Anthropic has restored access to its latest models. Fable 5 returned to users worldwide on July 1, while the Mythos 5 model was once again made available only to a group of U.S. organizations. The company also introduced Claude Science, a research tool for scientists.
Fable 5 is now available on the Claude platform, in Claude.ai, Claude Code, and Claude Cowork. For Pro, Max, Team, and selected enterprise accounts, Anthropic will count its usage against only up to half of the weekly limit through July 7, after which it will be accessible through user credits. The company promises to enable access through AWS, Google Cloud, and Microsoft Foundry as soon as possible.
New mechanisms to prevent misuse
On this occasion, the company explained the logic behind its safeguards. It uses a so-called defense-in-depth approach, which layers several mechanisms. None of them is perfect on its own, but together they make the model very difficult to misuse. Some safeguards teach the model to refuse dangerous requests, while others retrospectively analyze patterns of misuse.
At the core of the defenses are classifiers, smaller automated systems that monitor conversations to determine whether someone is asking the model to perform potentially harmful work. Anthropic deliberately set their sensitivity to block even requests that are probably harmless. It calls this buffer a safety margin. For Fable 5, the company made it wider than for any previous launch. Users will notice this when the model occasionally refuses even a reasonable and harmless request.
The margin also helps counter attempts to bypass safeguards. Most such attempts are limited, unlocking only one specific behavior and nothing else. Anthropic categorizes safeguard bypasses by severity. These range from minor bypasses that reach no further than the safety margin, through harmful ones, to the most dangerous, which unlock an entire class of harmful behavior. According to the company, no one has yet discovered a universal bypass for Fable 5.
Anthropic has also deployed an improved safety classifier. When a user’s request triggers this barrier, they receive a warning and their query is routed to the Opus 4.8 model. Researchers from the U.S. Center for AI Standards and Innovation (CAISI) evaluated both the old and new safeguards and agreed that they are exceptionally strong. However, tighter safeguards come at a cost. The new classifier is more likely to flag innocent requests during routine programming and code debugging as suspicious. The company promises to continue improving the system so that it can better distinguish genuine misuse from legitimate work.
Shared rules for the entire industry
Together with Amazon, Microsoft, Google, and other partners in the Glasswing program, Anthropic is developing a shared framework for assessing the severity of safeguard breaches. The industry currently lacks a standardized way to describe how dangerous a given bypass is. As a result, developers do not know which findings deserve the most attention, and governments lack guidance on when to intervene.
The proposal scores each breach according to four criteria: how many additional capabilities it gives an attacker compared with commonly available tools, how many different attack tasks the technique works for, how much human effort is required to turn it into a real-world attack, and how easily the technique can be obtained. In the most severe cases, such as an attack targeting power grids or banking systems, the company promises to deploy initial countermeasures immediately. It is also establishing a team that will continuously monitor the channels where people report new breaches. At the same time, it is launching a program on the HackerOne platform where security researchers will be able to submit their findings for review.
New Claude Science tool aims to make scientists’ work easier
Anthropic has introduced Claude Science, a workbench for scientists. It is an application that brings together tools and libraries commonly used in research, produces traceable outputs, and offers flexible access to computing power. The company launched it in a test version for Pro, Max, Team, and Enterprise plans on macOS and Linux systems.
Scientific work is often time-consuming. Researchers have to work through dozens of databases, each with its own structure, switching between tools such as PubMed, Jupyter, R, or a computing cluster terminal. Claude Science combines these scattered tools into a single environment. The user works with an orchestration agent that has access to more than sixty ready-made skills and connectors for genomics, proteomics, structural biology, and cheminformatics. The agent can launch additional agents and has its work overseen by a supervisory agent that checks citations and calculations and corrects errors on its own.
Traceability is one of its strengths. When the tool creates an image, it includes the exact code and environment that generated it, a description in plain language, and the complete message history. This makes the work verifiable and reproducible even months later. Users can request chart edits simply in words, such as asking for the grid to be removed or the axis to be switched to a logarithmic scale. For demanding computations, such as protein folding or genomic analyses of massive datasets, the tool creates a plan, asks for permission, and sends the task to infrastructure the laboratory already uses. Sensitive data does not have to leave the systems where it is stored. Only the portion needed for the current step is sent to Claude. The application uses skills from the NVIDIA BioNeMo Agent Toolkit and connects to models such as Evo 2, Boltz-2, and OpenFold3.
Scientific applications
Several teams tested the tool over a period of several months. Manifold Bio, a company that designs drugs targeting specific tissues, used it to select targets for its latest experiments. For each tissue and target, the tool assessed surface expression, trafficking, and safety, then ranked candidates according to criteria that Manifold derived from its own data. According to the company, it was able to handle the entire process from start to finish.
Neuroscientist Jérôme Lecoq of the Allen Institute used Claude Science to build a set of about twenty custom skills for writing lengthy review studies. Sub-agents read thousands of papers, extract their main claims and key numerical findings, and store them in a database. Writing such a review previously took his team up to two years. He now has around ten of them, many more than one hundred pages long, with citations verified by supervisory agents. Epidemiologist Stephen Francis of the UCSF Brain Tumor Center is using the tool to investigate the genetic basis of brain tumors. He can complete analyses that previously took weeks in roughly one-tenth of the time. His group independently verified the results and confirmed that the tool can work both quickly and reliably.
