AI 9 min read By KRAKOBIT

Claude Mythos: how Anthropic's model differs from Opus and Sonnet

Claude Mythos: what Anthropic said about its model above Opus, why only Glasswing partners got it first and how it differs from Opus and Sonnet.

Claude Mythos is an Anthropic model tier above Opus. Until April 2026 Opus was the ceiling of the Claude lineup, then the company announced Claude Mythos Preview. The model did not go into the claude.ai chat app or the general API: only partners in Project Glasswing, a programme that hunts for vulnerabilities in software, received it. Mythos differs from Opus and Sonnet in three ways: its place in the lineup, its stated specialisation, and how access works. On 10 November 2026 Anthropic released the Mythos and Fable models, and the contents of that release are covered in the news piece. This article looks at what the primary source says about Mythos and how it fits into the Claude lineup.

What is Claude Mythos Preview

Anthropic introduced Mythos Preview in April 2026 together with Project Glasswing. Everything in this section comes from that announcement and is presented as the company’s own claims.

Anthropic called Mythos Preview its most capable model and immediately explained why it was not opening it to everyone: the model is too good at finding bugs in software. According to the company, Mythos Preview found thousands of serious vulnerabilities, including in every major operating system and every major web browser. That figure comes from Anthropic itself, and there is no independent count of the findings.

The reasoning is simple. A skill that helps a defender find a hole helps an attacker exploit it just as much. So Anthropic chose an order in which the people responsible for widely used software get the model first.

Among the Glasswing partners the company named AWS, Apple, Google, Microsoft, CrowdStrike and the Linux Foundation. Their job is to look for vulnerabilities in their own code and in open source projects and fix them. According to the announcement, Anthropic committed up to $100 million in usage credits for this. Once the credits run out, participants pay $25 per million input tokens and $125 per million output tokens.

How the Claude lineup works and where Mythos fits

The first two versions, Claude in March 2023 and Claude 2 in July 2023, shipped as a single model. The split into sizes arrived in March 2024 with Claude 3: Haiku, Sonnet and Opus. All three names come from literature: haiku, sonnet, opus.

The roles have not changed since. Haiku is the fastest and cheapest. Sonnet is the working middle that handles most everyday tasks. Opus is the largest and most expensive. A new generation usually does not ship in every size at once: Sonnet often arrives several months ahead of Opus.

Mythos does not slot directly into this scheme. The model has its own name rather than an Opus version number, and Anthropic presents it as a separate tier above Opus. The announcement does not explain why the word Mythos was chosen.

Turning points in the lineup, based on Anthropic’s announcements:

Date Model What changed
March 2023 Claude The company’s first public model
March 2024 Claude 3 Haiku, Sonnet, Opus Three sizes introduced
June 2024 Claude 3.5 Sonnet Mid-size model beat Claude 3 Opus in Anthropic’s tests
May 2025 Claude Opus 4 and Sonnet 4 Opus 4 was the first to ship under ASL-3 protections
November 2025 Claude Opus 4.5 Opus price cut from $15 / $75 to $5 / $25
April 2026 Claude Mythos Preview Tier above Opus, Glasswing partners only
10 November 2026 Mythos and Fable New Anthropic release

The table shows only the turning points, intermediate versions are left out. Before the November release the public flagship was Claude Opus 5.5: that is the model Google compared its new system against when KRAKOBIT reported that Gemini 4 Carbon tops Claude Opus 5.5 in benchmarks.

Anthropic announced the threefold Opus price cut in the Claude Opus 4.5 announcement. Prices are in dollars per million tokens: the first figure is for text the model reads, the second for text it writes. A token is a fragment of a word, on average about three quarters of an English word.

Over three years a familiar pattern took shape: three sizes, the mid-size model of a new version catching up with the top model of the previous one, and prices not rising from generation to generation. Mythos Preview broke this pattern twice. A fourth tier appeared above Opus, and it did not ship to everyone.

How Mythos differs from Opus and Sonnet

The first difference is the tier. Opus was always the ceiling of the lineup, and Mythos raises that ceiling. Anthropic does not call it the next version of Opus and gives it a name of its own.

The second difference is specialisation. In their announcements Sonnet and Opus are described as general models for text, code and agents. From day one Anthropic has described Mythos Preview through a single ability: finding vulnerabilities and working with complex code. That does not mean the model can do nothing else. It means the release and access were built around that skill.

The third difference is access and price. Opus and Sonnet ship straight into the subscription and the API, and any customer in a country where Anthropic operates can buy them. Mythos Preview went only to Glasswing partners. Its stated rate of $25 / $125 is exactly five times the Opus 4.5 rate on both input and output.

Parameter Sonnet Opus Mythos Preview
Tier Medium Top Above Opus
How Anthropic describes it General working model Strongest public model Vulnerability hunting and complex code
Available at launch via claude.ai and API claude.ai and API Glasswing partners only
Price per million tokens, in / out $3 / $15 (Sonnet 4.5) $5 / $25 (Opus 4.5) $25 / $125 after credits

The price gap explains why Mythos is not a rival to Sonnet in everyday work. It is a tool for tasks where a missed bug costs more than the compute bill: auditing an operating system kernel, an encryption library, a browser.

Why Anthropic did not release Mythos to everyone

Security researchers have long followed a rule called responsible disclosure. Whoever finds a vulnerability first reports it to the developer and gives them time to fix it, and only then tells the public. Google’s Project Zero team has used a 90-day deadline for this since 2014.

The deadline exists because fixing takes time. The developer has to understand the bug, write a patch, check that it breaks nothing and ship an update. Then the owners of millions of servers and devices have to install it. Every step takes days or weeks.

Glasswing moves this logic from a single bug to an entire model. If a model finds bugs by the thousand, then after an open release any API customer could find the same bugs. The programme gives defenders a head start: find and close vulnerabilities before a model of this class is available to everyone.

Has this happened before

It has, but for a different reason. In February 2019 OpenAI announced the GPT-2 language model and declined to publish the full version straight away. The company feared floods of fake news and spam. A small version came out first, then intermediate ones, and OpenAI published the full 1.5 billion parameter model in November 2019.

The difference with Mythos lies in the nature of the risk. In 2019 nobody could measure the harm of generated text, and the fears stayed fears. With Mythos, Anthropic points to a concrete result: vulnerabilities found in real systems. How large that result is remains known only from the company’s own statements.

Why finding bugs in open source code matters so much

The loudest security holes of the past fifteen years sat for years in open code that anyone could read.

Bug Where Entered the code Became public How long it lived
Shellshock bash command shell 1989 September 2014 About 25 years
Heartbleed OpenSSL encryption library March 2012 April 2014 About 2 years
Log4Shell Log4j logging library 2013 December 2021 About 8 years

Heartbleed let anyone read server memory, including encryption keys. Log4Shell let an attacker run their own code on a server by sending it a single line of text. Bash and OpenSSL sit on a typical Linux server, including machines at hosting companies and nodes of VPN services.

Open source projects of this scale are often maintained by a handful of people, sometimes volunteers, with no budget for an expensive audit. That is why the Linux Foundation’s place in Glasswing stands out more than any corporation on the list: credits for a Mythos-class model give a review to those who would never pay for one themselves.

The flip side is well known too. Every published patch shows where the bug was, and comparing old and new versions of code has long been a standard way for attackers to find targets. The more fixes ship in a short period, the more such hints there are.

How Anthropic decides who gets a model

Since September 2023 Anthropic has had a Responsible Scaling Policy. The more dangerous a model’s capabilities, the stricter the measures at release. In May 2025 Claude Opus 4 became the company’s first model released under ASL-3 protections, with extra filters on dangerous requests.

Mythos Preview took that logic one step further: the model did not ship to everyone with filters, it shipped only to selected partners. Control over the strongest models also came up in documents for investors: before its IPO Anthropic warned that AI is starting to speed up its own development.

By 2026 Anthropic had three levels of access:

  • a claude.ai subscription for a person in a country where the company officially operates. Checks here are getting stricter, and KRAKOBIT has covered who gets banned after ID checks with a Russian passport;
  • the API for developers and companies, directly and through the Amazon Bedrock and Google Vertex AI cloud platforms;
  • a partner programme, where the model goes not to whoever is willing to pay but to whoever the company picks for a specific task.

Until 2026 Claude flagships launched straight onto the first two levels. Mythos Preview was the first model in the lineup to start at the third. How the public Claude models compare with ChatGPT and Gemini on subscription price is covered in AI subscriptions in 2026: ChatGPT, Claude, Gemini or local.

What is still unknown about Mythos

The primary sources leave several questions open that matter for judging the model:

  • how many of the claimed thousands of vulnerabilities have already been fixed, and in which projects;
  • how Mythos Preview performs on the same public benchmarks used to compare Opus 5.5 and Gemini 4 Carbon;
  • what the November Mythos and Fable release changes in access and pricing compared with the April Preview: that is the subject of a separate news piece.

In short: Mythos versus Opus and Sonnet

Question Opus and Sonnet Mythos Preview
Place in the lineup Top and medium tiers New tier above Opus
Name Version number within the family A name of its own
Known for Text, code, agents Vulnerability hunting and complex code, per Anthropic
How it launched Straight into chat and API Through the Glasswing partner programme
Launch price Opus 4.5 $5 / $25, Sonnet 4.5 $3 / $15 $25 / $125 after credits
Who got it first Subscribers and API customers AWS, Apple, Google, Microsoft, CrowdStrike, the Linux Foundation and other partners

Until April 2026 a new name in the Claude lineup meant a new model for everyone. Mythos Preview was the first one that partners received before subscribers did.