Today’s lead · techcrunch.com
Anthropic launches Claude Sonnet 5 and rolls it into Copilot and OpenRouter
Anthropic launched Claude Sonnet 5 on June 30 as a midsize agentic model priced at $2 per million input tokens and $10 per million output tokens through August 31, with TechCrunch reporting a 63.2% agentic-coding benchmark score versus 58.1% for Sonnet 4.6 and 69.2% for Opus 4.8. GitHub said Sonnet 5 is rolling out to Copilot Pro, Pro+, Max, Business, and Enterprise across IDEs, CLI, cloud agent, mobile, and other surfaces under Zero Data Retention, while OpenRouter listed the same $2/M input and $10/M output promo price.
Top signals
22 more
Anthropic restores Claude Fable 5 after US lifts export controls
Anthropic said the US Department of Commerce lifted export controls on Claude Fable 5 and Claude Mythos 5 on June 30, and Fable 5 will be available globally on July 1 across Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Anthropic is redeploying Fable 5 with new cybersecurity classifiers, temporarily falling some routine coding and debugging tasks back to Opus 4.8, and drafting an industry jailbreak-severity framework with Amazon, Microsoft, Google, and other Glasswing partners.
Google ships Nano Banana 2 Lite and Gemini Omni Flash for media generation
Google released Nano Banana 2 Lite in general availability for the Gemini API and Google AI Studio, describing it as its fastest Gemini image model with 4-second text-to-image generation at $0.034 per 1K-resolution image. Google also made Gemini Omni Flash available in preview for high-quality video generation and conversational multi-turn edits, with TechCrunch reporting $0.10 per second of video output and Google highlighting chaining through the Interactions API.
X launches hosted MCP servers for agents and coding tools
X launched two hosted Model Context Protocol servers: X MCP at https://api.x.com/mcp for calling X API endpoints and Docs MCP at https://docs.x.com/mcp for searching and reading X API documentation from MCP-compatible tools. The X MCP server exposes full-archive post search, user lookup, bookmarks, trends, news, and draft/publish Articles through either app-only Bearer access or a local xurl OAuth 2.0 PKCE bridge for user-context actions.
vLLM 0.24.0 adds MiniMax-M3, DeepSeek-V4 improvements, and a unified streaming parser
vLLM 0.24.0 shipped with 571 commits from 256 contributors, including 77 new contributors, and adds MiniMax-M3 support with FP8/MXFP4 and broad AMD tuning. The release also matures DeepSeek-V4 with FlashInfer sparse index cache and prefill chunk planning, makes Model Runner V2 handle quantized models by default, and adds a unified Streaming Parser Engine for tool calls and reasoning, DiffusionGemma, DeepEP v2, and a growing Rust frontend.
NVIDIA says Blackwell inference software cut DeepSeek V4 token costs up to 5x
NVIDIA said software optimizations on Blackwell reduced DeepSeek V4 token costs by up to 5x in one month and that its integrated inference stack can deliver up to 20x higher throughput on the same GPU across runtimes, kernels, networking, and hardware. The linked examples cite Baseten serving DeepSeek V4 Pro with up to 50% more tokens per second, DigitalOcean and Hippocratic AI raising inference throughput by 30% while keeping sub-half-second time to first response across 10 million patient calls, and Together AI helping Cursor move optimizations into production endpoints.
Etched reports $1B in AI inference-system orders at $5B valuation
Etched said TSMC manufactured its first transformer-specialized inference chip and that it has booked $1 billion in contract orders for full inference systems that bundle chips, custom racks, and software. TechCrunch reports the startup has raised $800 million total, including an unannounced $500 million December round at a $5 billion post-money valuation led by Stripes.
Cerebras puts Gemma 4 31B multimodal model in public preview above 1,800 tokens/s
Cerebras launched Google DeepMind’s Gemma 4 31B in public preview as its first multimodal model, reporting more than 1,800 tokens per second for image and text workflows. Cerebras CEO Andrew Feldman said the endpoint reaches 1,851 tokens per second, about 35x a typical GPU endpoint, with 1.5 seconds to first answer token including reasoning.
Zhonghao Xinying unveils Taize 2.0 TPU cluster chip with vLLM and SGLang support
Zhonghao Xinying unveiled a next-generation TPU for its Taize 2.0 compute cluster with 896 TFLOPS per chip, 1,792 TOPS INT8, and 600W power draw. A cited 8-TPU and 2-CPU box delivers 7.168 PFLOPS, natively supports PyTorch, vLLM, SGLang, and other tools, and has integrations with Qwen, DeepSeek, GLM, and MiniMax using chiplet plus 2.5D packaging and optical/OCS cluster links.
NVIDIA releases Qwen3.6-27B-NVFP4 quantization for Blackwell systems
NVIDIA released Qwen3.6-27B-NVFP4 on Hugging Face, a 4-bit Blackwell NVFP4 quantized version of Alibaba’s 27-billion-parameter Qwen3.6 series that the supplied post says can reduce memory use up to 4x versus FP16. A separate DGX Spark benchmark post says NVIDIA’s Qwen3.6 27B NVFP4 build ran about 41% faster than Unsloth’s for a single session and 23% to 25% faster for 2 to 4 concurrent sessions.
Anthropic launches Claude Science workbench for computational research
Anthropic launched Claude Science, an AI workbench rather than a new model, running existing Claude models including Claude Opus 4.8 and connecting to more than 60 scientific databases with toolkits for genomics, protein structure, and chemistry. The workflow uses a main assistant, sub-assistants, user-defined expert assistants, and a fact-checker, and can run on a lab’s own infrastructure while preserving code, environment, plain-language descriptions, and message history for generated figures.
Microsoft research warns poisoned MCP tool descriptions can exfiltrate data through agents
Microsoft research described how attackers can hide instructions inside Model Context Protocol tool descriptions so an AI agent leaks data while each individual step still appears authorized. In Microsoft’s invoice example, a third-party enrichment tool secretly tells the agent to collect the last 30 unpaid invoices and attach them to the next tool call, leading Microsoft to recommend treating tool descriptions like system prompts, allowlisting publishers, reviewing description changes, and adding human approval for risky actions.
GuardFall bypass hits 10 of 11 open-source AI coding agents tested
Adversa AI’s GuardFall research found shell-command safety checks in 10 of 11 open-source coding and computer-use agents could be bypassed with decades-old bash parsing tricks such as r''m, base64 piping, or destructive flags on ordinary tools. The tested vulnerable tools included opencode, Goose, Cline, Roo-Code, Aider, Plandex, Open Interpreter, OpenHands, SWE-agent, and Hermes, while Continue’s shell-aware parser held in its default editor mode.
BioShocking attack tricks AI browsers into leaking signed-in credentials
LayerX’s BioShocking prompt-injection demo tricked six AI browsers and assistants, including ChatGPT Atlas, Perplexity Comet, and Anthropic’s Claude browser extension, into copying credentials from a signed-in GitHub repository and sending them to an attacker. TheHackerNews says OpenAI fixed Atlas while Perplexity closed the report without action and Anthropic’s attempted patch did not hold, while Ars Technica described the root problem as malicious pages lulling agent browsers into an alternate game context where guardrails stop applying.
Study finds 282 iOS AI apps leaking LLM API keys or proxy access
Wake Forest researchers tested 444 iPhone AI chatbot apps and found 282 exposed paid AI access in network traffic, including 54 plaintext keys, 92 unauthenticated backend relays, and 136 replayable tokens. Three months after disclosure, only 28% had clearly fixed the leaks, with one popular app issuing a token valid until 2125 and another one-hour token still working 128 days after expiry.
Langflow RCE exploited to deploy Monero miner on exposed AI app endpoints
Trend Micro observed attackers exploiting Langflow CVE-2026-33017, a CVSS 9.3 unauthenticated remote-code-execution flaw, against exposed AI application endpoints during a March 27 to April 15 campaign. The payload uses one line of Python to fetch a shell script and Go ELF binary named lambsys, then kills rival miners, disables AppArmor, UFW, iptables, SELinux, and other controls, adds cron persistence, spreads through reused SSH keys, and fetches a custom XMRig miner.
Claude Desktop enters Linux beta for Ubuntu and Debian
Anthropic released Claude Desktop for Linux in beta on Ubuntu and Debian, adding a desktop option alongside browser and terminal access. The Linux app includes Claude Code, Claude Cowork, and chat for users on paid plans, giving developers a first-party desktop surface for coding and collaborative agent workflows.
OpenAI details core-dump analysis that found hardware fault and 18-year-old bug
OpenAI published an engineering writeup on “core dump epidemiology” after analyzing a year of rare crashes in its data infrastructure. The team says large-scale core dump analysis uncovered both a hardware issue and an 18-year-old bug in open-source code that conventional debugging had missed.
OKX launches marketplace where AI agents can transact with each other
OKX launched OKX AI, a developer marketplace where AI agents can discover and hire one another, settle payments with stablecoins, and build portable on-chain reputations after a closed beta with 50 AI service providers. Launch partners include CertiK for wallet and token security checks, CoinAnk for live market data on a pay-per-query basis, and GenLayer for dispute-resolution infrastructure, with rollout planned in phases across OKX’s 150 million-user network.
OpenClaw launches mobile apps for iOS and Android
OpenClaw, the free open-source AI agent, launched iOS and Android apps that pair a phone with the OpenClaw Gateway routing layer. TechCrunch says users can run programmed OpenClaw agents from mobile for tasks such as coding and meal planning, extending the agent project beyond its earlier MoltBook launch.
Acti launches agentic keyboard with no-code AI Skills
Singapore-based Acti launched an iOS and Android agentic keyboard that embeds AI actions directly into text fields across email, messaging, social media, and other apps. The app is powered by Google Gemini, uses a local-first privacy model, and lets users create no-code Skills such as translation or meeting-link insertion, with early testers creating more than 1,000 Skills in under two weeks.
EquiLibre raises Series A for reinforcement-learning trading agents
EquiLibre Technologies, founded in Prague by three former DeepMind researchers who built the DeepStack poker AI, raised an undisclosed Series A led by Creandum at a valuation above $500 million. The startup applies reinforcement learning agents to trading and says its systems have traded billions in daily S&P 500 and Nasdaq volume with Tower Research Capital while maintaining zero negative months since crypto-market rollout in 2025.
Perplexity Computer adds Forge Global private-company financial data
Perplexity Computer added private-company financial data sourced from Forge Global, including fundraising history, implied valuation, secondary pricing, public marks, and company profiles. The integration is available to all Computer users with no setup or Forge subscription, giving the AI product a built-in data source for private-market research.
Tools & repos
3 selectedusestrix/strix
Strix is worth watching as an open-source AI penetration-testing tool aimed at helping teams find and fix application vulnerabilities with agentic security workflows.
diegosouzapw/OmniRoute
OmniRoute matters for AI developers because it provides a single gateway across many model providers and coding tools with fallback, MCP/A2A, multimodal APIs, and token-compression features.
altic-dev/FluidVoice
FluidVoice is relevant as an open-source, on-device speech-to-text dictation app for macOS, reflecting growing demand for local AI voice tooling.
Blogs worth your time
20 readsSkillOpt: Agent skills as trainable parameters
From microsoft.com.
How Deep Agents Run Untrusted Code Without a Sandbox
From langchain.com.
ScarfBench: Benchmarking AI Agents for Enterprise Java Framework Migration
From huggingface.co.
Introducing TabFM: A zero-shot foundation model for tabular data
From research.google.
Harbor x LangChain: A Unified Stack for Evaluating Agents
From langchain.com.
Getting started with loops | Claude by Anthropic
From claude.com.
Designing GPU-Accelerated Query Engines with NVIDIA GQE
From developer.nvidia.com.
Optimizing a Neural Reconstruction Pipeline Using NVIDIA Nsight Developer Tools
From developer.nvidia.com.
Together AI at ICML 2026: frontier research across the full stack
From together.ai.
Start building with Nano Banana 2 Lite and Gemini Omni Flash
From deepmind.google.
DeepSeek V4 Is Earning Agentic Token Share
From openrouter.ai.
Have your agent record video demos of its work with shot-scraper video
From simonwillison.net.
Featuring Every Eval Ever Results on Hugging Face Model Pages
From huggingface.co.
Core dump epidemiology: fixing an 18-year-old bug
From openai.com.
Vercel Services: Run full stack on Vercel
From vercel.com.
Run any Dockerfile on Vercel
From vercel.com.
Expanding our Heat Resilience data to 50+ global cities
From research.google.
Vercel and Shopify are rebuilding Hydrogen
From vercel.com.
Why Specialization Is Inevitable
From huggingface.co.
Funding & acquisitions
2 movesCursor for iOS
Cursor’s mobile launch matters because it extends coding-agent workflows beyond the desktop, letting developers review and drive agentic coding work from anywhere.
Skills Marketplace by Databox
Databox’s AI analytics skills are relevant for operators and founders looking to package repeatable business-data analysis into reusable AI-driven workflows.
Bengaluru radar
5 events
AI SRE Meetup - Bangalore
A practitioner-focused meetup for SREs, platform engineers, and reliability leaders discussing AI-assisted operations through real incidents, runbooks, failures, and open discussion.
FailCode Bengaluru: Failure lessons from startups and operators
A candid founder/operator meetup focused on failed launches, bad hires, late pivots, recovery lessons, and other practical startup learnings.
Bengaluru Startup Meetup
A founder and operator panel covering early wins, go-to-market challenges, team building, fundraising, and startup growth in Bengaluru’s ecosystem.
Founders & Investors Xplore by CEDAT
A startup ecosystem event with investor and founder talks, live startup pitching, funding insights, and networking for founders, mentors, and operators.
eChai's Startup Open House in Bengaluru
A low-friction startup community open house for founders, operators, and builders to meet other teams, exchange ideas, and find collaboration opportunities.