Anthropic expands access to Claude Cowork for web and mobile devices

French scientists have created a non-invasive interface capable of converting brain signals into typed text on a keyboard. Magnetoencephalography (MEG) and electroencephalography (EEG) methods are used to collect data. The process of converting brain signals into characters is carried out using the Brain2Qwerty neural network model. The best results were achieved using MEG, where the average error rate was 29%, whereas using EEG, this figure reached 65%.

Advancements in Brain-Computer Interfaces

In recent years, significant progress has been observed in the field of 'brain-computer interfaces,' which allow people with severe paralysis or amyotrophic lateral sclerosis to regain the ability to communicate. These neurointerfaces read motor cortex activity and translate it into text at a speed close to that of normal speech. However, the most effective systems require surgical intervention to implant electrodes directly into brain tissue, which carries health risks and difficulties in long-term chip support. Therefore, current research is focused on creating systems that do not require surgery.

Comparison of Data Collection Methods

The most common method for non-invasively obtaining information about brain activity is EEG, which records electrical potentials on the scalp via a special cap. Nevertheless, the signal is significantly attenuated and distorted by the skull bones, which reduces the signal-to-noise ratio. MEG offers a higher signal-to-noise ratio because it measures the magnetic fields of the brain, which are subject to less distortion from other tissues. Although MEG has rarely been used for such tasks previously due to the bulky equipment, researchers decided to apply both methods in conjunction with machine learning.

Experimental Setup and Model

Under the direction of Jarod Levy from the Meta AI* division in Paris, work was conducted on non-invasive text decoding. Thirty-five healthy Spanish speakers participated in the experiment. Participants were sequentially shown short sentences in Spanish, one word at a time. After the text disappeared, participants had to recall and type the phrase on a QWERTY keyboard. A non-magnetic version was used to avoid interference from keyboard components. Brain activity was recorded during typing using EEG or MEG, and each key press was synchronized with the brain data. The collected dataset included hundreds of thousands of presses for both methods.

Brain2Qwerty Neural Network Architecture

To decipher the collected data, the authors trained the Brain2Qwerty neural network model, which consists of three parts. In the first stage, the algorithm segments continuous recordings of brain activity into half-second segments around each key press. These fragments first pass through a convolutional module, which extracts spatio-temporal features from the signals and accounts for individual differences among subjects. Next, the resulting representations are fed into a transformer architecture module, which analyzes the sequence of presses within the entire sentence, using context to refine predictions for individual characters. In the final stage, the system applies a pre-trained language model to correct potential typos. This three-stage structure ensures a direct translation of brain signals into text format.

Testing Results and Conclusions

Testing demonstrated a clear superiority of MEG over EEG. The model trained on magnetic signals showed an average symbol recognition error rate of 29% (which includes substitutions, insertions, and deletions to match the predicted string with the correct one). The corresponding rate for electrical signals was 65%, making EEG poorly suited for practical application. For the most successful participants using MEG, the error rate often dropped to 18%. Furthermore, the model successfully reconstructed entire phrases, even if they were not present in the original training set. Additional analysis showed that the neural network had well learned the layout of keys on a standard keyboard, and errors most often occurred when confusing adjacent letters, indicating the reading of brain motor commands. The authors also noted that the model effectively recognizes frequently used words while maintaining the ability to guess rare vocabulary.

Future Prospects of the Technology

Currently, the system only generates a result after the user has finished typing the entire sentence. Developers intend to modify the architecture to enable real-time operation. The next critical step will be testing the system on paralyzed patients who can only mentally form text. The main technical obstacle at present remains the need to use a large stationary magnetic field scanner. Nevertheless, the authors express hope that the development of compact sensors will eventually allow for the creation of wearable helmets for such interfaces.

A series of analyses focused on financial transactions and educational interest demonstrates that the paid use of the Claude assistant, developed by Anthropic, has shown constant growth among consumers in the United States.

Claude Growth Data

This information was extracted from millions of card transactions and technology education platforms. The data reveals that this advance is occurring in an environment of strong competition with ChatGPT, although the latter maintains a significant lead in terms of total user volume and revenue.

Growth trends are also visible in search indicators and course consumption, in addition to changes in the regulatory framework affecting the company's models, generating greater attention to the company's recent performance in the artificial intelligence sector.

Consumer Expansion in the US

Surveys conducted by the data intelligence company Indagari, based on billions of anonymous card transactions, indicate that subscriptions and payments linked to Claude have continuously progressed throughout the year 2026 in the United States. The study covers about 28 million consumers and points to a notable increase in the amount spent on services related to the Anthropic system.

According to this data, the accumulated growth of the paid consumer group for Claude reached approximately 75% since the beginning of 2026, showing monthly expansion. This movement remained stable even after a peak recorded in March, a period when the company had decided to limit military and surveillance uses in certain contexts.

Educational Interest and Search

Additionally, the digital learning platform DataCamp reported a significant increase in demand for materials related to Claude. The company stated that the term began to dominate internal searches, and interest in courses about the system grew substantially in a short time, surpassing even generic searches about artificial intelligence in some segments of users.

Comparison with ChatGPT

Despite this advance, ChatGPT remains the main artificial intelligence tool aimed at the end consumer, both in terms of user quantity and projected revenue. Reports from market intelligence companies confirm that the user base of OpenAI's product remains considerably larger than that of Claude.

Sensor Tower reports show that Claude demonstrated relevant expansion across various platforms throughout the year, but it is still far behind in scale compared to its rival. This disparity is attributed to ChatGPT's established reach in the global market. However, industry analysts observe that this difference may be decreasing in terms of growth speed and public attention, particularly among users looking for alternatives for specific tasks.

Regulatory and Business Impact

Anthropic's recent performance also occurs in a scenario of regulatory tensions. There is information that US authorities imposed restrictions on the use of the company's more advanced models in certain regions and contexts, resulting in the temporary removal of some versions from the market.

Among the affected models were the versions named Mythos 5 and Fable 5, which possess advanced cybersecurity capabilities. This determination affected the global availability of these systems. To date, Anthropic has not issued an official comment on the growth data or on the effects of these restrictions, but the collected information suggests a parallel expansion of consumer interest and regulatory pressure.