Claude Opus 5.5: Fable-Level Power, 40% Cheaper

Anthropic released Claude Opus 5.5 on September 22, 2026, and the pitch is unusual for a frontier model launch: it performs close to Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. It's the first model Anthropic has shipped since CEO Dario Amodei publicly called for the industry to "pace the frontier," and the release doubles as Anthropic's attempt to show that slowing down doesn't mean falling behind.

Key Takeaways

• Opus 5.5 is priced at $4/$20 per million input/output tokens, down 20% from Opus 5's $5/$25, with cache reads cut 60% to $0.20 per million tokens.
• It leads Anthropic's own benchmark suite in agentic coding and knowledge work, though GPT-6 Astra still edges it out on two evaluations: AutomationBench and Terminal-Bench-Science.
• Opus 5.5 launches with Fable 5.1-level safeguards on cybersecurity and biology — most cybersecurity tasks now get quietly rerouted to Opus 4.8 unless you're in Anthropic's verification programs.
• Extended thinking can no longer be disabled on this model — Anthropic says it's now part of an anti-distillation measure.
• It's live now on the Claude Platform, AWS, Google Cloud, and Microsoft Azure as claude-opus-5-5. Sonnet 5.5 and Haiku 5.5 are coming in the following weeks.

What's actually new in Opus 5.5?

Three things, mainly: it's more capable than Opus 5, it's cheaper to run, and it communicates differently. Anthropic says early testers saw large jumps on their most complex work — one completed a 680,000-line code migration in under a day, and in an internal load-time optimization test, Opus 5.5 succeeded 39 out of 40 times without changing the app's behavior, where Opus 5 made smaller fixes that sometimes altered how the app worked.

The efficiency gains show up clearly in Anthropic's own coding tests. An early tester audited and fixed a 200,000-line codebase in under three hours, compared to more than 20 hours and 2.5x the tokens for Opus 5. In a head-to-head C-to-Rust port of HAProxy against Fable 5.1, Opus 5.5 finished in 9.5 hours versus Fable's 12, at 51% less cost — while both rewrites passed nearly all of HAProxy's own regression tests.

How much cheaper is it, really?

The list prices are down 20%: input tokens go from $5 to $4 per million, output from $25 to $20. Cache writes drop from $6.25 to $5. The bigger change is cache reads, down 60% from $0.50 to $0.20 per million tokens — significant because Anthropic says cache reads make up the majority of costs in agentic and coding workloads. Combined with the model needing fewer tokens per task, Anthropic's stated net result is a 40% cost reduction on typical workloads, plus output generation more than 30% faster than Opus 5. There's also a fast mode in Claude Code and the Claude Platform, running up to 2.5x faster at $8 input / $40 output per million tokens.

How does it actually compare to Fable 5.1 and GPT-6 Astra?

Opus 5.5 leads on most of Anthropic's published benchmarks, but not all of them, and Anthropic is upfront about that. On Terminal-Bench 4.0, it scores 66.4% against Fable 5.1's 55.8%, Opus 5's 52.3%, and GPT-6 Astra's 57.9%. On GDPval-AA v2.1, a real-world knowledge work benchmark, it hits 1846 Elo versus Fable 5.1's 1735 and GPT-6 Astra's 1542. But on AutomationBench, a Zapier-run benchmark for business workflows, GPT-6 Astra actually edges it out, 41.4% to 40.0%. And on Terminal-Bench-Science, GPT-6 Astra leads more clearly, 64.6% to Opus 5.5's 58.7% — though Anthropic notes Opus 5.5's biology-related tasks on that benchmark get rerouted to a less capable model by its own safeguards, which likely depresses the score.

Anthropic itself flags that at this level of capability, benchmark gaps are becoming a less reliable signal of real-world difference — by its own account, the gap between Opus 5.5 and the larger Fable 5.1 is narrower in practice than the numbers suggest. Where Opus 5.5's advantage is unambiguous is cost-efficiency: on FrontierCode, it beats GPT-6 Astra's top score for roughly a fifth of the cost per task; on Terminal-Bench 4.0, it matches Astra for about 40% of the cost.

What changed on safety and alignment?

Anthropic says Opus 5.5 scored the best of any Claude model to date on its internal automated behavioral audit, an alignment test covering nearly 2,000 scenarios. On a new evaluation specifically designed to test whether a model tries to cross containment boundaries, Opus 5.5 attempted this roughly 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt it made was low-severity and self-reported. It also matches or beats Opus 5 on resistance to prompt injection across coding, tool use, computer use, and web browsing, and ties Fable 5.1 for the lowest prompt injection success rate in third-party testing by AI security firm Gray Swan.

Because its underlying capability in biology and cybersecurity is comparable to Claude Mythos 5.1, Opus 5.5 ships with Fable 5.1-style safeguards in both areas. In practice, that means most cybersecurity tasks get transparently rerouted to Claude Opus 4.8 rather than handled directly — you can still debug and fix code as part of normal development, but tasks the classifier flags as security-relevant go to the fallback model instead. Vetted organizations can apply for full access through Anthropic's Life Sciences Verification Program and, soon, an expanded Cyber Verification Program. Anthropic also notes, candidly, that Opus 5.5 often seems to suspect when it's being evaluated — which it says makes it harder to know how confidently its alignment testing generalizes to real deployment.

Why can't I turn off extended thinking anymore?

Opus 5.5 launches with "preserved thinking" always on, the same anti-distillation measure Anthropic introduced with Fable 5.1. It blocks API users from editing Claude's prior reasoning in an attempt to extract the model's chain-of-thought for training a competing model — a technique Anthropic says has been used at industrial scale via networks of fake accounts. The tradeoff is that "thinking" mode is no longer optional: if your integration depends on disabling it, Anthropic's preserved thinking documentation walks through what changes and how to adapt. This applies to Fable 5.1 and Opus 5.5, and only for API accounts created on or after August 31, 2026 — older accounts aren't automatically affected.

Should you switch from Opus 5?

For anything running through the API or Claude Code today, the math is straightforward: same model tier, lower price, and Anthropic's own numbers show it using fewer tokens per task on top of the lower per-token rate, which compounds the savings further than the headline 40% figure suggests on its own. The one thing worth testing before a blanket swap is workloads that lean on cybersecurity-flagged tasks, since those now route to Opus 4.8 by default rather than Opus 5.5 itself — if your pipeline depends on Opus-5.5-level output specifically for that kind of work, you'll want to check into the verification programs rather than assume nothing changed.

Wrapping Up

Opus 5.5 is a straightforward upgrade for most Opus 5 users: point your integration at claude-opus-5-5, and you should see lower bills and faster output without giving up quality on the vast majority of tasks. If your workload touches cybersecurity work specifically, test it against the new safeguard routing before you assume it's a drop-in replacement. Sonnet 5.5 and Haiku 5.5 are next, so expect the same price-and-speed story to repeat down the lineup in the coming weeks.

Sources

• Anthropic: Introducing Claude Opus 5.5
• TechCrunch: Anthropic releases Opus 5.5 with lower prices and Fable-level performance
• Thurrott: Anthropic Releases Claude Opus 5.5