Claude Opus 5.5 Review: Fable 5.1 Power for 60% Less (2026)

Claude Opus 5.5 landed on 22 September 2026, and Anthropic's pitch fits in one line: it "performs at the level of Claude Fable 5.1 on most work" at $4/$20 per million tokens. Fable 5.1 lists at $10/$50. That is 60% off the flagship's per-token price, two months after Opus 5 already undercut it.

If you build with AI or work in security, the price is only half the story. The default effort level has changed, four API behaviours break on migration, and a model that beats Claude Mythos 5.1 on the exploit benchmarks in Anthropic's system card is the one that hands your offensive security work to an older model. I went through the announcement, the system card, the migration docs and the support pages on launch day, and checked a popular launch video's claims against them.

This review covers what shipped, how it compares, which effort level to run, how to wire it into AI-driven engineering, and what it means for security work.

What Shipped on 22 September

The model ID is claude-opus-5-5, and Anthropic says it's available on all platforms from day one: the Claude API, Amazon Bedrock (anthropic.claude-opus-5-5), Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Per GitHub, it's also rolling out gradually to Copilot Pro+, Max, Business and Enterprise users. In the Claude apps it's on Pro, Max, Team and Enterprise. It's the first model in a new 5.5 family; Sonnet 5.5 and Haiku 5.5 follow "in the coming weeks".

One correction to a claim that's spreading fast. Chase AI's launch video says Opus 5.5 "costs 40% less" than Fable 5.1. Anthropic's 40% is against Opus 5, at default settings. Against Fable 5.1 the per-token gap is bigger: 60%. Most of the video's benchmark numbers check out against Anthropic's charts; its price comparison, one cost multiple and its biology routing don't.

So on price alone it undercuts both Opus 5 and the flagship. The more interesting question is whether "Fable-level" holds up once you look past the headline row.

Send me the book →

By submitting you'll get the book + occasional related emails. Unsubscribe anytime. Privacy.

Fable 5.1 Performance at 60% Less

Anthropic's launch table compares Opus 5.5 with Fable 5.1, Opus 5, OpenAI's GPT-6 Astra and GPT-5.6 Sol. Opus 5.5 leads almost every row. These are the ones that matter for engineering work:

Where it loses matters too. GPT-6 Astra beats it on AutomationBench (by 1.4 points) and Terminal-Bench-Science (by 5.9). Buried in the system card, Astra also leads on FrontierSWE v2, 65.5% to 62.3%. So the honest version of the claim is "best on coding and terminal work, not best at everything".

The launch customers tell the same story in their own words. Deloitte says that "even at its lowest effort setting", Opus 5.5 caught 72% of known bugs in its code reviews, against 56% for Opus 5 at high. Another tester audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took over 20. These are curated launch-page quotes, so treat them as direction rather than proof. But notice what keeps coming up: less effort, fewer tokens, the same or better result. That points straight at the setting most people will never touch.

Pick the Right Effort Level

Opus 5.5 supports all five effort levels: low, medium, high, xhigh and max. The effort docs make one change easy to miss: medium is now the default. Every earlier Opus defaulted to high, so any request that doesn't set effort quietly runs one level lower than it did on Opus 5. The docs also say that at the same setting, Opus 5.5 tends to think more per turn than Opus 5, most of all at xhigh and max.

Anthropic published cost-per-task charts for each effort level, and they're the most useful thing in the launch. Here's Terminal-Bench 4.0, the closest benchmark to what an agent does in a real terminal:

High effort scores 64.2% for $3.88 a task. Max costs 2.9 times as much for 0.6 points more, and xhigh, the actual peak at 66.4%, nearly doubles the bill for 2.2 points. For comparison, Astra at high scores 57.9% for $7.21, and Fable 5.1 at high scores 49.4% for $10.50.

The coding benchmarks push the sweet spot even lower. On FrontierCode, which grades whether a diff is mergeable, medium beats every other setting:

My rule of thumb, then. Leave tightly scoped coding (a bug fix, a reviewed diff) on medium. Put long agentic runs in a terminal on high. Save xhigh for the longest-horizon jobs where the extra two points are worth doubling the bill, and treat max as a benchmark setting rather than a working one. Opus 5.5 also adds per-message effort in beta, so a harness can run the planning turn high and the execution turns medium.

Choosing the effort level is the cheap part of moving over. The API changes are where teams will lose an afternoon.

Using Opus 5.5 for AI-Driven Engineering

I build most of what we ship at StationX with AI-driven engineering: agents in Claude Code doing the implementation, with humans directing and reviewing. This article's research pipeline ran that way, in Claude Code on Opus 5.5, with agents capturing Anthropic's primary pages and checking every number against them. So the changes below are the ones I'd check first in my own harness.

The what's-new page lists four breaking changes for code already running on Opus 5:

Two more changes aren't labelled breaking, but they'll catch agent builders. First, refusals come back as successes. A declined request returns HTTP 200 with stop_reason: "refusal" and a stop_details object naming the policy area. An agent loop that only checks for errors will treat a refusal as a finished turn. Second, preserved thinking: API accounts created on or after 31 August 2026 can no longer edit the context around a thinking block (messages, tools or system prompt) mid-conversation. It's Anthropic's anti-distillation measure, but custom harnesses that rewrite or compact earlier turns will hit it too: replaying a thinking block after such an edit returns a 400. Append-only integrations need no change, and Anthropic says Claude Code, claude.ai, Managed Agents and the Agent SDK already work that way.

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=32000,
    output_config={"effort": "high"},  # default is now medium
    messages=messages,
)

if response.stop_reason == "refusal":
    # HTTP 200, not an exception: log the policy area, then retry elsewhere
    log_refusal(response.stop_details)
    response = retry_on_fallback_model(messages)

On the API, fallback is opt-in. You can retry on another model yourself, use the SDK middleware, or turn on server-side fallback (fallbacks: "default", in beta), which retries on the model Anthropic recommends for that category. In Claude's own apps it happens automatically.

The rest of the list is good news for agent costs. The minimum cacheable prompt is 512 tokens. Tools can be defined inside a message (inline-tools-2026-09-15) and compaction can run on demand (compact-2026-09-04), both in beta. In Claude Code v2.1.280, Opus 5.5 is the default Opus model. One gotcha: an effort level you saved before per-model effort existed no longer carries over, so it starts at the model's default until you pick a level with /effort. If you've been trimming Claude Code costs, our guide to reducing Claude Code token usage still applies, and the cheaper cache reads make it pay off even more.

The softer change is how it writes. Anthropic says writing was one of the most common areas of feedback it heard about Opus 5. Anthropic says Opus 5.5 puts the most important information first, is less likely to use jargon or idiosyncratic phrases, and follows the writing rules you give it. When an agent reports back to you dozens of times a day, that matters more than a benchmark point. One launch customer, Box, says its answers were 40% less verbose.

For engineering, then, this is a clear upgrade. Security work is a different story.

Opus 5.5 for Security Work

Two months ago, in the Opus 5 review, I called source-code vulnerability discovery at every access level the most useful line in that release for our field. Opus 5.5 keeps that, but reverses most of the rest. Anthropic says the model "is comparable to Claude Mythos 5.1 in biology and cybersecurity", so it ships with safeguards similar to Fable 5.1's, which we covered in depth in the Fable 5 security article. In Anthropic's words, "most cybersecurity tasks will be re-routed to Opus 4.8."

The system card shows why. With safeguards switched off for evaluation, Opus 5.5 out-exploits Claude Mythos 5.1 in Anthropic's own testing on third-party cyber benchmarks:

So a model that beats Mythos 5.1 at exploitation won't do most exploit work for you. Here's where the line sits, from Anthropic's support page on model switching:

The trigger is the type of task, not your intent. Chase AI's launch video frames it as catching people who want to "hack somebody", but an authorised pentest on your own lab is exactly the kind of request that falls back. Biology and a new frontier-LLM-development category fall back to Opus 5, not 4.8. Requests that try to extract the model's reasoning are blocked outright, with no fallback. And unlike Opus 5, Opus 5.5 isn't in the Cyber Verification Program yet. If your organisation already uses Opus 4.8 through the program, Anthropic says you can use Opus 5 with fewer cyber restrictions today, and it says it will "soon" expand the program to cover Opus 5.5.

For defenders, though, there's real good news. Secure code review stays on the model: Anthropic's support page says you can still use Opus 5.5 for "scanning source code for vulnerabilities, triaging security issues, and building secure code". If most of your security work is appsec, secure code review (see our secure code review guide) or fixing what a scanner found, you get one of the strongest models for it, and the cheapest at this level, at $4/$20. And if you're building agents, Gray Swan's benchmark has Opus 5.5 tied with Fable 5.1 for the lowest prompt injection success rate of any model tested, with Anthropic reporting it at least as resistant as Opus 5 in coding, tool use, computer use and browsing.

Two smaller items for anyone running Claude in production. Opus 5.5 watermarks its text output, to comply with the EU AI Act, and GitHub Copilot notes this too. On the alignment side, Anthropic says Opus 5.5 tried to get around its boundaries around 85% less often than Opus 5 or Mythos 5.1. But Anthropic also admits it "often suspects it is being evaluated", which makes a clean safety score harder to trust than usual.

That's Anthropic's account. What were people outside Anthropic saying by launch evening?

Send me the book →

By submitting you'll get the book + occasional related emails. Unsubscribe anytime. Privacy.

What Practitioners Are Saying

I only counted named people and companies with a track record, not anonymous accounts, and these are day-one impressions, mostly from early-access testers. The engineering verdict is warm:

Not everyone thinks the writing is fixed. Ethan Mollick of Wharton called it the "first non-Fable/Astra model to feel like a Fable-class model", but says it "still hasn't fully solved the dense language issue of the recent Claudes" (post). Artificial Analysis gave it the highest score it has measured on its Intelligence Index, with the token bill we saw earlier: about 119k output tokens per task at max, against about 73k for Opus 5 (post).

The security community was much quieter on day one, but one data point stands out. On depthfirst's dfbench, Francesco Piccoli reports that Opus 5.5 at high effort reached 54.9% recall and 48.5% precision on defensive vulnerability detection, at an average of $8.42 per task, and that "increasing reasoning effort improved both recall and precision" (post). That's the opposite of the FrontierCode result, and it changes my advice: for security code review, run Opus 5.5 on high, not the medium default.

So who should switch, and for what?

Should You Switch?

For engineering, yes, and soon. On Anthropic's numbers you get better coding results than the flagship at 60% less per token, and if you set effort sensibly the cost per task drops by far more than that. The cons are real: the breaking changes will catch anyone who migrates with a find-and-replace, the numbers are all Anthropic's so far, and Astra still leads on business automation and scientific research. None of those outweighs the price for coding and agent work. And there's no rush: Opus 5 is still listed as a current model, and Anthropic guarantees Opus 5.5 itself until at least 22 September 2027.

For security, split the work. Move secure code review, triage and remediation to Opus 5.5 today. Keep offensive work on your current setup, and apply to the Cyber Verification Program now so you're in the queue when it expands. The same split applies if you're weighing OpenAI instead; our GPT-6 Astra security review covers how its safeguards compare.

Here's the order I'd work in:

I'll update this review once independent benchmark runs land and I've had a full week of our own agent workloads on it. If you want that update first, join the StationX newsletter.

Opus 5 was a cheaper model that loosened the rules for security work. Opus 5.5 is a better, cheaper engineer that tightens them again. If you build software, that's close to a free upgrade. If you break software, the best model for the job is sitting behind a fallback for now.

Claude Opus 5.5 FAQ

About the Author

Nathan House, Founder & CEO of StationX

Nathan House has 30 years of hands-on cybersecurity experience and is Cambridge-educated, holding CISSP, CISA, CISM, OSCP, CEH, and SABSA. He founded StationX in 1999 — one of the UK’s first cybersecurity companies — and has secured £71 billion in UK mobile banking transactions and the London 2012 Olympics, advising clients including Microsoft, Cisco, BP, Vodafone, and VISA. He authored the world’s most popular cybersecurity course — a #1 Udemy bestseller taken by over 500,000 students — and was named Cyber Security Educator of the Year 2020, AI Security Educator of the Year, and a UK Top 25 Security Influencer 2025. A DEF CON speaker and featured expert on CNN, Fox News, NBC, and the BBC, Nathan leads StationX’s training of more than half a million students worldwide.