Claude Opus 5.5: Pricing, Benchmarks, and Whether to Switch From Opus 5 or Fable 5.1

What is Claude Opus 5.5?

Claude Opus 5.5 is Anthropic's newest Opus-tier model, released on September 22, 2026. It replaces Opus 5 as the default recommendation for most workloads, and it is cheaper per token than the model it replaces. That's unusual for a flagship update.

The short version: it scores higher than Opus 5 on every benchmark Anthropic published, it costs $4/$20 per million input/output tokens instead of $5/$25, and Anthropic says it matches the far more expensive Fable 5.1 on most tasks. The figures below come from Anthropic's Claude Opus 5.5 launch page and the Claude models overview in the platform docs.

Claude Opus 5.5 pricing: what a real agent job costs

The list-price cut looks modest at 20% on input and output. The bigger change is in prompt caching, where cache reads fall by 60%.

This matters because agentic workloads are dominated by cache reads. A coding agent or a multi-step automation re-sends the same system prompt, tool definitions and growing conversation on every turn, and most of those tokens come back as cache hits. Here is an illustrative agent session. The token volumes are an example profile, and the prices are the published ones:

That is about 30% cheaper with no change to how many tokens the job uses. Anthropic also says Opus 5.5 uses fewer tokens to finish the same work and puts the saving at roughly 40% on typical workloads. Treat 40% as the vendor's estimate and 30% as the floor you get from pricing alone if your pipeline is cache-heavy. For a single-shot, uncached call the saving is a flat 20%.

Claude Opus 5.5 vs Opus 5: benchmark results

Anthropic's launch table compares the two directly:

The pattern matters more than any single number. The biggest gains are in long agentic loops: Terminal-Bench and CursorBench both reward a model that keeps its plan, recovers from failed commands and doesn't wander. The smallest gain is on FrontierCode, which is closer to one-shot hard problem solving. If your use case is "answer one hard question," the upgrade is real but incremental. If it's "run 40 tool calls unattended," it's the largest jump in the table.

Anthropic also reports that Opus 5.5 generates output more than 30% faster than Opus 5 and that it beats GPT-6 Astra on Terminal-Bench 4.0 at roughly a fifth of the cost per task. That last comparison is a vendor claim, so check it on your own tasks before you rely on it.

Opus 5.5 vs Fable 5.1: which one should you run?

Fable 5.1 is Anthropic's Mythos-tier model, and it costs 2.5 times as much per token:

Anthropic says Opus 5.5 matches Fable 5.1 on most tasks and beats it on agentic coding. On Terminal-Bench 4.0 the launch figures put Opus 5.5 at 66.4% and Fable 5.1 at 55.8%. Given the price gap, our read is:

• Default to Opus 5.5 for coding agents, automation pipelines, document work and anything that runs at volume.
• Keep Fable 5.1 for a small set of tasks where you have evidence that it wins on your own evaluation set. Don't keep it just because it's the top tier.
• Use Sonnet 5 for high-volume, lower-stakes steps such as classification, extraction and routing, where half of Opus pricing adds up quickly.

One trap when you migrate from Fable: the default effort is different. Fable 5.1 defaults to high and Opus 5.5 defaults to medium. If you swap the model ID and quality drops on your hardest prompts, raise effort to high before you conclude the model can't do the task. If you run the comparison without matching effort, you're testing two variables at once.

How to switch to Claude Opus 5.5 in the API

For most integrations, migrating means changing one string. Thinking is adaptive and always on, so you don't pass a separate thinking flag:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    messages=[
        {"role": "user", "content": "Summarize the failing tests and propose a fix."}
    ],
)
print(response.content[0].text)

Before you move production traffic, check these three things:

1. Re-run your eval set at the effort level you'll actually ship. Medium is cheaper and faster, and it may be enough for most steps. Save high or xhigh for the planning or review step of an agent loop.
2. Re-check your cache hit rate. The 60% cache-read discount only pays off if your prompt prefix is stable. Put tool definitions and system instructions first and keep volatile content at the end.
3. Check any output parsers that expect the old style. Anthropic changed how Opus 5.5 writes (see the next section). If downstream code relies on a particular preamble or layout, test it.

If you're building agents, chatbots or document pipelines on top of this and would rather have the evaluation and the migration handled for you, that's what our AI automation and workflow service does.

Is Claude Opus 5.5 better at writing?

This is the change most people will notice in daily use. Users have long complained about a recognisable "Claudish" style: long wind-ups, hedging, and the answer buried in the third paragraph. According to Anthropic, Opus 5.5 puts the most important information first, uses less jargon and follows the user's writing rules more consistently. Their before-and-after examples show shorter answers with clearer structure than Opus 5.

In practice, this means two things. First, style prompts that tried to suppress verbosity ("be concise, no preamble, lead with the answer") may now be redundant, and redundant instructions can overcorrect into terse output. Remove them one at a time and compare. Second, style rules in your system prompt, such as a banned-phrase list or a house tone, should now be followed more reliably. That makes it worth writing them precisely.

Safeguards that can change which model answers you

This is a practical detail you won't find in most coverage: Opus 5.5 is the first Opus model with safeguards that match Fable 5.1 for biology, cybersecurity and frontier LLM development. For cybersecurity, Anthropic routes most cybersecurity tasks to Opus 4.8 rather than letting Opus 5.5 answer them.

If you build security tooling, such as log triage, vulnerability write-ups or detection rules, some of your requests may be served by an older model even though you asked for claude-opus-5-5. You'll see this as a quality or style difference on security prompts compared with everything else. Anthropic offers a Cyber Verification Program, and a Life Sciences equivalent, for vetted organisations that need the restricted capabilities.

On the positive side, Anthropic reports that Opus 5.5 has the best score of any model to date on its automated behavioral audit and is less vulnerable to prompt injection. For agents that read untrusted web pages or emails, that is a meaningful improvement, although it doesn't remove the need for your own input sanitisation.

Should you switch to Claude Opus 5.5 now?

For almost everyone on Opus 5, yes. It's cheaper, faster and better on every published benchmark, and the migration is a one-line change. The only reasons to wait are an output parser that depends on Opus 5's writing style, or security workloads affected by the cyber routing.

For teams on Fable 5.1, run a side-by-side test at matched effort on your own tasks. For many workloads the 60% per-token saving will be hard to justify giving up.

On subscription plans, Anthropic has also raised the five-hour usage limits. Combined with the model's lower token use, heavy Claude Code users should get noticeably more work done per window.

If you're unsure which parts of your stack should move to Opus 5.5, Sonnet 5 or a cheaper model, or you need an AI feature built properly from the start, tell us what you're running and get a scoped quote.