Claude Mythos

Anthropic's most powerful AI cutting-edge model is not released to the public and is only available to partners through Project Glasswing.

large language model

https://claudemythosai.io/

Visit Website

In-depth Report

Claude Mythos Preview is a new cutting-edge AI model released by Anthropic on April 7, 2026. It is positioned as the highest level of the Claude product line, and its performance far exceeds the existing Opus series. The model scored 93.9% on SWE-bench Verified and 94.6% on GPQA Diamond, making it Anthropic’s strongest model to date. However, due to security concerns, the model is not publicly available and is only available to select partners through the Project Glasswing program.
Anthropic is an AI security and research company dedicated to building reliable, explainable, and controllable AI systems. The company was founded in 2021 and is headquartered in San Francisco, USA. The main founders are from key employees of OpenAI. Anthropic features AI-aligned research at its core, and its Claude family of models has been an industry leader in safety and inference capabilities. On April 7, 2026, Anthropic joined forces with 11 technology giants including Apple, Nvidia, Microsoft, Amazon Cloud Service (AWS), Google, and the Linux Foundation to launch the Project Glasswing project and released Claude Mythos Preview. The internal code name of this model is "Capybara". The name comes from the ancient Greek "Mythos", which means "narrative/discourse" and refers to the story system by which human civilization understands the world.
Claude Mythos Preview has demonstrated excellent capabilities in both security and general dimensions. In the security direction, this model can discover thousands of high-risk zero-day vulnerabilities in all mainstream operating systems and browsers. It can independently write browser vulnerability exploits and remote code execution exploits. It can connect multiple vulnerabilities in series to achieve a privilege escalation chain from ordinary users to root. It can also reconstruct and exploit desymbol table binaries. These abilities are not specially trained, but emerge as a natural result of the comprehensive improvement of coding and reasoning. In terms of general direction, Claude Mythos Preview performs equally well. The model scored 93.9% on SWE-bench Verified, ranking among the highest in the industry in terms of code generation and repair. In terms of academic reasoning, GPQA Diamond scored 94.6%, Terminal-Bench 2.0 scored 82.0%, BrowseComp scored 86.9%, and the token consumption was only one-fifth of Opus 4.6, reflecting significant efficiency improvements. Corresponding to technical capabilities are limitations in user experience. Users cannot access the Claude Mythos Preview directly and must go through a partner in the Project Glasswing program. Currently, it can be accessed through three channels: Google Cloud Vertex AI (Private Preview), Amazon Bedrock and Microsoft Foundry.
The pricing mechanism for Claude Mythos Preview is more complicated. According to the disclosed information, the standard price of the model is US$25 per million tokens for input and US$125 per million tokens for output, which is an increase of approximately 67% compared to Opus 4.6’s input of US$15 and output of US$75. However, this is the standard price after the $100 million quota is used up, and partners can initially use the quota provided by Anthropic. Anthropic has committed up to $100 million in model usage and is donating $2.5 million to Alpha-Omega and OpenSSF under the Linux Foundation, and $1.5 million to the Apache Software Foundation. This business model is mainly aimed at enterprise-level partners rather than ordinary individual users.
Since Claude Mythos Preview is not released to the public, ordinary users cannot experience it directly, so there are no extensive user reviews. However, looking at industry reaction, there are two distinct views on the model's release. Some argue that Anthropic is handling high-risk AI capabilities responsibly, and that this approach deserves recognition. Another view is that this is a marketing strategy to enhance brand value through "mystery". Others worry that this sets a precedent for AI companies to unilaterally decide which capabilities are “too dangerous” and not make them public. Another view is that the fundamental reason why Claude Mythos Preview is not made public may not be "too dangerous", but that Anthropic hopes to use it as a differentiated competitive advantage to obtain higher profits through corporate cooperation. In any case, due to the lack of actual usage feedback from ordinary users, these opinions remain at the level of speculation.
From an industry perspective, the release of Claude Mythos Preview marks a new stage in the competition for cutting-edge AI models. The model scored 93.9% on SWE-bench Verified, an improvement of 13.1 percentage points compared to Opus 4.6’s 80.8%. This gap is rare in the history of AI model evaluation. GPQA Diamond score of 94.6%, demonstrating excellent academic reasoning skills. However, the strategy of not releasing the model to the public has also triggered industry thinking. If the most powerful AI capabilities are only in the hands of a few technology giants, will this create a new technology monopoly? The 12 core partners of Project Glasswing include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks and Anthropic itself, covering many key areas such as cloud services, mobile phone chips, network security, and finance. Whether this cooperation model will exacerbate the Matthew Effect in the field of AI deserves continued attention.
The biggest controversy with Claude Mythos Preview is its "non-disclosure" strategy. Anthropic’s official explanation is that the model’s security capabilities are too strong, and the consequences of falling into the hands of attackers will be serious, so it is not suitable to be made public until the development of new security guardrails is completed. The plausibility of this explanation is controversial. Supporters say it's a responsible approach, opponents say it's a business ploy. From a risk perspective, if the most powerful AI system is controlled by only a few institutions, it may bring multiple risks. The first is the risk of technological monopoly, as a few companies may have a say in AI capabilities. Secondly, there is the risk of security review. AI companies can unilaterally decide which capabilities are “too dangerous” and not make them public. This lacks a transparent supervision mechanism. The third is the risk of capability gap. If the gap between cutting-edge models and public models continues to widen, it may lead to polarization in the AI field.
Claude Mythos Preview is not suitable for ordinary individual users. First, the model is not publicly released and cannot be directly accessed by individual users. Secondly, even if accessed through cooperative channels, its high price ($125 per million tokens output) is not suitable for individual users. Third, this model is mainly oriented to security research and enterprise-level applications, and it is difficult for individual users to find suitable application scenarios. For enterprise users, if your company is a partner of the Project Glasswing program or a critical infrastructure organization, you can consider applying for access. This model is suitable for enterprise-level application scenarios that require high-intensity code generation, vulnerability mining, and security research. For average business and personal users, the Claude Opus 4.7 or Opus 4.6 are the more practical choices, and these models remain at the top of the list among publicly available models.
Claude Mythos Preview is the most powerful AI model Anthropic has released to date, demonstrating outstanding capabilities in multiple dimensions such as code generation, academic reasoning, and security research. However, due to security concerns, the model is not publicly available and is only available to select partners through the Project Glasswing program. The strategy has sparked industry discussions about AI safety, technology monopoly and responsible innovation. For regular users who don't have access to this model, Claude Opus 4.7 is still the best option currently publicly available. With the development of AI safety guardrail technology, Anthropic is expected to gradually disclose some of its capabilities in the future, but this process may take a long time.

User Reviews

GammaGainRivera

—

The SWE-bench score of 93.9% is also terrible. Previously, Opus 4.6 only had 80.8%. This improvement is far ahead of competing products.

月光_13

—

Although Mythos is not used, when you see this 244-page system card, you know that Anthropic has really spent a lot of money this time. The technical report is more detailed than the paper.

WKimQ

—

To put it bluntly, I don’t dare to make it public for fear of being hacked and used to do bad things. I can understand this logic, but as a developer I’m still jealous.

DouglasButler_2020

—

I was convinced when I saw GPQA Diamond’s 94.6%, which is basically the strongest academic reasoning model on the planet.

j4b4ynj11

—

The funniest thing is that this model is so powerful that even my own company doesn't dare to use it, so I suggested renaming it Claude Don't Use.

GloriaCampbellII129

—

It feels like Anthropic is playing a big game. It will first make a super model that will not be made public, and then slowly release permissions. I give full marks to this wave of marketing.

Caroline979

—

BrowseComp only uses one-fifth of the tokens of Opus 4.6 to achieve higher results. This efficiency improvement is a bit outrageous.

悠然64

—

The GPT-5.4 Pro next door is still competing with Opus 4.6, but here Mythos is already playing its own game.

jbzo4i

—

After reading Huxiu’s analysis, it was said that this model has emotional response and ethical decision-making capabilities. This is the most terrifying part of thinking about it.

刘凤

—

It's good not to make it public, otherwise it will be penetrated by various prompt attacks.

Terry.Howard369675

—

After all, safety comes first, and Anthropic was scared by the previous incident.

MrFinnMoore_dev

—

It is strongly recommended that ordinary users give up. There is no place for individual users in the list of Project Glasswing partners.

BeverlyKing16822

—

I hope Opus 5 can inherit some Mythos capabilities in the future, even 10% will be enough.

Diana.Lee1684

—

Laughing to death, I can’t even afford the series: the output is US$125 per million tokens, a price that only major manufacturers can afford.

马涛军

—

In fact, it is right not to disclose it. If we really let it go, something might go wrong. AI safety is more important than anything else.

Deborah302

—

Very good, Claude Opus 4.7 is still the most powerful one we can use in public channels, and that’s enough.

Amber_RobertsX044

—

After looking at the evaluation data, Terminal-Bench 2.0’s score of 82% was higher than I expected.

DeFiGuru923

—

The five major technology giants jointly endorsed this, and this battle is too big.

兰花774

—

The internal code name is Capybara. I don’t know what the product manager thinks.

Christian.Foster520

—

It is said to be designed for defensive network security workflow, but what about the offensive direction? I dare not think about it at all.

Claude Mythos

In-depth Report

User Reviews

Similar Products