is worth something on its own, especially for teams that had been paralysed by the “which Claude” question — the answer is now Opus 5 by default, with the other two as deliberate exceptions rather than a menu you have to reason through every time.

3. The safety numbers are a genuine differentiator — with an asterisk. Anthropic claims Opus 5 posts its lowest misalignment score ever and needs 85% less classifier intervention than Fable 5. If that holds up, it’s a real advance: a more capable model that also behaves more predictably, requiring less heavy-handed gating. It’s consistent with Anthropic topping the FLI AI Safety Index and with the Cyber Jailbreak Severity framework it built. The asterisk: these are Anthropic’s own internal measurements, and “misalignment score” is a metric it defines. Credible, but not independently audited — treat it as a well-supported claim, not a settled fact.

4. “Fewer restrictions” is a quiet but real shift after the export-control saga. Fable 5 and Mythos carried heavy restrictions and a 30-day data-retention policy — legacy of the export-control shutdown that pulled them offline for 18 days in June. Opus 5 shipping without those constraints, and needing far less classifier intervention, suggests Anthropic found a way to deliver near-Fable capability inside a lighter-touch safety envelope — which is exactly what a business deploying Claude wants: high capability without the compliance drag. It’s arguably the most practically useful part of the launch.

5. The cadence itself is the strategic signal. Opus 4.8 in late May, Sonnet 5 June 30, Opus 5 July 24 — Anthropic is shipping frontier updates roughly monthly, each cheaper or more efficient than the last, all while reportedly racing toward an IPO. That pace is a message to buyers and investors alike: the model itself is a fast-depreciating asset, and the winner is whoever ships capability-per-dollar fastest. For a buyer, it’s a reason not to over-commit to any single model version — the next, cheaper one is weeks away.

What this means for you

• Max and Pro users: Opus 5 is your new best model — start using it as the default for serious work. On Max it already is the default.
• If you run Claude at volume: re-price your workloads. Near-Fable quality at $5/$25 likely cuts your cost meaningfully versus Fable 5, and it’s stronger than Opus 4.8 at the same price.
• When to still use Fable 5: only for the genuinely hardest reasoning/strategy tasks where the absolute ceiling matters and cost is secondary. For most work, Opus 5 is the better trade.
• Free users: you’re still on Sonnet 5 — excellent for most tasks, but Opus 5 is a real reason to consider Pro if your work rewards frontier quality.
• Verify before you commit: the benchmark and safety numbers are Anthropic’s own. Test Opus 5 on your actual tasks against GPT-5.6 and your current model before re-platforming; see the best AI chatbots guide.

The honest caveats

• Every benchmark here is Anthropic’s own. Frontier-Bench 43.3%, ARC-AGI-3 30.2%, “>2× Opus 4.8 on terminal coding,” and “beats Fable 5 on several benchmarks” all come from Anthropic’s announcement. No independent evaluation existed at launch. Treat as claims to verify.
• “Lowest misalignment score” is a self-defined, self-measured metric. It’s a meaningful signal given Anthropic’s track record, but it’s not an external audit, and the metric’s definition is Anthropic’s. Read it as directional, not definitive.
• “Automatic Fallbacks” has a hidden cost. Silently routing a blocked request to a weaker model avoids errors but means you may get a lower-quality answer without realizing it. If you need consistent model behavior (e.g., in production pipelines), understand when fallbacks trigger.
• “Fewer restrictions” isn’t “no restrictions.” Cyber safeguards remain (no binary scanning), and the lighter classifier touch is relative to Fable 5, not absolute. Don’t assume Opus 5 is unrestricted.
• The lineup could shift again fast. At a roughly monthly cadence, today’s positioning (Opus 5 as everyday frontier, Fable 5 as ceiling) may change within weeks. Don’t architect long-term around a specific model version.

The grounded summary: Claude Opus 5 is the most consequential Claude release since Fable 5’s return — not because it’s the most powerful, but because it puts near-Fable-5 capability at half the price, with fewer restrictions and, per Anthropic, better behavior. For most Claude users it’s the new default and a genuine cost cut. Verify the numbers on your own work, keep Fable 5 in your back pocket for the hardest jobs, and expect the next, cheaper flagship sooner than you’d think.

Questions or corrections? Email Pick Right. Want the full list? See all news.