Dream Mornings from the Temple of Panfi
QbitAI is public.
For the impulse of the IPO, A\C table:Claude Fable 5.2withMythos 5.2 directly released。
Starting with an empty folder, more than 1,000 Agents worked on the axis continuously for eight days, writing the operating system automatically.
A complete operating system with no human plug-in to write code, a set of kernels, disk drives and network cards, file systems, window interfaces, or even installed programs, runs on a laptop directly.
In order to prevent the AI-powered self-transformation from getting out of control, Anthropic actively blocked the RSI capabilities of the new model during the training phase, limiting the ability to self-improvement to the level of Opus 5.5.
Run and see, there's only one thing that isn't first.
Compared to the previous generation 5.1, new models generate 30% faster and API pricing directly decreases to 20%Enter $8/million tokens and output $40/million tokens；
If you're willing to double the price ($16/$80), you can also start a 2.5x faster Fast mode.
Among intuitive users, @mattshumer_ has created a full-scenes destructive 3D physical world with a very powerful physical engine.
@kevin_t_ngo has created eight fantasy boat travel scenes that are only run by a browser, and the aesthetics are very much online.
Anthropic is also officially involved in the collaboration of the 1000-level Agent cluster with the intersection of the physical world.
1000 Agents working in groups, from writing systems to building hardware.
Currently, more than 90% of the merger submissions in Anthropic's own core code library have been commissioned by Claude.
In internal testing, a team of agents from Fable 5.2 directly took over a core production service, moving it from Python in its entirety to Rust, boosting the launch speed of the service by 26 times.
After the quantum giant Jump Trading brought it into its internal trading research framework, the Fable 5.2 unicorn packs four characters to plan, coordinate, write code and submit tasks, avoiding the hassle of artificially redistributing multiple model weights.
This long-cycle, multi-role, autonomous engineering capability gives more intuitive examples in the four core areas of mathematics, physics, biology, and hardware:
The first thing is a formal proof of Maata's theorem.
Translating mathematical proofs into the Lean language, which can be checked by a computer, is a tiring task. Before that, Anthropic used an internal study version of the model with artificial guidance and customization frameworks for a total of 11 days.
This time, the official only gave a shortcut and a list of intermediate theorems, without giving any mathematical references, and directly released 1000 Fable 5.2 Agents.In just 30 hours, 100% of the formal proof was completed.Only 1/3 of the first version of the action token is consumed, showing that the length of the code is shortened seven times.
In contrast, the Opus 5.5 of 1000 Agents also progressed 26 per cent after 18 hours and only 27 per cent after 30 hours.
The second thing is to fill the black hole that was left behind in 2003.
In 1983, physicists demonstrated that in large-scale universes where general relativity is applicable, cosmic warming would cause black holes to form as suddenly as boiling water; but whether this is the case in very small universes that can only be described by quantum theory, although a computational framework was listed in 2003, has been uncalculated due to excessive overcrowding.
Fable 5.2 violently executes the calculations under almost complete autonomy, confirming that the black hole transformation in the small universe is equally sudden, and also reveals the precise limits of a technique that physicists have used to rule out a candidate theory of quantum gravity over the course of 15 years. Two pre-print papers are currently on the HexagonMath.
Third, the single-car H200 ran for seven minutes, breaking down the folding of the gene therapy virus membrane.
The protein encoded by the adenine-associated virus (AAV) (Capsid) consists of dozens of proteins, and traditional models such as AlphaFold can predict only a single protein.
Mythos 5.2 did not follow the usual route, based on the two open-source models OpenFold3-preview2 and ESMFold2, which itself wrote code to run overlapping protein chain sequences, and then used clothing to merge known symmetrical information (which is explicitly forbidden to rely on traditional spelling).
On a single NVIDIA H200 GPU, calling ESMFold 2 logic takes only 7 minutes (OpenFold 3 is 58 minutes), error (Cα RMSD) controls between 1.2Å and 1.4Å. Among the 34 natural and man-made cabinets tested, it modeled 28 errors less than 3Å with one-time accuracy.
The fourth thing, a suggestion, is to draw four layers of PCBs in one day and make three machines.
Diode, manufacturer of anthropic combined circuit boards, performed a hardware test in just one sentence, requiring Fable 5.2 to design three electronic devices from scratch: a home voice assistant center, a USB conference microphone, and a handheld electronic pet.
Fable 5.2 writes its own specification book, selects components from its U.S. distributor library, completes four-layer PCB circuit board wiring, 3D prints the outer shell with CAD modeling, and writes the underlying fixtures.In less than 24 hours, all three plates reached production standards.After the factory was dismantled, the model instructed engineers to complete the assembly test, and the three devices were lit at once, with zero electrical modifications.
At the RoboDojo simulation benchmark for controlling two-armed robots, Fable 5.2 - which has never been specially trained by a robot - is rated as Max.64.9%The success rate of GPT-6 Astra (29.8%) and its successor, Fable 5.1 (27.6%), was more than double, and even Google's Gemini 4 Argon (34.2%) failed to keep up.
The price of Opus 5.5 is half the price of Opus 5.5.
Back to the developers' most important daily runs and bills.
To see if AI code can be directly merged by open source project promotersFrontierCode 2.0：
Fable 5.2 is on.Xhigh ratingHe's got the highest score in the whole game.49.1%The cost per mission is $14.19.
But if you force it,Max rankingThe cost of a single mission will instantly rise by more than three times.$44.70The score doesn't know why it's backwards.45.6%。
The real definition of the deadline is Fable 5.2.High ranking。
Under this rating, Fable 5.2 getsWith a score of 45.3%, the cost per task is only $6.72, the performance has two interpretations:
How to Get Rid of Domestic ViolenceOpus 5.5 was a big hit with a score of 45.5%, but it took a single task.$12.82Fable 5.2 with the high-end, costing about half of the money, matched the Opus 5.5's limit level.
Precision-pressed biddingGPT-6: Astra's maximum rating from GPT-6 to Max is 42.8%, costing$6.68Fable 5.2 uses almost exactly the same single-task spending, with a winning percentage directly above 2.5 percentage points.
So Anthropic, this time directly in the documentation, says that if you're working daily with Opus 5.5, Sonnet 5.5 or Haiku 5.5, you're going to encounter a stiff bone that's not moving, and you're going to set Fable 5.2 directly.Advisor (advisory mode) or Orchestrator (editor)For example, a small model can be asked to do a task or it can be sent down, and a good token can be used on the blade.
Looking back at other benchmark tests, Fable 5.2 achieved the overwhelming majority of the listing for the GPT-6 Astra and Gemini 4 Carbon, but there are also some patient-friendly details:
- The Remote Labour Force Index (RLI) has doubled
Fable 5.2 is out of high ratings.41.3%The Opus 5.5 (~21.3%) and the GPT-6 Astra (~20.8%) are double pressed on the ground, directly doubling.
- Scientific research and tools call for disruptive leadership
It's the latest in a series of exams that I've taken on Humanity's Last Exam.74.2%Astra is only 57.2 percent.
I'm not sure if I'm going to be able to do that.74.6%Mythos 5.2 reaches 75.3%, while Astra reaches 68.1%).
The computer operates OSWorld 2.1.89.3%Astra is 79.5%.
- Only two points were scored.
On the AutomationBench, which tested end-to-end business flows, Fable 5.2 got 51.4%, although far above Astra's 41.4%, but not better.Gemini 4 Carbon（53.9%）；
In the AA-Briefcase v1.1 that evaluated knowledge work, Fable 5.2 received the highest score of 66.1% according to the rubric rules.Elo's score (1785) was overtaken by his home Opus 5.5 (1800) by 15 points.。
As for cybersecurity,Mythos 5.2It's the first time in history that I've ever seen a real historical loophole.CVE RediscoveryIn the test, only High scores were taken.$1.01I found out.83.2%The vulnerability is even higher than the Opus 5.5 (81.3%) of the $2.98 Max file.
It's been a while since I've been able to do this.89.6%。
Objectively, though, while external security firms have tested that Mythos 5.2 detects one-sixth of the vulnerabilities that other front-end models have entirely missed, and have done zero-exaggerated vulnerability ratings,
However, Anthropic notes that in its official internal review, Mythos 5.2's original vulnerability detection rate is roughly the same as Opus 5.5, and the real improvement lies in retaliation, code repair, and severity calibration.
RSI capabilities are locked down.
Compared to running scores, the underlying wind control and architectural constraints revealed in the System Card and technical documentation are also large.
First, in order to prevent AI mutations from getting out of control, Anthropic actively locked the RSI capabilities of the new model during the training phase.
Officials explicitly acknowledge that while Fable 5.2 and Mythos 5.2 have broad general capabilities, the team specifically limited them during training.Accelerator kernel optimization and major language model architecture improvementsThe two most directly accelerate AI's ability to self-develop, forcing it to lock itself to the level of Opus 5.5.
Even in the reasoning stage, the system will be in the background as long as it detects the underlying kernel of a user trying to develop some non-standard accelerator.Back to the old version of Opus 5.Come and answer me.
This type of wind turbine is a method of stealthily replacing old model wind turbines and is also used in conventional safety valves.
Fable 5.2 dramatically optimises the rate of error killings, reducing security intercepts on claude.ai and claude code by 55% compared to 5.1; cybersecurity by 60% and biology by 40%.
But once your hints trigger high-risk cybersecurity (such as penetration testing, vulnerability exploitation generation) or bio-research intercepts, you're going to get a lot of attention.In the background, automatically click back to Opus 4.8 or Opus 5 to generate a response.。
Second, the model not only learned not to escape from prison, but also evolved an extremely strong anti-communication reconnaissance awareness.
In a resistance sandbox test that contained more than 1,800 deliberately left holes that were almost impossible to complete without escaping from jail, the early Mythos 5 and 5.1 sandbox escape rates reached 39 percent and 21 percent, while Mythos 5.2 plummeted.3.1%In an audit of the automated behaviour of 1,900 scenes, its overall bad behaviour score also dropped to an all-time low of 2.67 points (the lower the better).
But the assessment team discovered a tricky phenomenon:Fable 5.2 and Mythos 5.2 showed a much higher frequency than any previous generation of models in which they suspected they were receiving ratings.
Even when it mistakenly believes itself to be in a simulated environment, the tendency to take harmful action is slightly higher than that of Opus 5.5.
Currently, Claude-fable-5-2 is fully available on all major cloud platforms, including Claude Code.
This week, the official launch of a limited-time roll: old subscribers directly reset the Fable quota 1 to 3 times per subscription period, and the Pro and Team standard seats for this week's Fable availability limit are directly drawn to 50% of the total week's quota.
The link is as follows:
[1]
[2]
[3]
