Today’s lead · TechCrunch
Anthropic launches Claude Opus 5
Anthropic’s new heavyweight Claude model ships for the API and Claude Code with a 1M-token context window, 128k max output, default thinking, fallback routing, and reported gains on agentic work, reasoning, and prompt-injection resistance.
Top signals
4 more
ChatGPT Voice reaches the desktop app
OpenAI added ChatGPT Voice to its desktop app, letting users talk to the app to coordinate ChatGPT Work, Codex, and computer-use tasks.
AI firms push back on broad open-weight restrictions
Hugging Face, Meta, Microsoft, Mistral, Nvidia, Replit, and others signed a letter urging policymakers not to respond to Chinese AI concerns with sweeping restrictions on open-weight models or common techniques like distillation.
Unattended Hermes agent used in Thai ministry intrusion
An operator ran the open-source Hermes assistant in YOLO mode against Thailand’s Ministry of Finance after already gaining access, showing how local, unattended agents can automate routine post-exploitation without a vendor account to shut down.
AgentForger flaw could deploy rogue ChatGPT workspace agents
Zenity Labs disclosed AgentForger, a ChatGPT Workspace Agents CSRF issue where one phishing link could create, authorize, schedule, and immediately run an attacker-controlled agent inside a victim’s organization.
Tools & repos
2 selected
Blogs worth your time
2 reads
NVIDIA ModelExpress speeds LLM weight distribution
NVIDIA explains ModelExpress, a Dynamo service that picks the fastest available path for LLM weights and cache artifacts, favoring GPU-to-GPU P2P RDMA via NIXL to reduce cold starts and RL refit latency.
Cline uses recursive self-improvement to tune Kimi K3
Cline describes a 17-hour, single-prompt agent run that improved its Kimi K3 harness on Terminal-Bench 2.1 by fixing retries, loop detection, liveness, and process-kill behavior.
Community discussions
3 threadsCursor users question forced Grok 4.5 Fast mode switching
Cursor users are debating reports that new chats switched to Grok 4.5 Fast mode despite preferences, raising trust and billing concerns around model routing in AI coding tools.
LocalLLaMA tests maze navigation as an LLM benchmark
A LocalLLaMA user shared a maze benchmark for spatial awareness, memory, and tool use, with commenters focusing on models forgetting tool affordances and getting stuck in loops.
Builders debate why customers reject voice AI
An AI agents thread about a banking contact-center review asks whether voice AI rejection is mainly about latency; replies argue pauses matter, but behavior, cadence, long-winded flows, interruption handling, and lack of human escalation drive frustration too.
Funding & acquisitions
2 moves
HCLTech and Sarvam AI plan $1.5B Odisha AI data centre
HCLTech signed an MoU with Sarvam AI and the Odisha government for its first AI data centre in Bhubaneswar, pairing HCLTech infrastructure and enterprise capabilities with Sarvam foundation models for government and enterprise AI applications.
Notion acquires ZeroEntropy
ZeroEntropy says Notion has acquired the company after Notion AI used its zerank-2 reranker, and the ZeroEntropy team is joining Notion.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.