Claude Opus 5 Review: Everything New in 2026

Anthropic released Claude Opus 5 on 24 July 2026, following quickly on Opus 4.8. Same $5 input / $25 output pricing as 4.8, half of what Fable 5 costs, and Anthropic's own system card states plainly that it is not more capable overall than Fable 5. A model priced below the flagship it sometimes beats: if that leaves you confused, the Hacker News launch thread is full of developers asking the same question.

We've reviewed every release in this series (the 4.7 review and the 4.8 review), and this one follows the same method: launch day spent in the announcement, the 194-page system card, and the migration docs, checked against what developers are actually reporting.

The plan: what shipped, why the smaller-but-cheaper positioning matters, the migration changes that can break or silently alter your API calls, the security story nobody else is leading with, and whether you should switch. So let's get into it.

What Actually Shipped on 24 July

If you're just after the Claude Opus 5 release date: 24 July 2026, with Anthropic saying it's "available today on all platforms". Their docs list the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry. The model ID is claude-opus-5 (anthropic.claude-opus-5 on Bedrock), and you'll see it written as "Claude 5 Opus" in places (same model).

The spec sheet, from Anthropic's docs:

The specs are straightforward and the price familiar. The line that actually explains this launch sits outside the benchmark table entirely: a sentence in the system card about what Opus 5 is not.

Send me the book →

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

Smaller Than Fable 5, and It Beats It Anyway

The RSP section of the system card says it outright: "Claude Opus 5 is not more capable overall than our most capable general-access model, Claude Fable 5." TechCrunch framed the launch the same way: a smaller model that still beats Fable 5 on several benchmarks. Anthropic shipped a model below its own flagship and priced it at half of Fable 5's $10/$50.

On Anthropic's own numbers, the smaller model gets uncomfortably close. Within 0.5% of Fable 5's peak CursorBench 3.2 score at half the cost per task. Beating Fable 5's best OSWorld 2.0 result at just over a third of the cost. Cursor's Sualeh Asif, quoted in the launch post, called it "near Fable 5 intelligence at Opus speed and cost". It's a curated marketing quote, but it matches the pattern the ratios describe.

The honest counterweight, and it's Anthropic's own: Opus 5 trails Fable 5 on SWE-bench Pro, 79.2 against 80.0. Close, but the flagship is still the flagship.

One number needs defusing. Zapier's Wade Foster says in his launch testimonial that Opus 5 "hit 100%". That refers to one churn-prevention workflow in Zapier's own testing, not a benchmark. On Zapier's AutomationBench, the system card shows 26.0% against a next-best of 18.1. Big relative gap, modest absolute score. Anyone telling you Opus 5 scored 100% on Zapier's benchmark has confused a testimonial with a result.

The developer reaction split exactly along this line. The Hacker News launch thread (1,100+ points and 600+ comments within hours of launch) has datakan asking "Ok then so what's the point?" while zuzululu answers it in effect: "so almost fable 5 with 50% cheaper cost? sign me up". My take, as someone who builds agent infrastructure on Claude Code daily and actually reads the token bill: zuzululu has it right. For agentic work billed by the token, price/performance is the product, and "nearly the flagship at half price" is a better deal than a marginal capability bump would have been.

Therefore the upgrade should be a no-brainer. It would be. Except this time, unlike 4.8, swapping the model ID can break your code.

Thinking On by Default: The Breaking Change

In the 4.8 review we wrote "swap the model ID and your 4.7 code keeps working." Not this time.

Claude Opus 5 ships with thinking on by default, and the effort ladder (low, medium, high, xhigh, max) no longer needs a beta header. The loud breaking part: if you disable thinking and request xhigh or max effort, the API now returns an HTTP 400. Anthropic's docs call this "a breaking change from Claude Opus 4.8", unusually direct labelling from them. If any of your pipelines run thinking-off with high effort settings, a common cost-tuning pattern, those calls fail the moment you swap the ID.

Strictly, the migration guide counts two breaking changes, and the quieter one costs money rather than throwing errors: requests that send no thinking field at all, which ran without thinking on 4.8, now run with adaptive thinking. Your max_tokens still caps total output, thinking plus response text, so a limit tuned for a thinking-free 4.8 workload can now truncate responses mid-answer. Revisit max_tokens on any call you migrate, or pass thinking: {"type": "disabled"} (at high effort or below) to keep the old behaviour.

There's a quieter migration note that will save you money too. The docs advise removing verification instructions carried over from earlier models, lines like "include a final verification step" or "use a subagent to verify", because on Opus 5 they cause over-verification. We've all got prompts with those exact sentences in them; they were the standard fix for models that shipped unchecked work. On Opus 5 they make it check twice at your expense. Delete them.

Two prompt-file audits, then, before you migrate anything. Still, for this audience the change I'd rank highest has nothing to do with API behaviour. It's what the model will now agree to work on.

Security Work Unblocked

I've spent 30 years in security, and the last two of them watching frontier models refuse legitimate security work. It's been a running theme in this series since the "malware reminder" complaints in the 4.8 launch thread, and the subject of our Fable 5 security article. Every StationX student who's had a model refuse to explain a vulnerability class they're studying knows the frustration.

Opus 5 is the first release where Anthropic has moved meaningfully the other way. Anthropic's announcement says it expects the cyber safety classifiers to intervene around 85% less often than they do for Fable 5. Note both qualifiers: relative, and expected. That's not a measured refusal rate in the wild. Still, an 85% reduction against their own flagship is not a tweak.

More concretely: vulnerability discovery in source code is permitted at all access levels. No enterprise tier, no special agreement: code review for security flaws is now sanctioned work for everyone. Given how much of modern appsec, bug bounty triage, and secure-code training is exactly that, this is the single most useful line in the release for our field.

What's still gated: binary-based vulnerability scanning and exploit generation remain blocked, and penetration testing is blocked by default, but it's available to members of Anthropic's Cyber Verification Program. That exception matters. If pentesting is your day job, the path forward is verification, not jailbreaks.

Some launch-day reactions misread this. HN user not_a9 concluded "Okay so it's worse than Opus 4.8 for my purposes I guess?" But the documented change loosens the safeguards for source-code work; unless your work sits in the still-blocked categories, the docs point the other way.

One behaviour to know before your users hit it: on consumer surfaces (Claude.ai, Claude Code, Cowork), flagged requests fall back to Opus 4.8 by default: you ask for Opus 5, you sometimes get 4.8. On the API, though, fallback works differently — and it arrived alongside a set of API changes worth five minutes of your attention.

Fast Mode and the API Changes

Four things shipped alongside the model, and all four change how you'd wire it up.

Fast mode: $10/$50, research preview, first-party Claude API only. Yes — the price the rumour mill assigned to Opus 5 itself turned out to be its fast mode (and Anthropic's pricing page lists fast mode for Opus 4.8 as well, so it's a platform feature, not an Opus 5 exclusive). It's not on Bedrock or the other clouds, and in Claude Code it draws usage credits. Research preview means behaviour and pricing can move.

Server-side fallback is opt-in beta on the API (server-side-fallback-2026-07-01), covering safety-refusal cases only. Read that against the previous section: on consumer surfaces fallback to 4.8 is the default; on the API it only happens if you ask for it, which is the right design.

Mid-conversation tool changes no longer invalidate the prompt cache (beta header mid-conversation-tool-changes-2026-07-01). If your agents swap tools mid-session, this is real money on a 1M-token context.

The prompt-cache minimum drops to 512 tokens, from 1,024. Short-prompt callers (chat frontends, fast-turn agents) can now cache what previously fell under the bar. And if you batch, the 300K output header from the spec sheet stacks with all of this.

A long parts list, and cheaper on almost every line. Whether it adds up to a switch depends on which of four camps you're in.

Send me the book →

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

Should You Upgrade?

First, the comparison half the search traffic wants: Opus 5 vs GPT-5. The honest day-one answer is that we haven't seen a trustworthy head-to-head yet. Anthropic's launch material compares Opus 5 against its own models and unnamed "next-best" scores, and until a third party publishes a like-for-like run, any Opus-5-vs-GPT-5 verdict this week is vibes. We'll update this section when real numbers land, as we did for 4.8's late benchmark table.

Against Anthropic's own lineup, the picture is clear enough to commit:

My answer: switch your agentic and security workloads now. The price/performance case is strong and the security unblock genuinely useful. The cons are real: the consumer-surface fallback to 4.8 means you won't always get the model you asked for, the breaking changes will bite anyone who migrates by find-and-replace, and every impressive number in this launch is still Anthropic marking its own homework. None of those outweighs half-price near-flagship performance for the work we actually do. And Opus 4.8 is listed Active with no retirement date; the deprecation page guarantees it can't retire before 28 May 2027 on Anthropic-operated platforms (the clouds set their own schedules). So the cost of me being wrong is a config rollback, not a stranded stack.

That's who should move. What follows is what I'd actually do, in order.

What I'd Do First

I'll update this article once I've run Opus 5 against a full week of my own agent workloads, particularly the security prompts in step 3, and the moment any independent benchmark reproduction lands. If you want that update first, subscribe to the StationX newsletter or join the Circle community.

Last release, 4.8's pitch was a model that stopped hiding its mistakes. Opus 5's pitch is narrower and, for us, better: nearly the flagship, at half the price, finally allowed to hunt vulnerabilities in source code. That's a trade I'll take.

Claude Opus 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.