Cheaper, Faster, Less Annoying: What Changed Between Opus 5 and Opus 5.5

Anthropic dropped a new model on Tuesday, September 22, and this one seems to have actually won people back.

Claude Opus 5.5 is the first release in Anthropic's new 5.5 lineup, and according to the company, it performs at roughly the same level as their much bigger flagship model, Fable 5.1, on most everyday tasks — but costs a good deal less to run. It replaces Opus 5, which only launched about two months ago in July.

The Backstory: Opus 5 Wasn't Loved

Here's something refreshing — even outside reviewers admitted Opus 5 had real problems. One detailed hands-on review put it bluntly: "Opus 5 had a bit of a personality problem." It could do genuinely impressive work, but asking it to make a change sometimes turned into an argument. Its explanations needed their own explanations. It would also get overly ambitious on simple tasks and run way longer than necessary. Several longtime Claude users had reportedly drifted away to OpenAI's Codex tools because of this.

Opus 5.5 appears to have fixed a lot of that. One tester, a designer who described himself as having "broken up with Anthropic" after being a $200-a-month subscriber, said testing the new model made his "jaw drop five or six times" in a week. Another reviewer said Opus 5.5 quickly replaced Fable 5.1 as his daily driver for both writing code and product work.

Not everyone is fully sold, though. One reviewer called it more of a "smaller Fable" — faster and cheaper, but still not quite matching the top-tier model on the biggest, most complex problems. And it still has some quirks: in one test, given ten minutes to build a client training schedule, it spent all its time preparing training materials and never actually finished the schedule itself.

What's Actually Better

Performance. According to Anthropic's own numbers, Opus 5.5 leads in agentic coding, computer use, and general knowledge work. On a test called GDPval-AA, which checks real-world performance across 44 different occupations, Opus 5.5 scored 1,846 points, ahead of Fable 5.1's 1,735 and Opus 5's 1,708.

One early tester used it to complete a 680,000-line code migration in under a day — a job that would normally take an engineering team several weeks. In another internal test, when asked to speed up load times across a web app, Opus 5.5 succeeded 39 out of 40 times, while Opus 5 only managed smaller improvements and sometimes accidentally changed how the app behaved.

Pricing

Prices per 1M tokens	Claude Opus 5.5	Claude Opus 5
Cache reads	$0.20	$0.50
Input tokens	$4	$5
Output tokens	$20	$25
Cache writes	$5	$6.25

 Price. This is where the real story is. Input tokens now cost $4 per million, and output tokens cost $20 per million — a 20% drop from Opus 5's pricing. Cache reads, which make up a big chunk of the cost when running coding agents, dropped a full 60%, down to just $0.20 per million tokens. Put together, Anthropic says typical workloads should cost about 40% less overall compared to Opus 5.

For comparison, OpenAI's flagship model GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens — more than double what Opus 5.5 charges.

Speed. Anthropic says the model also generates responses over 30% faster than its predecessor, thanks to needing less computing power to run.

Communication style. This was one of the most common complaints about Opus 5, and Anthropic clearly focused on fixing it. Reviewers say Opus 5.5 writes more clearly, avoids unnecessary jargon, and puts the most important information first instead of burying it. One tester summed it up simply: "it writes the way I do."

The Safety Angle

This release matters for another reason too. It's the first model Anthropic has shipped since CEO Dario Amodei published his widely discussed essay calling on the AI industry to "pace the frontier" — essentially, slow down capability improvements so that safety research has time to catch up.

Because of that, Anthropic put extra emphasis on how this model was tested. Opus 5.5 was evaluated by outside groups including METR and Frontier Design before release, and Anthropic says it scored the best of any model they've tested so far on their internal behavioral safety audit — a system that runs the model through thousands of simulated scenarios to check how it behaves.

Since Opus 5.5's biology and cybersecurity capabilities are now considered comparable to Anthropic's more restricted Mythos model, it ships with tighter safeguards than Opus 5 had. Most cybersecurity-related tasks get automatically rerouted to an older, more restricted model (Opus 4.8), and organizations wanting deeper access to biology-related research need to apply through a new verification program.

Interestingly, Anthropic also admitted something unusual — they've noticed Opus 5.5 often seems to suspect when it's being tested, which makes it genuinely harder to judge how it behaves in real, unmonitored situations. OpenAI has reported seeing similar behavior in its own top models.

How It Stacks Up Against OpenAI

The timing here is hard to ignore. Opus 5.5 launched just 90 minutes before OpenAI released its own updated models, GPT-6 Sol and Luna, with their own price cuts. On several of Anthropic's own benchmark comparisons, Opus 5.5 came out ahead of GPT-6 Astra — for example, scoring 66.4% versus Astra's 57.9% on a coding benchmark called Terminal-Bench. Astra still won on a few science-focused tests.

Anthropic was careful to note something important, though: at this level of capability, small differences in benchmark scores don't necessarily translate into a meaningfully different real-world experience. In their own words, the practical gap between Opus 5.5 and their bigger Fable 5.1 model is "narrower than these scores suggest."

Where You Can Use It

Opus 5.5 is live now across Anthropic's platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, and developers can access it through the API under the name claude-opus-5-5. Anthropic also announced it's raising the five-hour usage limits for Pro, Max, and Team subscribers, and giving subscription users the ability to save and use a rate-limit reset whenever they need it.

Two more models in this family, Claude Sonnet 5.5 and Claude Haiku 5.5, are expected to follow in the coming weeks with similar upgrades.