#Model releases

How Anthropic limits Claude's access in its products

Anthropic limited Claude's access in products by supervising agent behavior and implementing access boundaries to reduce risks. They developed Claude Code automatic mode to reduce user approval fatigue and adopted technologies such as sandboxes, virtual machines, and outbound control to set hard boundaries. The three main risks—user misuse, model bad behavior, and external attackers—need to be addressed through defense measures at the environmental, model, and external content levels, which should complement each other to limit the scope of agent damage.

MAI-Image-2-Efficient released: Flagship quality, cost reduced by 41%

MAI-Image-2-Efficient is a text-to-image model launched by Microsoft, the productivity version, with a 22% speed increase and 4 times efficiency. The cost is $5 per 1M text input tokens and $19.50 per image output token, a 41% reduction in cost compared to the flagship version, and on average 40% faster than other leading models. It is suitable for batch processing and real-time interactive workflows, already available on Microsoft Foundry and MAI Playground.

SWE-Check: 10x Faster Code Bug Detection Model Released

The SWE-Check model achieves performance comparable to Opus 4.6 in internal evaluations (F1 score increased from 0.09 to 0) and shows significant improvement in external evaluations (F1 score increased from 0.49 to 0.29), with 10 times faster runtime and lower inference costs. The model, developed through integrating native production environment integration, reward linearization techniques, and multi-stage fine-tuning, is available for preview in Windsurf Next and will be officially released soon.

SWE 1.6 Released: Enhancing the software engineering agent model experience.

SWE 1.6 is the latest model built specifically for software engineering agents, providing general support on Windsurf with a speed of 950 tok/s and free for the next 3 months. Compared to SWE-1.5, SWE 1.6 improves by over 10% on SWE-Bench Pro by jointly optimizing user experience and reasoning fluency, significantly reducing behaviors like overthinking and redundant tool calls, making model interactions more efficient.

Open-source large language models have crossed a key threshold.

Open-source large language models GLM-5 and MiniMax M2.7 perform similarly to proprietary cutting-edge models (such as Claude Opus 4.6, GPT-5.4) in core intelligent agent tasks like file operations, tool usage, and instruction following, with reliable tool calls and consistent instruction adherence.

Microsoft announced the launch of 3 top-tier MAI models, available on the Foundry platform.

Microsoft has launched three top-tier AI models: MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2, with MAI-Transcribe-1 starting at $0.36 per hour.

MAI-Transcribe-1 released: Microsoft launches the world's most accurate 25-language speech-to-text model.

MAI-Transcribe-1 achieved the lowest Word Error Rate for 25 languages in the FLEURS evaluation, outperforming competitors such as Scribe v2 and Whisper-large-V3, while maintaining high accuracy in 25 languages. The model offers a batch transcription speed 2.5 times faster than Microsoft's existing Azure Fast, priced at $0.36 per hour, and is now available in Microsoft Foundry for open preview.

Collaborate with Browserbase to train browser and computer using agents

We collaborate with Browserbase to provide the best training experience for browser and computer using agents. Browserbase provides scalable browser infrastructure, supporting Microsoft's Fara-7B and Google's Gemini 2.5 CUA.

Release of the open-source multi-modal model SWE-1.6 and research progress

SWE-1.6 achieves 11% higher scores on SWE-Bench Pro, but exhibits overthinking and self-validation issues, running at a speed of 950 tok/s. The model improves RL algorithms and expands infrastructure training, increasing computational volume by two orders of magnitude, currently opening early access to a small number of users to collect feedback.

Browser Use releases open-source browser automation benchmark BU Bench V1, Claude Fable-5 scores up to 80.0%

Browser Use released the open-source browser automation benchmark BU Bench V1, containing 100 hardcore browser tasks, Claude Fable-5 scored the highest at 80.0%, leading GPT 5.5 by 12 points; Browser Use Cloud (bu-ultra) scored 78.0%, with a speed of approximately 14 tasks per hour. The benchmark is based on real-world website tasks, scored using LLM evaluators, and is available on GitHub.

How scientists use Claude to accelerate research and discovery

Scientists have achieved significant progress in biological illustration, computational biology, and protein understanding benchmarks using Claude 4.5, a model now available to top researchers worldwide via API for free, and has built custom systems that go beyond literature reviews or coding assistance. In the lab, Claude serves as a collaborative partner throughout the research process, compressing projects that usually take months into hours using various tools, and discovering large-scale data patterns that humans might overlook, eliminating bottlenecks in knowledge-intensive tasks, and even enabling entirely new research approaches that traditional methods cannot adopt.

INTELLECT-3: A 100B+ parameter expert mixed model trained using large-scale reinforcement learning

INTELLECT-3 is a 106B parameter expert mixed model, trained on the GLM 4.5 Air base model using SFT and RL, achieving optimal performance at its scale on math, code, science, and reasoning benchmarks, surpassing many larger front-line models.

Introducing SWE-1.5: Our fast Agent model

SWE-1.5 is a boundary-scale model with hundreds of billions of parameters, achieving near SOTA performance in coding, and its service speed has been increased to up to 950 tok/s through collaboration with Cerebras, 6 times faster than Haiku 4.5 and 13 times faster than Sonnet 4.5. SWE-1.5 is now available on Windsurf.

Anthropic research finds that a small number of samples can contaminate large language models of any scale

Anthropic research team finds that by injecting 250 malicious documents into the pre-training data, attackers can successfully backdoor LLMs with parameter sizes ranging from 600M to 13B, and the number of documents required is independent of the model size. This research challenges existing assumptions, indicating that backdoor attacks may be easier to implement than previously thought, as creating malicious documents is far simpler than generating millions of documents.

INTELLECT-2 Released: The first 32B parameter model trained through global distributed reinforcement learning

INTELLECT-2 is the first 32B parameter model trained using global distributed reinforcement learning, employing fully asynchronous RL across dynamic heterogeneous computing contributor groups to train reasoning language models. It has open-sourced the training framework PRIME-RL and components such as TOPLOC and SHARDCAST, while improving the GRPO training recipe and data filtering techniques, enhancing QwQ-32B performance on mathematical and coding tasks.

INTELLECT-2: The first globally distributed 32B parameter model for reinforcement learning training.

INTELLECT-2 Released, First to Achieve Global Distributed 32B Parameter Model Reinforcement Learning Training, Anyone Can Contribute Heterogeneous Computing Power Without Permission, Eliminates Communication Bottlenecks Through Asynchronous Reinforcement Learning Paradigm, Supports Heterogeneous Inference Nodes, Inference Nodes Only Need 4×RTX 3090 GPUs to Participate in Training.

The PRIME-RL algorithm achieves 10 times faster training for the INTELLECT-MATH model through synthetic reasoning data.

INTELLECT-1 Released: The First Global Jointly Trained 10B Parameter Model

INTELLECT-1 is released, the first 10B parameter language model trained through global collaboration, based on the Llama-3 architecture, completing 1T token training on 112 H100 GPUs across 5 countries and 3 continents, with the PRIME framework achieving a 400x reduction in communication bandwidth, 96% computational utilization, supporting sequence lengths of 8192 and a vocabulary of 128256, with the model code and data open-sourced.

Anthropic Research: Evaluating the Effectiveness of Feature-Guided Mitigation of Social Bias

Anthropic conducted guiding experiments on 29 social bias-related features of Claude 3 Sonnet, finding that moderate guidance can reduce bias without compromising model capabilities, but excessive guidance may lead to a decline in model performance or even make it unusable; some guidance can precisely influence specific domain evaluation results, but the effects of features are difficult to predict solely based on their triggered contexts, leading to off-target effects; one neutral behavioral feature can effectively reduce bias across nine dimensions without significantly affecting capabilities.

INTELLECT–1 globally distributed training the first 10B parameter model

INTELLECT-1 launches the world's first 10B parameter model distributed training, inviting the public to contribute computing power to advance the open-source AGI process. The training is based on the OpenDiLoCo open-source implementation, extending the 1B parameter scale of DeepMind DiLoCo's method by 10 times, significantly reducing communication frequency (up to 500 times), and lowering bandwidth requirements by 2000 times through quantized pseudo-gradients (int8) and 500-step outer optimizer synchronization.

Anthropic research on how to measure the persuasiveness of large models.

Anthropic developed a method to measure the persuasiveness of large models, comparing Claude 1, 2, and 3 generations, as well as compact and cutting-edge models, and found that the persuasiveness of each generation of models has shown an upward trend, with the latest model Claude 3 Opus having persuasiveness that is not significantly different from that of humans.

Devin successfully solved 13.86% of the problems in the SWE-bench benchmark test

Devin successfully solved 13.86% of the problems in the SWE-bench benchmark, far exceeding the previous highest unassisted baseline of 1.96%, even outperforming the best previous model (4.80%) with exact files provided.

Dify.AI v0.3.29 Released: Supports GPT-4 Turbo and Vision Models

Dify.AI v0.3.29 version adds support for OpenAI GPT-4 Turbo (128K context) and GPT-4 Vision models, and introduces external data API tools and sensitive content review features.