Claude Opus 5.5: What's New, Benchmarks, and What It Means for Business Chatbots

Last updated: September 24, 2026 | By Omar El Bahr

• Released on September 22, 2026, as the first model in Anthropic's new Claude 5.5 family
• 1M-token context window with up to 128K output tokens per response
• Priced at $4 per million input tokens and $20 per million output tokens on the Claude API
• Thinking is always on, and the default effort setting is now medium
• Scored 94.3% average accuracy across 42 languages on the Global MMLU benchmark
• Available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry Claude Opus 5.5 is Anthropic's newest Opus model, released on September 22, 2026, and the first in the Claude 5.5 family. It is built for long coding and knowledge work, costs $4 per million input tokens and $20 per million output tokens, and scored 94.3% average accuracy across 42 languages.

Most coverage of this release is written for developers. But many people searching for it run a business and want to know one thing: does this matter for me? To answer that, I went through Anthropic's announcement, the developer docs, independent test results, and the full 230-page system card, and pulled out what actually affects a business running a chatbot.

What Is Claude Opus 5.5?

It is the new top Opus model from Anthropic and the replacement for Claude Opus 5, which came out on July 24, 2026. The Opus 5.5 release date was September 22, 2026, and it went live the same day on Anthropic's own platform and on the big cloud providers.

Anthropic's pitch is simple. The company says the new model performs at the level of Claude Fable 5.1 on most work, while costing 40% less to run than Opus 5 on typical workloads. Fable 5.1 sits in Anthropic's higher Mythos tier, so that is the main story of this launch.

It is also Anthropic's first release since it called for "pacing the frontier", meaning AI progress should move at a speed safety work can keep up with. METR and Frontier Design tested it before release, and Claude Sonnet 5.5 and Claude Haiku 5.5 are due in the coming weeks.

What stood out to me is the pace. Opus 5 launched on July 24, Fable 5.1 on September 1, and this model on September 22. That is three frontier releases in under two months. A small support team does not have time to rebuild its setup every three weeks, and it should not have to. I come back to this later.

Want a chatbot that answers your customers from your own content, in their language? Start free with Agentency and have your first agent live today.

What's New in Claude Opus 5.5

The short version comes down to five changes: thinking is always on, the writing is clearer, it uses fewer tokens per task, it ships with stricter safeguards, and a few API behaviors changed. Below I go through the Claude Opus 5.5 features one by one and explain what each one means in practice.

Always-On Thinking and Opus 5.5 Effort Levels

Thinking can no longer be switched off. According to Anthropic's docs, a request that tries to disable thinking now returns an error. Instead, you control how hard the model thinks with the effort setting, which has five levels: low, medium, high, xhigh, and max.

With Opus 5.5 adaptive thinking, the model decides how much reasoning each task needs. The default effort is now medium, while Opus 5 defaulted to high. Anthropic also notes that at the same effort setting, the new model tends to think more per turn than Opus 5, especially at xhigh and max.

That matters more than it sounds. Developer Simon Willison ran his "pelican riding a bicycle" test on max effort, and the model kept reasoning until it hit the 128,000-token output limit without answering. It happened twice, and each attempt cost him $2.56 and took close to 20 minutes. My advice: start at medium, and only raise effort for the steps that need it.

Clearer Writing

This is the change I noticed most. Anthropic says the model puts the most important information first, uses less jargon, and follows the writing rules you give it, a direct fix for one of the most common complaints about Opus 5. Deloitte, an early tester, reported that even at the lowest effort it caught 72% of known bugs in their code reviews, compared with 56% for Opus 5 at high effort, with fewer false alarms and far less output.

Fewer Tokens per Task

Anthropic's 40% cost figure comes from a lower price per token plus fewer tokens per task. An early tester audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and used 2.5 times as many tokens. In an internal test, the model rewrote HAProxy from C into Rust in 9.5 hours, compared with 12 hours for Fable 5.1, at 51% lower cost. Output is also more than 30% faster than Opus 5.

Not every tester saw the same savings. CodeRabbit reported about 49% higher token usage than its current production setup in its lower-effort configuration, while catching slightly more known bugs, 51 out of 80 versus 49. Measure it on your own work before you assume the saving.

Opus 5.5 Safeguards

Because Anthropic rates the model as comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships with safeguards similar to Fable 5.1. In practice, most cybersecurity tasks are rerouted to Claude Opus 4.8, and some biology requests are handled by a different model too. Vetted organizations can apply to Anthropic's Life Sciences Verification Program for broader biology access, and Anthropic says its Cyber Verification Program will expand to this model in the coming weeks.

The safety numbers are strong. On a prompt injection benchmark run by the security firm Gray Swan, the model tied Fable 5.1 for the lowest attack success rate of any model tested. In a new test of whether a model tries to cross containment boundaries, it made those attempts about 85% less often than Opus 5 or Claude Mythos 5.1. It also launches with watermarking to comply with the EU AI Act, and it remains available with zero data retention.

Breaking Changes for Developers

Moving code from Opus 5? Four things break: thinking cannot be disabled, forced tool use returns an error, thinking blocks are tied to the model and the conversation, and the older computer_20251124 tool is no longer accepted on the Claude API and Google Cloud. Also, notes the model writes between tool calls now come back inside thinking blocks, so apps that stream them as progress updates can go quiet.

Claude Opus 5.5 Benchmarks

Here is Anthropic's published comparison. All scores use adaptive thinking at max effort, except Terminal-Bench 4.0, which was run at xhigh.

Benchmark	New Opus	Fable 5.1	Opus 5	GPT-6 Astra
Terminal-Bench 4.0 (agentic coding)	66.4%	55.8%	52.3%	57.9%
FrontierCode v1.1 Main (coding)	54.4%	50.3%	48.0%	53.3%
CursorBench 4.0 (coding)	57.8%	51.8%	46.6%	Not reported
GDPval-AA v2.1 (knowledge work, Elo)	1846	1735	1708	1542
AutomationBench (business workflows)	40.0%	31.4%	26.9%	41.4%
Humanity's Last Exam (with tools)	67.7%	65.6%	63.6%	57.2%
Terminal-Bench-Science 0.1	58.7%	52.6%	29.0%	64.6%
OSWorld 2.0 (computer use, partial)	81.8%	80.7%	74.0%	Not reported

The independent picture points the same way. Artificial Analysis placed the model at the top of its Intelligence Index with a score of 58 at max effort, ahead of Claude Fable 5.1 and GPT-6 Astra at 53.

Anthropic itself admits that at this level, benchmark margins are a less reliable guide to real-world differences, and that the gap with Fable 5.1 is narrower in its own use than the scores suggest. That is the right way to read any table like this.

Where the New Model Is Not the Best

This is the part most coverage skips. GPT-6 Astra still leads on AutomationBench (41.4% versus 40.0%) and on Terminal-Bench-Science (64.6% versus 58.7%).

The more interesting result is buried in the system card. On Toolathlon, a test of 108 real tool-use tasks across office work, ecommerce, operations, data analysis, and web research, the new model scored 77.8%. Claude Opus 5 scored 80.6% and Claude Opus 4.8 scored 79.9%. So on this kind of task, the newest model was not the strongest one.

That is not a knock on the release. It is proof that "newest" and "best for your job" are different things, and the only way to know is to test on your own questions. Our guide on how to evaluate an AI agent walks through how to do that properly.

How It Compares with Opus 5, Fable 5.1, and GPT-6

Here is a quick side-by-side of the models people are most likely comparing right now, based on published list prices per million tokens.

Model	Input / Output price	Context window	Best known for
New Opus	$4 / $20	1M	Long coding and knowledge work at lower cost
Claude Opus 5	$5 / $25	1M	The previous Opus flagship
Claude Fable 5.1	$10 / $50	1M	Anthropic's Mythos-tier model
GPT-6 Astra	$10 / $50	Not compared here	Business automation and science tasks
GPT-6 Sol	$2 / $10	Not compared here	OpenAI's cheaper tier, launched the same day

Against Opus 5, it is cheaper and scores higher on every benchmark Anthropic published for both. It beats Fable 5.1 on every published row at a lower price, and it wins most rows against GPT-6 Astra, but not all.

Claude Opus 5.5 Pricing and Availability

Here are the Claude API list prices per million tokens, with Opus 5 for comparison:

Price per 1M tokens	New Opus	Opus 5
Input	$4	$5
Output	$20	$25
Cache reads	$0.20	$0.50
Cache writes (5 minutes)	$5	$6.25

Input and output are 20% cheaper, and cache reads are 60% cheaper, which matters for long agent sessions where most input is context the model has already seen. The Batch API halves input and output prices, and a 1-hour cache write costs $8 per million tokens. There is also Opus 5.5 fast mode, which runs up to 2.5 times faster at $8 per million input tokens and $40 per million output tokens. On the Claude API it is offered as a research preview.

For availability, developers call the model as claude-opus-5-5 on the Claude API, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, and as anthropic.claude-opus-5-5 on Amazon Bedrock. Anthropic says it will not retire the model before September 22, 2027. On the consumer side, Anthropic also raised five-hour usage limits on its Pro, Max, Team, and seat-based Enterprise plans.

If you are trying to budget a support bot rather than an API bill, token prices are only part of the picture. We broke down the full cost in our guide on how much an AI chatbot costs.

Claude Opus 5.5 Multilingual Performance

On Global MMLU, a knowledge test translated into 42 languages that range from French and German to Yoruba, Igbo, and Chichewa, the model averaged 94.3% accuracy. Fable 5.1 scored 94.0%, Opus 5 scored 92.5%, and Claude Sonnet 5 scored 89.2%. On MILU, which covers ten Indian languages plus English and tests culturally grounded knowledge, it scored 93.1%.

The Claude Opus 5.5 system card also notes that the model usually replies in the user's language, and that output quality still varies by language. Anthropic's safety tests were run in seven languages: Arabic, English, French, Hindi, Korean, Mandarin Chinese, and Russian.

A customer writing in Arabic at 11 pm, or in Hindi from a phone, expects the same answer quality as someone writing in English. These scores are a good sign, but test your own questions in the languages your customers actually use.

What It Means for Customer Support Chatbots

Coding benchmarks say little about how a model handles a customer asking about a refund, but the system card has several findings that apply directly to support. If you are still comparing tools, our list of the best AI chatbot for customer service is a good place to start.

Your Instructions Do Real Safety Work

This was the most useful number I found. When Anthropic tested harmful requests with no system prompt at all, the model gave a harmless response 94.50% of the time. Inside claude.ai, with Anthropic's own system prompt in place, that rate went up to 99.51%. At the same time, it refused harmless requests only 0.03% of the time on the API.

So the instructions you give your chatbot are not just about tone. They change how safely it behaves. Take the time to write instructions that cover your rules, your banned topics, and what the bot should do when it does not know the answer.

Keep Your Knowledge Base Clean

The system card describes an issue Anthropic found in early versions of the model: it was more likely than Opus 5 to follow instructions hidden inside text a user pasted in, such as a README or an email. Anthropic says the final version, together with product changes, mitigated this.

Your bot reads whatever you put in its knowledge base, so only feed it content you trust and review imported documents before training.

Honesty and Citations

Anthropic reports that the model is its strongest or tied-strongest model on most honesty measures in its behavioral audit. On a closed-book factual test it posted a net score of 0.58, the highest of any Claude model, though within the error bars of the Mythos models. One nuance: on MASK, which tests whether a model holds its position under pressure, it scored below Opus 5 and Sonnet 5. So keep answers grounded in your own content, with sources your team can check.

Tone and Brand Voice

According to the system card, the model is especially good at keeping a consistent character and avoiding condescension, but it lags other Claude models on humor and creative output. For support, that is a fair trade. Calm and clear beats clever when someone is waiting on an order. And if your bot also books appointments or checks orders, it is closer to an agent than a chatbot, which we explain in AI agent vs chatbot.

Healthcare and Regulated Industries

On HealthBench Professional, the model scored 65.6, ahead of GPT-6 Astra at 63.4 and Opus 5 at 59.8, according to the Claude Opus 5.5 system card. That does not make any chatbot a medical advisor, but it is a useful signal for clinics using a bot for bookings and general questions.

Want to test different AI models on your own customer questions before your visitors ever see them? Create your free Agentency account and set up a testing model in minutes.

Opus 5.5 Around the World

Here is what stands out for four markets searching for this model.

United States

According to Anthropic's system card, the company worked with the US Center for AI Standards and Innovation (CAISI), part of the National Institute of Standards and Technology, to measure the model's cyber and biology capabilities and safeguards. You can read more about CAISI's role on the NIST CAISI page. For US businesses that need data to stay in the country, Amazon Bedrock offers a US inference profile and Google Cloud offers a US multi-region.

United Kingdom

UK businesses can use the EU inference profile on Amazon Bedrock or the Europe multi-region on Google Cloud. Anthropic also added watermarking to comply with the EU AI Act, which is relevant if you sell into the EU. If your chatbot handles customer data, the UK Information Commissioner's Office explains how data protection law applies in its guidance on AI and data protection.

India

For Indian businesses, the MILU result is the one to watch. The model scored 93.1% across ten Indian languages, including Hindi, Bengali, Tamil, and Telugu, plus English. Hindi was also one of the seven languages used in Anthropic's safety testing. If your chatbot collects customer details, it is worth reading the Ministry of Electronics and Information Technology's material on the Digital Personal Data Protection framework.

Australia

Amazon Bedrock offers an Australian inference profile, and the model also runs in the Asia Pacific (Melbourne) Region on Bedrock's newer endpoint. For privacy, the Office of the Australian Information Commissioner has published guidance on privacy and the use of commercially available AI products, which applies directly to businesses running a third-party chatbot.

What I Noticed Using the New Opus on Real Work

I used Anthropic Opus 5.5 to help research this article, so here is what I actually saw rather than what the launch page promised.

First, it handled long documents well. It read the full 230-page system card, about 62,000 words, and pulled out findings none of the top-ranking articles mention, like the 42-language score and the Toolathlon result.

Second, it pushed back. Before writing anything, it pointed out that my planned angle contradicted one of our own help articles. That saved me from publishing something wrong.

Third, it was careful with data. It noticed that most rows in my keyword export were about older model comparisons, not this release, and told me instead of treating them as real demand.

Choose the AI Model Behind Your Agentency Chatbot

Back to the release speed. A new model every few weeks is great for AI, but exhausting for a business. You should be able to try a new model without rebuilding your chatbot, and see what it costs before you switch. That is why Agentency now lets you pick the AI model for chatbots on a per-agent basis, because the right AI model for chatbots depends on your customers, not the leaderboard. Each agent has two slots:

• Live model: the model that answers your visitors
• Testing model: an optional second model used only in your Playground and Test Chat, so you can compare answers while visitors keep getting the live one Every model in the list shows what one reply costs in message credits. A 1X model uses one credit per reply, a 2X model uses two, and an 8X model uses eight. Today the list includes the Agentency AI model, which is our recommended default, along with options such as GPT-OSS 120B, GPT-6 Luna Pro, Qwen3.8 Flash, GLM-5.3 Flash, DeepSeek V4.1 Flash, and Gemini 3.8 Flash. The Free plan stays on the Agentency AI model.

Conversations also keep their model. A chat stays on the model that answered its first message, so changing the model never switches a conversation halfway through, and every transcript shows which model answered.

The Toolathlon result is why this matters. The best AI model for chatbots is the one that answers your customers correctly, in their language, at a cost that fits your volume, and the only way to find it is to test.

How to Choose an AI Model in Agentency

Step 1: Open your chatbot's settings

Open the chatbot you want to change from your dashboard and find the AI model card, which shows the model that writes its replies.

Step 2: Turn on testing with a different model

Switch on "Test with a different model". From that point, replies in your Playground and Test Chat use the testing model, while visitors keep getting the live one.

Step 3: Pick a testing model and check its cost

Choose a model and check its credit label, such as 1X, 2X, or 8X, so you know what each reply costs.

Step 4: Ask your hardest questions

Run the questions your team is tired of answering, in every language your customers use, and compare them with your live model. The Unanswered questions tab in Knowledge shows what visitors asked that your agent could not answer, which makes a ready-made test list.

Step 5: Promote the winner to live

When a model clearly does better, set it as your live model and save. New conversations use it, while existing ones keep theirs. Choosing an AI model for your chatbot this way means you switch only when the numbers are on your side.

Final Thoughts

Claude Opus 5.5 is a real step up from Opus 5: cheaper per token, faster, clearer, and strong across languages. But its own system card shows it is not the best model at every task, and that is the most useful lesson here for a business. Do not chase the newest name. Test on your real questions and keep the model that serves your customers best.

Ready to find the right model for your customers? Start free with Agentency, connect your website, and test your first agent today.

About the Author

Omar El Bahr is a Senior Digital Growth Specialist working on SEO, content, and digital growth, and a contributor to the Forbes Communications Council. He writes about AI chatbots, search, and how growing businesses can put new technology to practical use. Connect with Omar on LinkedIn.

Disclaimer

This article is for general information only. Model features, prices, and benchmark results come from Anthropic and third-party sources as of September 24, 2026, and may change. Always check the official documentation before making technical or purchasing decisions.