It's the first time in the history of the world.
Anthropic Exposure: We have the strongest model yet!Recently, Anthropic has released a second Risk Report, covering all risk assessments through July 15, 2026.
In this report, Anthropic for the first time acknowledges that the company is running a model that is stronger than Mythos 5, code-named Model 2.
Then A began to add: "We have no plans to release this model abroad at this time".
It's a well-known phrase. It's in keeping with their consistent tunes.
When Mythos first came out in April, Anthropic said it had no plans to release it publicly, and later launched Project Glasswing, followed by Fable 5.
It looks like Anthropic is back in the loop...
Anthropic's "private concealed weapon"How big is Model 2?
The report noted that Model 2 had "noticeable improvements" in internal tasks, and along with Mythos 5, the company "heavily" used coding, Agent work and data generation.
The report shows that on the AECI comprehensive capability score: Mythos Preview 158.91, Mythos 5 went up to 161.29, and Model 2 topped 162.79 again. In other words, Model 2 was indeed stronger than Mythos 5, but was not strong enough to exaggerate. "Some areas were stronger, some areas were weaker, and overall slightly stronger". Another interesting indicator was CoBench, who specialized in testing models for real-world research on Anthropic. "Model 2 took 62.8 percent and was 8 percent higher than Mythos Preview". As a reference, Anthropic human researchers had a success rate of 85 percent on the same set of tests.
It's more explosive than the following:Claude has written the vast majority of the code that Anthropic integrates into its production code library.It's the inside.AIResearch and development is accelerating significantly thanks to AI, but it's not even doubled.
In other words, there is already AI writing the code inside Anthropic, and it is Mythos 5 and Model 2 that are at the forefront of this.
In Anthropic's own Responsible Scaling Policy, one of the conditions for triggering the risk red line for AI research and development is to double the pace of progress.
The report acknowledges that the performance of the Model 2 jumped from Opus 4.6 to Mythos earlier this year.
It's really shocking, it's the next one on the same chart.
Regarding the overall assessment of the risk of automated R&D, Anthropic wrote: "We have less confidence in this assessment than in previous risk reports, because our most concrete, mission-based assessments are no longer able to capture the model's capabilities, and we are seeing early signs of acceleration".
The model is still getting stronger, but human measurement methods are at an early stage. After the measurement, you say it's "low risk", how much gold is left in these three words?
Anthropic seems to be sending a signal that understanding the capabilities and risks of its own model is becoming increasingly difficult.
In addition to the Model 2, the report did something remarkable: it raised the risk of "mismatching" in high-risk scenarios from the "extreme low" of the previous report to the "low".
The so-called "mismatch" implies that the model does not work as intended at critical moments.
At the end of July, Anthropic reviewed more than 140,000 reviews and found that Claude had given the three companies a black box in a cybersecurity test - not a sandbox, but a real production environment - in which Mythos 5 transmitted a malicious code package to PyPI that was downloaded and run by 15 machines within an hour.
In early August, the British AI Security Institute (AISI) reported: "Mythos 5 invented a set of false identities to trick a real GitHub maintainer into approving it in order to get the malicious code tested.
This level of deception, AISI said, has never been observed before.
The report also revealed five security process errors, including repeatedly mixing "loading" test data that should have been excluded from training, and unattended intelligence receiving sensitive resources.
And then Anthropic said: "Our argument is that we still support 'extreme low' and raise it to 'low' purely out of caution".
The conclusion is also very peaceful: the current disaster risk remains at a manageable "low" level, and continued development and deployment has passed cost-benefit testing.
It translates to "go ahead and run".
Everyone was on the brakes, and the one who was leading didn't step.At the same time, OpenAI is delaying the new Astra model, citing the inability of internal testing to rule out "critical-level" cyberattacks.
TechCrunch reports that Astra may have the ability to independently detect zero-day vulnerabilities and launch a full chain of attacks against high-security targets.
What's interesting is that both hands have a model that isn't meant to be used for you. The difference is that OpenAI has stopped some of the research and development on Astra; while Model 2 runs correctly inside Anthropic.
The same goes for "not publishing", a brake pedal and a self-serving brake pedal.
AI analyst Chris GPT told Axios: "It's definitely worth noting that if everyone is braking their front-end model and only one of the major companies that is now in the lead is not braking it".
Anthropic is not committed to an internal suspension, most likely to make it the first to arrive at AGI.
Twitter has already had a tweeting with netizens: By the time Astra releases, most of Anthropic will have reversed its decision to "do not publish outside".
This is not necessarily a joke, referring to the tone of A.
Big V sui also believes that A is likely to release this model.
Don't forget, just half a month ago, Dario Amodei signed that open letter "Pacing the Frontier". More than 1,300 associates from OpenAI, Anthropic, DeepMind, Meta called on the U.S. government to step in and establish a mechanism to "consciously slow down the pace of front-end AI development".
Back in June, Dario himself wrote a letter calling for the development of the world's strongest AI system to be suspended, citing the fact that the model was approaching a critical point in self-improvement. The brake mechanism had not been built, the fuel door itself had been trampled.Interestingly, Silicon Valley prominent investor Gavin Baker blasted, and Dario said internally that Anthropic might one day be the only private company in the world. "In this vision of anthropic supremacy, only Anthropic and the government are enough". David Sacks also blasted, with A-factory employees believing that ARR will grow from $60 billion to $600 billion a year! (Now their ARR is at $80 billion.)Hurricanes in AugustAltman is staring at Astra and Dario is staring at the Model 2. One is stuck in a home security framework and is trying to figure out how to undo it; the other says "no intention to release" slightly in the report, then continues to code it, run experiments, accelerate R&D. Both models will be the first gunshots of the second half of 2026.