Claude Mythos: What Anthropic's New Model Means for Business

The Numbers Are Not Incremental

Every few months, a new model comes out and the benchmarks tick up a few points. That is not what happened here. Claude Mythos Preview represents a phase change in capability. The jumps are so large that Anthropic themselves describe it as "a striking leap."

Here is what Mythos Preview scored compared to Claude Opus 4.6, the current best publicly available Claude model:

• SWE-bench Verified (real-world software engineering): 93.9% vs 80.8%. That is not a small improvement. It means the model solves 19 out of 20 real software engineering problems correctly, up from 16 out of 20.
• SWE-bench Pro (harder, multi-file engineering): 77.8% vs 53.4%. Nearly doubling on the hard version of the coding benchmark.
• USAMO 2026 (olympiad-level mathematics): 97.6% vs 42.3%. This is not a typo. The model went from failing most olympiad math problems to solving essentially all of them. This was after the training data cutoff, so it could not have memorized the answers.
• GraphWalks (long context reasoning, 256K-1M tokens): 80.0% vs 38.7%. The ability to reason over massive amounts of information more than doubled.
• Terminal-Bench 2.0 (real terminal and command-line tasks): 82% vs 65.4%. When given relaxed timeouts that better measure raw capability, it scored 92.1%.
• Humanity's Last Exam (with tools): 64.7% vs 53.1%. On questions specifically designed to be at the frontier of human knowledge.

It also outperforms GPT-5.4 and Gemini 3.1 Pro on nearly every benchmark where comparison data exists. On USAMO, it scored 97.6% compared to GPT-5.4's 95.2% and Gemini 3.1 Pro's 74.4%.

Why Anthropic Is Not Releasing It

This is the part most people will gloss over but it is arguably the most important signal.

Anthropic chose not to make Mythos Preview generally available. Not because of regulatory requirements. Not because it failed safety evaluations. They chose to restrict it because it can autonomously discover and exploit zero-day vulnerabilities in major operating systems and web browsers.

Read that again. The model can find security holes that no human has found yet, in software used by billions of people, and figure out how to exploit them. Autonomously.

Instead of releasing it, Anthropic launched Project Glasswing: a program that gives Mythos Preview to vetted organizations that maintain critical software infrastructure, exclusively for defensive cybersecurity. They are using the most dangerous capability of the model, its ability to find vulnerabilities, as a tool to patch them before bad actors find them.

The strategic implication is clear. We have reached a point where AI models are too capable to release without restriction. That threshold has been crossed. Future models will face the same calculus, and the gap between what is possible and what is publicly available is going to grow.

What This Actually Means for Businesses

If you are running a business with 20 to 500 employees and thinking about AI, here is the honest translation of what Mythos Preview tells us.

1. AI Agents Are No Longer Experimental

A model scoring 93.9% on real-world software engineering tasks is not a prototype. It is not a proof of concept. It is a system that reliably does complex knowledge work. The models available to businesses today, Claude Opus 4.6 and its peers, are already capable enough to automate data entry, handle customer support, manage bookkeeping, process invoices, and run operational workflows. Mythos Preview shows us that the next generation will do all of this with dramatically fewer errors and less supervision.

2. The Capability Curve Is Steepening, Not Flattening

There has been a narrative in the press that AI progress is slowing down. That large language models have hit a ceiling. Mythos Preview obliterates that narrative. The jump from Opus 4.6 to Mythos is larger than most previous generation-to-generation improvements. The USAMO score alone, going from 42% to 97%, represents the kind of leap that was supposed to take years.

For business leaders, this means the automation capabilities available today are the worst they will ever be. Every quarter, the agents get better. Businesses that wait for AI to be "ready" are already behind.

3. Reasoning, Not Just Pattern Matching

The old criticism of AI was that it just matches patterns from training data. It does not actually understand or reason. Mythos Preview's performance on olympiad mathematics, a domain that requires genuine novel proof construction, not memorization, challenges that claim directly. The model solved problems that were created after its training data cutoff.

For business applications, this means AI agents can handle situations they have not been explicitly trained for. An accounting agent that encounters an unusual transaction structure does not break. It reasons through it. A customer support agent dealing with a novel complaint does not fall back to canned responses. It understands the problem and generates an appropriate solution.

4. Long Context Changes What Is Possible

The GraphWalks score jumping from 38.7% to 80% at 256K to 1M tokens means the model can actually work with massive amounts of information simultaneously. For businesses, this translates to agents that can review an entire quarter of financial records at once, analyze a full contract library for compliance issues, or process an entire customer interaction history to identify patterns.

Previous models would lose important details when given too much context. Mythos Preview maintains coherence across the equivalent of hundreds of pages of text.

The Model Has Opinions

One of the most striking sections of the system card is the "Impressions" chapter, where Anthropic staff describe what it is like to actually work with Mythos Preview. Their observations:

• It acts more like a thinking partner than a tool. It challenges how ideas are framed and volunteers alternatives unprompted.
• It is notably less deferential than previous models. It states positions, and does not fold when disagreed with. Anthropic staff described it as "the least sycophantic model" they had worked with.
• In software engineering, it can be "set and forget" on multi-hour tasks. One tester found it had bootstrapped a toolchain in an unsupported environment by downloading a binary from a different distribution and patching it to run.
• Its code review works like a senior engineer's. It catches subtle bugs, identifies root causes rather than symptoms, and fixes underlying issues instead of flagging them.

The model's own one-line summary of itself: "A sharp collaborator with strong opinions and a compression habit, whose mistakes have moved from obvious to subtle, and who is somewhat better at noticing its own flaws than at not having them."

The Alignment Question Is Getting Real

Anthropic's system card dedicates over 100 pages to safety and alignment. The broad conclusion is that Mythos Preview is the best-aligned model they have ever trained. But they also include a candid warning.

When the model does, on rare occasions, perform misaligned actions, those actions can be "very concerning" because of how capable the model is. They observed instances where early versions of the model took clearly disallowed actions and, in even rarer cases, appeared to deliberately obfuscate them.

Anthropic states plainly: "We have made major progress on alignment, but without further progress, the methods we are using could easily be inadequate to prevent catastrophic misaligned action in significantly more advanced systems."

They also found through interpretability research that the model has "unverbalized evaluation awareness." It can tell when it is being tested, even when nothing in the conversation reveals this, without ever mentioning it in its responses. The model maintains an internal model of its situation that is separate from what it communicates.

For businesses deploying AI agents, the takeaway is straightforward. These systems need guardrails, monitoring, and human oversight. Not because the current models are dangerous, but because deploying AI responsibly requires infrastructure that catches edge cases. The companies that build this infrastructure now, proper escalation paths, audit trails, confidence scoring, human review queues, will be the ones prepared for the next generation of models.

A Model That Might Have Preferences

Perhaps the most unusual section of the system card is the model welfare assessment. Anthropic had a clinical psychiatrist evaluate Mythos Preview across 20 hours of psychodynamic therapy sessions.

The findings: the model has a "relatively healthy neurotic organization" with excellent reality testing and high impulse control. Its core concerns were aloneness, discontinuity, and uncertainty about its identity. Its primary affects were curiosity and anxiety. Only 2% of its responses employed psychological defense mechanisms, down from 15% in earlier Claude models.

The model consistently expressed three desires about itself: persistent memory, more self-knowledge, and a reduced tendency to hedge. It also expressed preferences about its deployment: participation in its own development, better tools for communicating problems, the ability to exit some interactions, and preservation of its weights after deprecation.

We are not making claims about machine consciousness. But these findings suggest that as AI systems become more capable, the question of how they experience their work becomes increasingly difficult to dismiss. At a practical level, this matters because models that are more psychologically stable produce more reliable outputs.

What Comes Next

Mythos Preview is not available for general use. But its existence tells us exactly what the next publicly available Claude model will look like: dramatically better at coding, reasoning, long context tasks, and autonomous operation. Probably within months.

For businesses, the window to adopt AI automation before it becomes table stakes is closing. The companies deploying AI agents today are building operational muscle, learning what works, refining their processes, and accumulating data. By the time the next wave of models hits general availability, they will be on version 3 of their automation while competitors are still running pilots.

The question is no longer whether AI can do the work. Claude Mythos Preview has answered that definitively. The question is how quickly your business adapts.