AI news — July 11, 2026
The 20 top stories PulseAugur surfaced that day, ranked by signal across labs, papers, and developer communities.
-
OpenAI's GPT-5.6 Sol Ultra solves 50-year-old math problem
OpenAI's GPT-5.6 Sol Ultra has reportedly solved the Cycle Double Cover Conjecture, a mathematical problem that has remained unsolved for 50 years. The model achieved this feat in under an hour by utilizing 64 parallel subagents. However, the proof's elementary nature and lack o…
-
Zhipu AI CEO outlines AGI focus post-trillion valuation
Zhipu AI founder Tang Jie announced in an internal letter that the company will focus on advancing AGI capabilities rather than short-term commercialization. Following a period of significant growth, including joining the "trillion Hong Kong dollar club," Zhipu AI is prioritizin…
-
Google Research releases TimesFM 2.5 for zero-shot time-series forecasting
Google Research has released TimesFM 2.5, an open-source foundation model for time-series forecasting. This model, with 200 million parameters and a context window of up to 16,384 points, can predict future trends without requiring task-specific training data, operating in a zer…
-
OpenAI's GPT-5.6 pricing under scrutiny as cost-per-task emerges as key metric · 1 source tracked
OpenAI has released its new GPT-5.6 model family, featuring Sol, Terra, and Luna tiers with significantly lower per-token pricing. However, the actual cost per task may be higher due to increased token usage, a phenomenon highlighted by previous analyses of GPT-5.5 and Claude Op…
-
Meta's Muse Spark 1.1 surpasses GLM-5.2 in coding, cuts costs and hallucinations
Meta's Muse Spark 1.1 has shown significant improvements, scoring 51 on the Artificial Intelligence Index and demonstrating a notable leap in coding capabilities with a score of 71.3. This new version outperforms GLM-5.2 in coding tasks and operates at a reduced cost of $0.26 pe…
-
Chinese GLM-5.2 model outperforms GPT-5.5 on coding benchmarks, offers lower cost · 1 source tracked
The Chinese AI model GLM-5.2, developed by Z.ai (formerly Zhipu AI) and released on June 13, 2026, has demonstrated superior performance over OpenAI's GPT-5.5 in specific coding benchmarks, achieving a score of 62.1 on SWE-bench Pro compared to GPT-5.5's 58.6. This open-weight m…
-
OpenAI's GPT-5.6 pricing revealed: Sol, Terra, Luna models vary fivefold
OpenAI has released GPT-5.6 with three distinct models: Sol, Terra, and Luna, offering varying price points for API access. Sol, the flagship model, is priced at $5 per million input tokens and $30 per million output tokens, the same as GPT-5.5. Terra is half the price of Sol, a…
-
Grok 4.5 release with 500K context window challenges OpenAI's GPT-5.6
Grok 4.5 has been released with a 500,000-token context window, significantly altering the competitive landscape for large language models. This advancement in context window size, coupled with its pricing structure, is expected to influence the development and strategic directi…
-
OpenAI's GPT-5.6 Sol code model faces scrutiny over benchmark claims
OpenAI has released GPT-5.6 Sol, a new flagship AI model focused on coding and complex reasoning, on July 9, 2026. While OpenAI claims impressive benchmark scores of 88.8% and 91.9% on Terminal-Bench 2.1, these results are from internal testing and not yet reflected on independe…
-
Anthropic's Claude Sonnet 5 offers near-Opus quality at lower cost, but with caveats
Anthropic has released Claude Sonnet 5, positioned as a balance between speed and intelligence, offering near-Opus quality at a reduced price. While the sticker price for Sonnet 5 is 40% lower than Opus 4.8, the actual cost savings are less significant due to a new tokenizer tha…
-
SpaceXAI launches Grok 4.5, a 1.5T parameter coding-focused model
SpaceXAI, formerly xAI, has released Grok 4.5, a new MoE model with 1.5 trillion parameters and a 500,000 token context window. This model, trained on data from the code editor Cursor, is positioned as a competitor to models like Claude Opus and GPT-5.5, offering a lower price p…
-
OpenAI launches new model to challenge Anthropic's collaborative AI
OpenAI has launched a new model designed to compete with Anthropic's offerings, particularly in the realm of collaborative AI tools. This new model reportedly incorporates features aimed at improving coding capabilities and directly challenges Anthropic's existing workflow. The …
-
Homoglyph Evasion Exploits LLM Prompt Filters
A security vulnerability has been identified where LLM prompt filters can be bypassed using homoglyph evasion techniques. Attackers can substitute Cyrillic or other non-ASCII characters that visually resemble standard ASCII characters into prompts, rendering them undetectable by…
-
Text-to-SQL benchmarks flawed: Over half of answer keys incorrect
A recent audit of text-to-SQL benchmarks revealed significant inaccuracies in answer keys, with over 50% of annotations being incorrect in prominent datasets like BIRD Mini-Dev and Spider 2.0-Snow. This suggests that current performance metrics for AI agents in this domain may b…
-
OpenAI's GPT-5.6 Solves 50-Year Math Conjecture with 64 Sub-Agents
OpenAI's GPT-5.6 model has reportedly solved the 50-year-old Cycle Double Cover Conjecture in graph theory within an hour. The model utilized 64 sub-agents and a detailed 700-word prompt that focused on defining success criteria and boundaries rather than dictating specific prob…
-
Rust accelerators boost LangGraph agent performance up to 10x
A new set of Rust accelerators, named fast-langgraph, has been developed to significantly speed up critical operations within the LangGraph framework. These accelerators, designed as drop-in replacements, target bottlenecks such as state serialization, state management, and redu…
-
Perplexity: A Key Metric for Evaluating Language Models
Perplexity is a key metric for evaluating the performance of Language Models (LMs), especially Large Language Models (LLMs), by measuring how well they predict text. A lower perplexity score indicates a model's greater accuracy and reduced uncertainty. This metric is vital for c…
-
Model Context Protocol integrates AI agents with GitHub, databases, and web search
The Model Context Protocol (MCP) has released a suite of officially maintained servers designed to integrate AI models like Anthropic's Claude Code with various tools and data sources. These servers, available via npm, enable agents to interact with local files, manage GitHub re…
-
Understand HNSW: Why Vector Search Returns Garbage
This article delves into the Hierarchical Navigable Small World (HNSW) graph structure, a key component in modern vector databases that enables efficient Approximate Nearest Neighbor (ANN) search. The author highlights that many developers building retrieval-augmented generation…
-
Mem0, Letta, Zep: AI Agent Memory Frameworks Compared
Three open-source AI agent memory frameworks—Mem0, Letta, and Zep—offer distinct approaches to memory management. Mem0 provides a universal CRUD API for vector embeddings, suitable for simple integrations into existing agents. Letta functions as a stateful agent operating system…