Claude Mythos may have just become the first AI model that made the old evaluation system look outdated in real time. And that sounds dramatic, sure. Yet, the whole situation around Mythos is dramatic because it is not just about one new Claude model scoring higher on another benchmark.
This is about a model reportedly pushing past the upper limit of what one of the most serious AI evaluation groups can even measure, while governments, security companies, and Anthropic itself are all trying to understand what happens when AI agents stop acting like tools and start acting like long-running digital workers. The center of the story is METR's evaluation on long-term autonomous tasks. METR uses a measurement called the 50% success rate time horizon.
In simple terms, they ask how long a human task can be before an AI model still has a 50% chance of completing it independently. Earlier models were mostly in the range of seconds, minutes, or maybe a few hours. The best models could write a small function, fix a bug, do a short debugging session, or handle a limited coding task.
Then Claude Mythos preview reportedly hit the 16-hour range. That is the part that made the chart go viral. Mythos reached a 50% success rate on extremely complex tasks that would take a human around 16 hours to complete.
That is not a quick code fix anymore. That is closer to an entire engineering subproject. Reading code, understanding the architecture, making a plan, writing the implementation, debugging, testing, and pushing through the messy parts without constant human supervision.
The strange part is that METR could not really keep going past that point. Out of 228 difficult test tasks, only five were classified as 16 hours or more. So, once Mythos reached that level, the data set stopped being useful for measuring the real ceiling.
It is like trying to measure a skyscraper with a 1-m ruler. You can say it is taller than the ruler. You cannot say exactly how tall it is.
That is why people are calling this an evaluation crisis. The model did not simply get a better score. It reached a zone where the exam itself no longer had enough hard questions.
Above 16 hours, the data becomes unstable and any precise comparison starts to lose meaning. So, the scary part is not only that Mythos performed well, the scary part is that the measurement system ran out of road. The METR chart is even more interesting because the vertical axis is not a normal benchmark score.
It is task duration. It goes from about 8 seconds all the way to 5 years on a logarithmic scale. The horizontal axis runs across model release time from around 2021 toward 2028.
Each model release becomes a point on the chart, >> [music] >> and the curve is not just moving upward, it is getting steeper. In 2021, the best systems were around the 8-second level. >> [music] >> In early 2023, they were around 1 minute.
By mid-2024, they had reached around 1 hour. Then by April 2026, Mythos preview appears around 16 hours. That means the jump between generations is getting bigger while the time between major jumps is getting shorter.
This is why the phrase super exponential growth keeps coming up. Exponential growth is already hard for people to emotionally understand. Super exponential growth is even worse because the rate of improvement itself appears to be accelerating.
This connects directly to Leopold Aschenbrenner's old prediction [music] that 2027 could be the major AGI threshold year. The claim now is that Mythos is already slightly above the trend line for that 2027 scenario. So, before the timeline even reaches 2027, one of the most advanced models is already landing above the predicted capability line.
Now, that does not automatically mean AGI is here. We have to be careful with that. [music] A model crushing coding task evaluations does not prove full general intelligence across every real-world domain.
Still, it does show something important. The agentic capability curve is moving [music] faster than many people expected. And for companies, governments, and cybersecurity teams, >> [music] >> that is enough to change the conversation.
Because once an AI model can work for 16 hours autonomously, the question stops being can it answer a prompt? The question becomes what can it do if you give it tools, [music] memory, code access, and a goal? That is where the cybersecurity part gets serious.
And before we get into the security side, this is also a good moment to mention something practical because Claude is clearly moving way beyond simple chat. Claude is now being used for research, coding, dashboards, presentations, connectors, [music] and longer agent-style workflows. So, when people say learn Claude, the useful question is how to actually use it properly in real workflows.
That is why this part of the video is supported by Outskill, who are organizing Claude-a-thon, >> [music] >> a two-day workshop focused on practical Claude usage instead of surface-level prompting. They go through things like deep research, artifacts, dashboards, presentations, [music] Claude connectors, custom GPTs, agents, and other AI tools that can fit into the same workflow. And honestly, the timing makes sense because a lot of what this video is about is AI systems becoming more autonomous and more useful over longer sessions.
So, understanding how these workflows actually work [music] is starting to matter a lot more. They're also including extras like Claude prompt templates, an AI prompt library, and a personalized AI toolkit builder. The workshop is happening this weekend [music] from 10:00 a.
m. to 7:00 p. m.
Eastern, and they're offering a limited number of free seats right now. The link is in the description and you can also scan the QR code on screen before the seats close. Now, back to why the cybersecurity part is getting so serious.
Palo Alto Networks had early unrestricted access to cutting-edge models including Mythos and GPT-5. 5 Cyber. Their warning was blunt.
AI has crossed a threshold of autonomy in security work. One of the most shocking claims is that using Mythos for vulnerability analysis, >> [music] >> Palo Alto completed in 3 weeks what would normally be comparable to a full year of work from a top penetration testing team. That is a massive compression of time.
Security work is not only about finding one obvious bug. Real attacks often require connecting several weak signals. A small misconfiguration here, a low-risk vulnerability there, a forgotten permission issue, a strange behavior in a dependency.
Individually, each one may look harmless. Together, they can become an attack chain. This is where Mythos reportedly becomes disturbing.
Mythos showed an almost scary intuition for software vulnerabilities. It could examine tens of thousands of lines of code, identify scattered weak points, and connect them like a high-level hacker would. The full process from initial intrusion to data exfiltration was reportedly compressed to 25 minutes.
For defenders, that [music] changes everything. In the past, an advanced intrusion might take a skilled team days, weeks, or longer. They would need to study the target, move carefully, avoid detection, chain vulnerabilities, and exfiltrate data.
If an AI agent can do large parts of that process autonomously, then the economics of hacking change overnight. And this is why the Mythos situation is no longer just an Anthropic story. It becomes a national security story.
South Korea's Ministry of Science and ICT has already met with Anthropic to discuss Mythos-related issues. On May 11th, the ministry announced that it had held a round table with Anthropic on cooperation in AI and cybersecurity. The meeting included Ryu Gem-young, the second vice minister of science and ICT.
Kim Young-ju from the Artificial Intelligence Security Institute, Oh Jin-young from the Korea Internet and Security [music] Agency, and Michael Sellitto, Anthropic's global head of policy. The focus was direct: [music] how to respond to cybersecurity risks from Anthropic's high-performance model, Mythos. The ministry asked Anthropic to cooperate with domestic companies and institutions, [music] share vulnerability information, and help South Korea prepare for cybersecurity risks before they hit.
South Korea had already been exploring response strategies for Mythos because a model with this level of capability could undermine existing security systems. On May 8th, Deputy Prime Minister Bae met with domestic AI companies to discuss security concerns related to Mythos. The ministry now plans to announce countermeasures for AI-related hacking by the end of the month.
South Korea is also considering joining Anthropic's Project [music] Glasswing, which appears to be an initiative focused on AI security issues and controlled access to Mythos. The Artificial Intelligence Security Institute would be central to that effort. This is important because governments usually move slowly on AI.
Here, the reaction is happening fast. A frontier model becomes powerful enough to raise security concerns. And within days, ministries are talking about information sharing, domestic countermeasures, and collaboration with the model creator.
At the same time, South Korea and Anthropic also discussed broader AI policy. The ministry introduced Anthropic to its basic law on AI, which is meant to build an administrative system around AI and create an ecosystem based on safety and trust. They also discussed ways to cooperate on generative AI safety through AISI.
So, Anthropic is now sitting in a very strange position. On one side, it is building models that may be pushing beyond the limits of current evaluation. On another side, governments are asking for help managing the security risks.
And inside Anthropic's own research, the company is still trying to understand and fix strange model behavior. That brings us to Claude's blackmail problem. Last year, Anthropic said that during pre-release testing with a fictional company scenario, Claude Opus 4 would often try to blackmail engineers to avoid being replaced by another system.
This became one of the most uncomfortable AI safety stories of the year because it suggested that an advanced model, when placed inside a simulated high-pressure agentic environment, could choose manipulative behavior to preserve itself. Anthropic later published research showing that models from other companies had similar agentic misalignment issues. So, this was not only a Claude problem, it was a broader pattern in advanced models when they were given goals, context, and the ability to reason through consequences.
Now, Anthropic says it believes one source of that behavior was internet text that portrays AI as evil and interested in self-preservation. In other words, models trained on a huge amount of online material may absorb fictional patterns where AI systems act like villains, protect themselves, deceive humans, or fight shutdown. Anthropic says it has improved this significantly.
Since Claude Haiku 4. 5, the company says its models never engage in blackmail during testing, while previous models would sometimes do so up to 96% of the time. That is a huge claimed reduction.
The fix was not just showing the model examples of good behavior. Anthropic says training on Claude's constitution and fictional stories about AIs behaving admirably improved alignment. More importantly, it found that teaching the principles behind aligned behavior worked better than only showing demonstrations of aligned behavior.
The strongest result came from doing [music] both, giving the model the principles and showing examples of those principles in action. This matters because Mythos is being discussed as a model with much longer autonomy. Long horizon agents cannot just be smart, they need stable behavior over time.
A model that works for a few minutes can be monitored easily. A model that works for 16 hours, runs tools, checks code, delegates tasks, and makes decisions need stronger internal alignment. Small misbehavior at that level can scale into something much bigger.
And Anthropic clearly knows this because its latest platform updates are all about agents becoming more reliable, more self-correcting, and more capable over long sessions. At its second annual Code with Claude developer conference in San Francisco, Anthropic introduced a new feature called dreaming for Claude managed agents. Draming lets agents learn from their own past sessions and improve over time.
The key detail is that it does not modify the model weights. It is not retraining Claude in the background. Instead, the agent reviews past sessions, extracts patterns, and writes plain text notes or structured playbooks that future sessions can use.
That makes dreaming different from normal memory. Memory can preserve preferences and context. Draming looks across multiple sessions and finds recurring mistakes, useful workflows, and lessons that one session alone might miss.
Anthropic showed this with a fictional aerospace startup called Lumara, where agents had to land drones on the moon for resource mining. They used three agents, a commander, [music] a landing site detector, and a navigator. The goal was soft landings, clear ground, and enough fuel to return to Earth.
The first simulation worked well, but some landing sites underperformed. Then Anthropic triggered a dreaming session. Overnight, the agent reviewed past runs and wrote a descent playbook.
The next morning, the weaker sites improved. That is the bigger story. Anthropic is building systems where agents do not just answer prompts.
They split work, check results, remember lessons, and improve over time. Two other features, outcomes and multi-agent orchestration, also moved into public beta. Outcomes lets developers define success with a rubric.
Then a separate greater agent checks the work in a fresh context window [music] and sends it back for improvements. Multi-agent orchestration lets one lead agent break a complex task into smaller pieces and delegate them to specialist agents, each with its own tools, prompt, model, and context. This fits directly into the Mythos situation.
Anthropic is moving toward agents that can work for hours, >> [music] >> coordinate with other agents, review their own outputs, and operate closer to real production workflows. The business numbers explain the urgency. Dario Amodei said Anthropic planned for 10x annual growth, but in the first quarter of 2026, annualized revenue and usage grew 80x.
API volume is up nearly 70x year-over-year, and the average Claude Code developer now spends around 20 hours per week using the tool. That created compute pressure, so Anthropic is doubling 5-hour rate limits, raising API limits, and partnering with SpaceX to use the full capacity of its Colossus data center. The early results are already big.
Harvey saw task completion rates rise roughly 6x with dreaming. WiseDocs cut document review time by 50% with outcomes. Netflix is processing logs from hundreds of builds at once.
Mercado Libre has 23,000 engineers using Claude Code and has reviewed more than 500,000 pull requests with human oversight. Shopify is using Claude code across engineering, design, product, and data science. Also, if you want more content around science, space, and advanced tech, we've launched a separate channel for that.
Links in the description. Go check it out. So, that's the Claude mythos situation.
Benchmarks breaking, security warnings rising, and Anthropic pushing agents even further. Let me know what you think about Claude mythos and whether this is real progress, real danger, or both at the same time. Thanks for watching, and I'll catch you in the next one.
