Anthropic Slows the Price War: Opus 5.5 Costs 40% Less

⏱️ Lectura: 10 min

Anthropic introduced Claude Opus 5.5 on September 22, 2026, the first model in its new 5.5 family. The core fact is simple: it costs 40% less to run than Opus 5 and performs at the level of Fable 5.1 on most tasks.

📑 En este artículo

1. TL;DR
2. What Happened
3. Context and History
4. Technical Details and Performance of Claude Opus 5.5
5. Getting Started
6. Impact and Analysis
7. What’s Next
8. Frequently Asked Questions
   1. What is Claude Opus 5.5?
   2. How much does Claude Opus 5.5 cost compared to Opus 5?
   3. Do I need special verification to use it for biology or cybersecurity?
   4. When will Claude Sonnet 5.5 and Claude Haiku 5.5 arrive?
   5. What does it mean that Anthropic is slowing the pace of the frontier?
   6. How do I confirm a response came from Opus 5.5 and not another model?
9. References

It’s the company’s first release since it publicly called for pausing the pace of the AI frontier, and it arrives pre-evaluated for production by external firms like Frontier Design and METR.

TL;DR

• Anthropic launched Claude Opus 5.5 on September 22, 2026, the first model in the 5.5 family.
• The new model performs at the level of Fable 5.1 on most tasks, according to Anthropic.
• It costs 40% less than Opus 5 for typical workloads, with input tokens at $4 and output tokens at $20 per million.
• Cache reads drop to $0.20 per million tokens, 60% less than Opus 5.
• It generates text more than 30% faster than Opus 5 and completed a 680,000-line code migration in under a day.
• It scored the best result so far in Anthropic’s automated behavioral audit.
• It’s comparable to Claude Mythos 5.1 in biology and cybersecurity, so it uses the same safeguards as Fable 5.1.
• Sonnet 5.5 and Haiku 5.5 will arrive in the coming weeks with similar improvements.

What Happened

Claude Opus 5.5 arrives as Anthropic’s flagship model, replacing Opus 5 at the top of its catalog. The company describes it as comparable in performance to Claude Fable 5.1 on most tasks, but with significantly lower compute cost.

The announcement comes after Anthropic called for slowing the pace of development at the AI frontier, a public stance on the speed at which increasingly capable models are deployed. Opus 5.5 is the first test of that stance. Before its release, it went through external evaluators like Frontier Design and METR, in addition to Anthropic’s internal automated behavioral audit, the most extensive one the company applies.

According to Anthropic, Opus 5.5 achieved the best results in that audit among all the models it has tested so far. It’s less prone to taking hard-to-reverse actions or acting outside the limits set for it, and it resists prompt injection attempts better than Opus 5.

Context and History

The Claude 5.5 family succeeds the generation that included Opus 5, Fable 5.1, and Mythos 5.1 throughout 2026. Every release in that cycle came with verification programs for sensitive domains, biology and cybersecurity, where a more capable model is also a model with greater potential for misuse.

Opus 5.5 inherits that framework. Anthropic says its performance in biology and cybersecurity is comparable to Claude Mythos 5.1, so it deploys the model with the same safeguards it uses for Fable 5.1: open access for general use, but verified access for high-risk research. Organizations can already apply to the Life Sciences Verification Program to use Opus 5.5 in biological research, and in the coming weeks the Cyber Verification Program will expand for verified cybersecurity professionals.

sequenceDiagram
participant O as Organization
participant A as Anthropic
participant M as Claude Opus 5.5
O->>A: requests access to the verification program
A-->>O: validates credentials and approves access
O->>M: sends research queries
M-->>O: responds with Fable 5.1 safeguards active
Note over O,M: the same flow applies to the Cyber Verification Program

That pattern, growing capability paired with growing verification, is how Anthropic has been managing the tension between releasing more useful models and preventing those same models from enabling serious harm.

Technical Details and Performance of Claude Opus 5.5

In internal tests, an evaluator completed a code migration of 680,000 lines in under a day, a job that would have taken an engineering team weeks. Anthropic also tested Opus 5.5 on cutting load times across every page of a web application, and it succeeded in 39 out of 40 attempts, while Opus 5 achieved smaller improvements and, in several cases, altered the application’s behavior while doing so.

In a third test, different Claude models built a video game from a single prompt. Opus 5.5 scored the highest of all for the quality of its graphics and the polish of the final result.

The following table summarizes the benchmarks Anthropic published, comparing Opus 5.5 against Fable 5.1, Opus 5, and competing models (GPT-6 Astra and GPT-5.6 Sol):

Benchmark	Opus 5.5	Fable 5.1	Opus 5	GPT-6 Astra	GPT-5.6 Sol
Terminal-Bench 4.0 (agentic coding)	66.4%	55.8%	52.3%	57.9%	37.3%
FrontierCode v1.1	54.4%	50.3%	48.0%	53.3%	47.5%
CursorBench 4.0	57.8%	51.8%	46.6%	N/A	41.7%
GDPval-AA v2.1 (knowledge work)	1846	1735	1708	1542	1588
AutomationBench	40.0%	31.4%	26.9%	41.4%	28.8%
Humanity’s Last Exam (with tools)	67.7%	65.6%	63.6%	57.2%	N/A
Terminal-Bench-Science 0.1	58.7%	52.6%	29.0%	64.6%	22.4%
OSWorld 2.0 (computer use)	81.8%	80.7%	74.0%	N/A	N/A

Results in multidisciplinary reasoning also rose: Opus 5.5 reached 67.7% on Humanity’s Last Exam using tools, up from 63.6% for Opus 5. In chart recognition (Chartography), it went from 83.4% to 89.0% with tools. These are gains of just a few points over Fable 5.1 on nearly every benchmark, which confirms what Anthropic itself warns in its comparison. At this level of capability, the company says, benchmark margins have become a less reliable guide to real-world differences in use, and in its internal experience the gap between Opus 5.5 and Fable 5.1 is smaller than these numbers suggest.

On safety, Anthropic expanded its alignment tests to cover longer tasks, tasks impossible to complete well, and scenarios modeled closely on real incidents. Full details of that evaluation are in the Opus 5.5 System Card, published alongside the model.

💭 Key takeaway: Anthropic warns that, past a certain level of capability, the gap between models on benchmarks stops accurately reflecting the real gap in day-to-day usefulness.

Getting Started

Opus 5.5 is used through the Anthropic API with the model identifier claude-opus-5-5. The first step is saving the API key as an environment variable.

On Windows (PowerShell):

setx ANTHROPIC_API_KEY "your-api-key"

On macOS and Linux (bash or zsh):

export ANTHROPIC_API_KEY="your-api-key"

With the key set, a minimal API call looks like this:

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2026-01-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Explain what this deploy script does"}]
  }'

That call returns a JSON object with the model’s response. For more realistic, agentic use, it’s better to handle it through the Python SDK and explicitly confirm which model responded before trusting the result:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    messages=[
        {"role": "user", "content": "Review this module and suggest how to reduce its load time"}
    ],
)

assert response.model == "claude-opus-5-5"
print(response.content[0].text)

The assert on response.model is the simplest way to confirm, on every call, that Opus 5.5 is actually being used and not a different model due to a configuration error or a silent fallback.

💡 Tip: if your application caches long prompts (system prompts, documentation, repo context), check the cache read cost in your usage dashboard. At $0.20 per million tokens, that’s where the price cut shows up the most.

Impact and Analysis

The price cut isn’t even across token types. Input and output tokens dropped 20%, to $4 and $20 per million respectively, but cache reads fell 60%, to $0.20 per million. Anthropic notes that cache reads make up most of the cost in agentic and coding work, so the real savings for an agent running for hours over the same repository is greater than the 40% the headline announces.

This isn’t just an incremental model improvement, but a price correction designed for long-running agentic compute, for agents that read the same context over and over during a long session.

For development teams in Latin America who already pay for these services in dollars, the price cut matters twice over. It reduces the direct cost per token and, since Opus 5.5 is faster, it also reduces the billed compute time per task. An agent that used to take minutes to iterate over a large monorepo now finishes faster and uses fewer cache reads per session.

Anthropic paired the launch with two product changes. It raised the five-hour usage limits on Pro, Max, Team, and Enterprise plans per seat, and added a rate-limit reset that paid plan users can save and use whenever they want, instead of losing it if they don’t use it.

The model’s communication also changed. Early users who tried it described its writing as clearer and better organized than Opus 5’s, with the most important information placed up front. One user quoted by Anthropic summed it up this way: “it writes like I write.” For Anthropic, that’s not just a style detail, but a safety benefit: a model whose work is easier to follow and review is also, in practice, a model that’s easier to oversee.

The most real limitation of this launch is access to the highest-risk capabilities. Using Opus 5.5 for cutting-edge biological research or cybersecurity work requires going through the verification programs, something that isn’t achieved by simply creating an API account. And although the model generated text 30% faster than Opus 5, it’s still the most expensive model in Anthropic’s catalog. The price cut brings it closer to Fable 5.1; it doesn’t make it the cheap option.

⚠️ Heads up: Sonnet 5.5 and Haiku 5.5 haven’t been released yet. If your use case doesn’t need top-tier capability, it’s probably worth waiting for those versions before migrating your whole pipeline to Opus 5.5.

What’s Next

Anthropic confirmed that Claude Sonnet 5.5 and Claude Haiku 5.5 will arrive in the coming weeks, with much of the same performance, efficiency, and safety improvements that Opus 5.5 brings. That will determine whether Opus 5.5’s price cut becomes the floor for the whole family or remains a special case for the largest model.

Also still pending is the expansion of the Cyber Verification Program, promised for the coming weeks, which will broaden who can use Opus 5.5 without restrictions for cybersecurity work.

📖 Summary on Telegram: View summary

Try it yourself: run the curl command above with your own API key and compare the response.model field against what Opus 5 returned last week.

Frequently Asked Questions

What is Claude Opus 5.5?

It’s the first model in Anthropic’s Claude 5.5 family, introduced on September 22, 2026. It performs at the level of Fable 5.1 on most tasks and costs 40% less to run than Opus 5.

How much does Claude Opus 5.5 cost compared to Opus 5?

Input tokens cost $4 per million and output tokens $20 per million, 20% less than Opus 5. Cache reads dropped 60%, to $0.20 per million tokens.

Do I need special verification to use it for biology or cybersecurity?

Yes, if the use is high-risk research. Anthropic considers Opus 5.5 comparable to Mythos 5.1 in those areas, so it applies the same safeguards as Fable 5.1, with verified access through the Life Sciences Verification Program and, soon, the Cyber Verification Program.

When will Claude Sonnet 5.5 and Claude Haiku 5.5 arrive?

Anthropic said they’ll launch in the coming weeks, with similar performance, efficiency, and safety improvements.

What does it mean that Anthropic is slowing the pace of the frontier?

It’s Anthropic’s public stance on slowing the pace at which increasingly capable models are deployed. Opus 5.5 is the company’s first release since adopting that stance, and it went through external evaluators like METR before launch.

How do I confirm a response came from Opus 5.5 and not another model?

By checking the model field of the API response object, which should be exactly claude-opus-5-5.

References

• Anthropic: Introducing Claude Opus 5.5: official announcement with benchmarks, pricing, and System Card details.
• METR: one of the external organizations that evaluated Opus 5.5 before its release.
• Anthropic API Documentation: reference for endpoints, models, and parameters to integrate Claude Opus 5.5.
• Wikipedia: Anthropic: general background on the company and its Claude model family.

📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Imagen destacada: Foto de Google DeepMind en Unsplash