A KV cache, or key-value cache, speeds up autoregressive transformer inference by storing attention information calculated for tokens that have already been processed. Without the cache, the model would recompute those keys and values each time it generates another token. Reusing them reduces repeated work and improves response latency, especially for long prompts and outputs. The tradeoff is memory: cache size grows with sequence length, batch size, layer count, and model dimensions. Production systems manage this cost with paging, quantization, eviction, prompt caching, and request scheduling. KV caching changes inference efficiency, not the model’s learned knowledge, and cached data must be protected when prompts contain sensitive information.
SpaceXAI’s Grok 4.5 tied GPT-5.6 Sol for first on the SWE-Atlas-QnA coding benchmark, a narrow win that highlights the model’s real edge: efficiency and low cost.
A South Korean developer says Anthropic’s billing system tried to charge him $16.6 million despite being on the free plan with zero API usage, one of several billing complaints.
Anthropic extended free Claude Fable 5 access to July 19, its second extension in a week, as OpenAI removed usage caps and made GPT-5.6 Sol cheaper to run.
Apple sued OpenAI for trade secret theft, alleging former employees now at OpenAI stole confidential hardware information to build its first AI device. OpenAI denies it.
SK Hynix raised $26.5 billion in its Nasdaq debut, the largest US listing ever by a foreign company, as investors rushed to buy into the AI memory chip boom.
OpenAI is discontinuing ChatGPT Atlas, its standalone AI browser, on August 9, folding its agentic browsing into the new ChatGPT desktop app just nine months after launch.
Anthropic launched Reflect, a beta dashboard that shows users how they use Claude and nudges them toward healthier habits, a rare feature that can encourage using the product less.
OpenAI launched ChatGPT Work, a GPT-5.6 agent that pulls context from a user’s apps to produce finished docs, slides and sites, as it merges Codex into a unified desktop app.
SpaceXAI released Grok 4.5, a coding and agentic model trained jointly with Cursor that undercuts rivals on price and token use, though it trails Fable 5 on benchmarks.
Meta plans to begin manufacturing its in-house Iris AI chip in September as it aims to double computing capacity to 14 gigawatts next year and cut its reliance on Nvidia.