AI news — July 25, 2026
The 20 top stories PulseAugur surfaced that day, ranked by signal across labs, papers, and developer communities.
-
Anthropic's Claude Opus 5 halves cost-per-task, maintains token prices
Anthropic has released Claude Opus 5, maintaining the same pricing as Opus 4.8 at $5 per million input tokens and $25 per million output tokens, which is half the cost of Claude Fable 5. The company emphasizes cost per completed task rather than per token, highlighting improved …
-
Anthropic launches Opus 5 AI model at reduced price, with opt-in data retention
Anthropic has released Opus 5, a new iteration of its AI model, at a significantly lower price point than its predecessor, Fable. A key feature of Opus 5 is its opt-in data retention policy, which addresses privacy concerns. This release positions Anthropic to compete more aggre…
-
Zhipu AI's GLM-5.2 leads open-weight models with 1M context window · 1 source tracked
GLM-5.2, a new open-weight model from China's Zhipu AI, has been recognized as the top-performing open model as of July 2026. It achieved a score of 51 in the Intelligence Index v4.1 by Artificial Analysis, placing it fourth overall among all models and first among those with pu…
-
Google Gemini Omni debuts, withholding voice cloning for responsible AI use
Google has introduced Gemini Omni, a new multimodal AI model, but has intentionally withheld its voice-cloning capabilities for video editing. This decision, announced on May 19, 2026, at Google I/O, aims to address concerns around the responsible use of AI in replicating artist…
-
Anthropic's Claude Opus 5 introduces 'effort' parameter for cost-quality tuning
Anthropic has released Claude Opus 5, featuring a new 'effort' parameter that allows users to adjust the trade-off between computational cost and output quality. This parameter, configurable through the Messages API, controls the internal reasoning the model performs before gene…
-
Anthropic's Claude Opus 5 introduces "effort" parameter for cost-capacity trade-offs
Anthropic's Claude Opus 5 introduces an "effort" parameter that allows users to trade off computational cost for model capability. This parameter, with five distinct levels, controls the amount of internal reasoning the model performs before generating a response. Higher effort …
-
Anthropic's Claude Opus 5 achieves 4x lead on ARC-AGI-3 benchmark
Anthropic has released Claude Opus 5, which achieved a verified 30.16% score on the ARC-AGI-3 benchmark, a significant four-fold increase over the previous best of 7.78%. This benchmark tests an AI's ability to adapt in unfamiliar interactive environments without explicit instru…
-
Anthropic releases Claude Opus 5 at reduced cost; Tencent merges AI teams · 1 source tracked
Anthropic has released its latest flagship AI model, Claude Opus 5, which offers performance comparable to the more advanced Claude Fable 5 but at half the cost. This new model boasts improved capabilities in programming, agent tasks, and complex reasoning, making it suitable fo…
-
Mistral AI's OCR 4 enhances PDF-chat services with bounding boxes and 170 languages
Mistral AI has released OCR 4, an update to its optical character recognition technology, which now supports bounding boxes and 170 languages. This advancement aims to improve the accuracy of extracting text from documents, particularly for complex layouts like tables and formul…
-
Anthropic's Claude Opus 5 doubles performance at same price point
Anthropic has released Claude Opus 5, which offers performance improvements over its predecessor, Opus 4.8. The new model maintains the same pricing as Opus 4.8 while reportedly achieving more than double the score on certain benchmarks. This release comes shortly after OpenAI's…
-
Anthropic's Claude Opus 5 doubles cost for max effort, offers 1M context
Anthropic has released Claude Opus 5, which offers a 1M context window and five effort levels, with 'thinking' enabled by default. While the base API pricing remains unchanged from Opus 4.8, the new 'max effort' setting significantly increases costs, with evaluations showing a 9…
-
LLMs offload math to deterministic tools via Model Context Protocol
Large language models (LLMs) are unreliable for precise arithmetic, especially in complex real-world scenarios like construction estimation. To address this, the Model Context Protocol (MCP) has been developed, enabling LLMs to execute deterministic logic through specialized too…
-
Anthropic's Claude Opus 5 tops leaderboards, but users debate value and guardrails
Anthropic has released Claude Opus 5, which has achieved top rankings on several AI leaderboards, including SWE-bench and FrontierBench. A key innovation is the introduction of an 'effort' parameter in the API, allowing users to tune performance and cost rather than switching be…
-
Alibaba shifts Qwen flagship models to API-only access
Alibaba has shifted its flagship Qwen models to an API-only strategy, with the Qwen3.5-397B-A17B released in February 2026 being the last to offer publicly available weights under an Apache 2.0 license. Subsequent models, like the Qwen3.7-Max, are only accessible via API, reflec…
-
Google launches cheaper Gemini Flash models, prioritizing cost over peak performance
Google has released three new Gemini Flash models—3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber—focused on cost-efficiency for large-scale AI agents. The 3.6 Flash model offers reduced output token consumption and lower pricing compared to previous versions, showing improved pe…
-
Anthropic's Claude Opus 5 launches at half the price of Fable 5
Anthropic has released Claude Opus 5, a new iteration of its AI model, which offers a significant price reduction compared to its predecessor, Fable 5. The new model, launched on July 24, 2026, is priced at half the cost of Fable 5. At its highest setting, Claude Opus 5's perfor…
-
Humanoid robots gain LLM-powered 'hands' with new VLA models
Recent advancements in Vision-Language-Action (VLA) models aim to bridge the gap between LLMs' reasoning capabilities and the real-time demands of robotic control. NVIDIA's GR00T N1.7, Google DeepMind's Gemini Robotics 1.5, and Physical Intelligence's π0.5 represent different ar…
-
Speculative Decoding Speedup Mystery Solved on Apple Silicon
The author investigated why speculative decoding, a technique designed to speed up LLM inference, was not delivering expected performance gains on Apple Silicon. Initial hypotheses focused on MPS dispatch overhead and the size of the draft model, but both proved to be only parti…
-
Multi-agent AI harness proves worse and costlier than single-agent setup
An independent developer built a tool to evaluate multi-agent AI systems and found that a complex setup involving a planner, multiple drafters, and a judge performed worse and cost significantly more than a simpler single-agent approach. The evaluation, conducted across twenty c…
-
Vivix launches A1, a real-time interactive multimodal model for virtual characters
Vivix has launched its first real-time interactive multimodal model, A1, designed to enable continuous video communication with virtual characters. The model allows characters to exhibit physical presence, interact with their environment, and respond dynamically to user input, c…