Anthropic has two frontier models: Claude Opus 4.7, which is public, and Claude Mythos Preview — a model so powerful that Anthropic has decided not to release it publicly due to potential cybersecurity risks.
Mythos is accessible only through Project Glasswing as of the time of this writing, and only 52 invited organisations can use it.
Nevertheless, thanks to leaks and official messages we have Mythos benchmarks, so in this article we’re comparing it to Opus 4.7, Opus 4.6 and GPT 5.4 to try and imagine what the next generation of AI is likely going to be like.
Mythos wins every shared benchmark, but the margin is uneven. On coding, the gap is 6–13 points, while on scientific reasoning the gap is sometimes 0.4 points — effectively tied.
Mythos is probably the basis for the next Claude flagship. Anthropic has stated that Opus 4.7 is a testbed for safeguards that will eventually allow Mythos-class models to ship more broadly — possibly labeled as Claude 5.
The biggest gap is in cybersecurity. According to CyberGym, a cybersecurity AI benchmark, Mythos is about 24.8% more capable than Opus 4.6 — it has discovered thousands of zero-day vulnerabilities, including a 27-year-old bug in OpenBSD.
But raw benchmark numbers don’t give you the full picture, so let’s break them down and look at what they mean in practice.
SWE-bench Verified
What is measures: How many bugs out of 500 the model can successfully solve, given a curated set of human-validated GitHub issues from popular Python repositories.
What the model has to do to complete it: Read the codebase, understand the bug report or feature request, and implement a working patch.
Mythos score: 93.9%
Opus 4.7 score: 86.6%
Opus 4.6 score: 80.8%
Opus 4.7 was already able to handle over 85% of bugs flawlessly, but Mythos pushes that to nearly 94%. In other words, it can resolve almost any real-world bug on the first try — and that’s a big deal in practice.
For example, if fixing a bug used to take 3 attempts at ~2K tokens each (~6K total), Mythos can often solve it in one pass (~2K tokens). That’s a ~65–70% reduction in token usage.
SWE-bench Pro
What it measures: Harder variant of SWE-bench — multi-language codebases (Python, JavaScript, Go, Rust, Java), full engineering pipelines, less handholding.
What the model has to do to complete it: Navigate polyglot codebases, work across language boundaries, handle complex dependencies.
Mythos score: 77.8%
Opus 4.7 score: 64.3%
Opus 4.6 score: 53.4%
GPT-5.4 score: 57.7%
Top publicly available models rarely clear 60% here, and that’s too hit and miss for the real-world. Mythos will be able to handle codebases where previous models had failed: complicated Kotlin and Rust services, Java and COBOL banking backends. We’ll see even wider enterprise adoption of AI code in places where companies were perhaps apprehensive to add it.
Terminal-Bench 2.0
What it measures: Command-line and sysadmin proficiency. Package installs, service configuration, CI/CD debugging, devops automation.
What the model has to do to complete it: Operate inside a terminal like a human engineer would, chaining commands, and debuting outputs based on error messages.
Mythos score: 82.0% (92.1% with 4-hour timeouts)
Opus 4.7 score: 69.4%
Opus 4.6 score: 65.4%
GPT-5.4 score: 75.1%
This is the one coding benchmark where Opus 4.7 trails GPT-5.4 — by almost 6 points. For agents that live in the terminal — Kubernetes debugging, server provisioning, log analysis — GPT-5.4 is currently the better tool. Anthropic has positioned Opus 4.7 higher up the stack: code generation, review, multi-tool orchestration.
OSWorld-Verified
What it measures: Autonomous desktop GUI operation. Clicking buttons, navigating menus, filling forms, reading screens.
What the model has to do to complete it: Interpret what's on screen, decide where to click, handle UI that changes as the task progresses.
Mythos score: 79.6%
Opus 4.7 score: 78.0%
GPT-5.4 score: 75.0%
Mythos is ahead by 1.6 points, but that’s unlikely to have a noticeable impact in practice. More interestingly, the reason Opus 4.7’s gained points in this benchmark over Opus 4.6 is that its vision resolution was increased by 3×. Since we’re not seeing any meaningful gains between Mythos and Opus 4.6, perhaps the two models share the same vision architecture?
BrowseComp
What it measures: Agentic web research and synthesis. The model searches, reads, correlates sources, produces a summary.
What the model has to do to complete it: Build a research strategy, execute searches, read and evaluate sources, synthesize across conflicting information.
GPT-5.4 score: 89.3%
Mythos score: 86.9%
Opus 4.6 score: 83.7%
Opus 4.7 score: 79.3%
The outlier in this comparison. Opus 4.7 actually regresses 4.4 points from Opus 4.6, and GPT-5.4 leads everything — including Mythos. Anthropic hasn't publicly explained the regression, but the Opus 4.7 system card mentions that MRCR (the long-context benchmark) is being phased out in favour of Graphwalks, which may reflect a broader tradeoff in how the model handles very long research contexts.
GPQA Diamond
What it measures: Graduate-level science questions in physics, chemistry, biology. Written by PhD experts.
What the model has to do to complete it: Answer questions that most university-educated non-specialists would get wrong.
Mythos score: 94.6%
Opus 4.7 score: 94.2%
Opus 4.6 score: 91.3%
GPT-5.4 score: 94.4%
The benchmark is approaching saturation. For any real-world science reasoning task, all four models are effectively equivalent.
CharXiv Reasoning
What it measures: Scientific figure and chart interpretation.
What the model has to do to complete it: Read graphs, tables, and diagrams from arXiv papers and answer questions about what the figure shows.
Mythos score: 93.2%
Opus 4.7 score: 91.0%
Opus 4.6 score: 84.7%
Opus 4.7's 6.3-point jump over 4.6 is backed by the 3× resolution support — the model can now read chart annotations that were too small for 4.6 to see.
Humanity's Last Exam
What it measures: How an AI handles questions that sit at the edge of human expertise. 3,000 questions written by domain experts — professors, researchers, Olympiad coaches — specifically designed to be unsolvable by current AI models
What the model has to do to complete it: Two versions of the test. Without tools, the model has to answer purely from what it learned during training. With tools, the model can search the web, run code, look things up — the way a human expert would when faced with a hard problem.
Mythos score (with tools): 64.7%
Mythos score (without tools): 56.8%
GPT-5.4 score (with tools): 58.7%
Opus 4.7 score (with tools): 54.7%
Opus 4.6 score (with tools): 53.3%
Mythos without tools (56.8%) is only 2 points below GPT-5.4 with tools (58.7%) — which means Mythos reasons nearly as well from pure memory as GPT-5.4 does with the whole internet to help.
One caveat worth understanding: HLE questions are publicly available on the internet, which means AI training data probably includes some of them. Anthropic flagged this directly — they noticed Mythos performs surprisingly well on HLE even when running at "low effort" (spending less compute per answer), which is a red flag for memorization rather than genuine reasoning.
In practice, for work that requires expert-level reasoning in specialised domains — legal research, academic writing, technical due diligence — Mythos will be measurably better than anything else on the market, as of the time of this writing.
FAQ
What is Claude Mythos?
Claude Mythos is Anthropic's unreleased frontier model, confirmed in April 2026. It's more capable than Claude Opus 4.7 and sits in a new internal tier Anthropic calls Capybara — a step above the Opus line.
Who can access Claude Mythos?
At the time of writing, Mythos is only accessible to Project Glasswing participants — AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, or one of the 40+ additional organisations, who were invited to test critical software infrastructure.
What is Project Glasswing?
Project Glasswing is Anthropic's initiative to give select organisations early access to Claude Mythos for defensive cybersecurity work. Launch partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. An additional 40+ organisations that maintain critical software infrastructure also have access. Anthropic committed $100M in usage credits and $4M in donations to open-source security groups. Participants pay $25 / $125 per million input/output tokens after the initial research preview.
Will Mythos eventually be released publicly?
Anthropic hasn't committed to a timeline, but has stated intent to release Mythos-class capabilities broadly once safeguards are in place. Based on Anthropic's release cadence and Polymarket odds, a Mythos-derived public flagship is most plausible in Q2–Q3 2026.
Which is better, Claude Mythos or Claude Opus 4.7?
Claude Mythos is better than Claude Opus 4.7 by quite a margin. The new model beats Opus 4.7 almost in every benchmark, in some cases, like coding, by 20–25%.
Bottom Line
Without knowing what major AI companies are working on behind closed doors, it’s hard to say definitively, but it’s quite possible that Mythos is currently the most advanced text generation AI model — we had a rare opportunity to see what it could do early on due to multiple leaks from Claude, which allowed us to compare it directly with the current flagship, Opus 4.7. Unsurprisingly, Mythos outperforms it by a noticeable margin.
Key Takeaways
We’ve compared Claude Mythos against Opus 4.7, and Mythos leads every shared benchmark by quite a margin, particularly in terms of coding and cybersecurity.
However, Mythos is currently restricted to 52 organisations due to its cybersecurity capabilities, so it is impossible to test them side by side in the real world at present.
Opus 4.7 is Anthropic's best publicly available model.
A model with Mythos-class capabilities is likely to be released in Q2–Q3 of 2026.
.webp)