Anthropic Launches Claude Sonnet 5 — TechCrunch / Anthropic Anthropic released Claude Sonnet 5 today, naming it the new default model for Claude Free and Pro plans. The model features a native 1M-token context window and is priced at $2/$10 per million tokens (in/out) through August 31 — positioning it as a cheaper alternative to Opus 4.8, GPT-5.5, and Gemini Pro for agentic workloads. Claude Code 2.1.197 ships Sonnet 5 as its default model simultaneously.

🤖 Frontier Models & Agents

Anthropic launches Claude Sonnet 5 as a cheaper way to run agents — TechCrunch Sonnet 5 delivers near-Opus-4.8-level performance on reasoning, tool use, coding, and browser-based tasks at a fraction of the price ($2/$10 per Mtok through August 31, rising to $3/$10 after). The 1M-token context window makes it particularly suited to long-horizon agentic workflows. Claude Code 2.1.197 ships with Sonnet 5 as its default model and adds org-level model defaults for admins.

Devin Fusion — Cognition Cognition’s new multi-model harness dynamically routes tasks between frontier and cost-effective models using a dual-agent architecture (main agent + sidekick), cutting costs 35% on the FrontierCode benchmark while maintaining top-tier performance. Integration with Fable 5 pushes savings to 41%.

Sakana Fugu Launches With 93.2 LiveCodeBench Score After Claude Ban — Implicator Sakana’s Fugu Ultra beats Fable on LiveCodeBench and opens at $5/million input tokens. The launch follows Sakana’s ban from Claude API access, which it worked around by training on alternative infrastructure.

Gemini’s personalized AI image generation is now free for US users — TechCrunch — Google opens its Nano Banana-powered personalized image generation (which tailors outputs to individual user preferences without explicit prompting) to all eligible free US users.

Google introduces a faster, cheaper image generator with Nano Banana 2 Lite — TechCrunch — An updated, faster, and more cost-efficient variant of Google’s image generation model aimed at high-volume creator use cases.

Vibe-coding platform Base44 launches own model — TechCrunch

🔬 Science & Research

Anthropic’s Claude Science bets on workflow, not a new model, to win over scientists — TechCrunch Claude Science, now in beta on macOS and Linux, unifies databases, code environments, and compute into a single AI-driven research workbench across genomics, proteomics, structural biology, and cheminformatics. A separate reviewer agent checks citations and math; every figure ships with the exact code and environment that produced it. Available to Pro/Max/Team/Enterprise users, with discounted academic seats and up to $30,000 in credits for 50 funded research projects (applications open through July 15).

From Brain Waves to Words: Brain2Qwerty Offers a New Path to Communication Without Surgery — Meta AI Meta’s Brain2Qwerty v2 achieves 61% word-level accuracy decoding sentences in real time from magnetoencephalography (MEG) — a non-invasive device worn like a helmet — up from 8% for prior non-invasive approaches. Trained on ~22,000 sentences from 9 volunteers, the full training code for v1 and v2 and the v1 dataset are now public.

DeepSeek open sources DSpark, a new framework to speed up LLM inference by up to 85% — VentureBeat DSpark acts as a speculative scout that runs steps ahead of the main model and lets the larger model quickly verify which predicted steps are safe, dramatically accelerating throughput without changing model outputs. Released as open source.

DiScoFormer: One transformer for density and score, across distributions — HuggingFace / AllenAI — A single transformer model that estimates both density and score functions in one pass without retraining, reducing score error by 6.5× and density error by 37× over kernel density estimation at 100 dimensions.

RoadmapBench: Evaluating Long-Horizon Agentic Software Development Across Version Upgrades — arXiv — A new benchmark covering 115 multi-file coding tasks grounded in real version upgrades across 17 repositories, with a median change of 3,700 lines across 51 files.

🛠️ Developer Tools

Build from anywhere with Cursor for iOS — Cursor Cursor for iOS is now in public beta. Developers can launch or manage cloud or local agents from the mobile app, receive progress updates via iOS Live Activities, and merge PRs on the go — targeting incident response and async engineering workflows.

X now offers an MCP server to make its platform easier for AI tools to use — TechCrunch — X launched a hosted MCP server, lowering the integration barrier for AI applications connecting to its API.

Getting started with loops — Anthropic — New Claude documentation and blog post on building persistent agentic loops.

Build a document processing workflow in 30 minutes — Mistral — Mistral Workflows launches as a durable, fault-tolerant orchestration platform for multi-agent AI pipelines.

💰 Enterprise & Business

Nvidia competitor Etched hits $5B valuation, $1B in sales for AI chip — TechCrunch Etched emerged from stealth today having raised $800M across multiple undisclosed rounds (most recently $500M at $5B post-money in December) and booked over $1B in customer contracts for its transformer-specialized inference clusters. Backers include Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Jane Street, and Two Sigma. The chips and accompanying rack systems target frontier inference at better power efficiency than GPU-based alternatives.

Amazon seeks cheaper AI alternatives as Anthropic shifts to token-based pricing — The Next Web A renegotiated contract moves Anthropic billing to per-token pricing, which could substantially raise Amazon’s costs. Amazon is reportedly exploring cheaper alternatives — even as Salesforce (which spends ~$300M/year on Anthropic tokens and holds ~1% equity) simultaneously promoted Claude Tag inside Slack alongside its own Agentforce platform, underscoring the tangled strategic relationships in the AI supply chain.

Amazon launches new $1 billion FDE org, following OpenAI and Anthropic — TechCrunch — Amazon’s new Field Deployment Engineers organization will embed engineers within enterprises to deploy purpose-built agents, focusing on fast deployment and customer self-sufficiency.

Google Cloud will sell specialist AI models built for science — The Next Web — Google Cloud adds SandboxAQ’s large quantitative models — trained on scientific equations and lab data — to its marketplace for drug discovery, materials science, and semiconductor manufacturing use cases.

Rocket Lab Buys Satellite Operator Iridium in Bid to Challenge SpaceX — WSJ — Rocket Lab acquires Iridium’s 66-satellite LEO constellation and spectrum rights, aiming to develop an upgraded constellation to compete with Starlink.

In San Francisco, Even $180,000 Tech Salaries Are No Longer Enough — NYT — A new AI wealth class is outcompeting traditional tech workers on housing; OpenAI, Anthropic, and SpaceX IPOs combined could create 20+ new billionaires.

🎬 AI & Media

Netflix is using an AI-generated Gene Wilder voice in its Willy Wonka reality show — The Verge A teaser for Wonka’s The Golden Ticket (premiering September 23) confirms the voiceover is AI-generated from Gene Wilder’s voice — raising immediate questions about estate consent, AI likeness rights, and the precedent for deceased performers in entertainment. The sets are practical, but the iconic narration voice is synthetic.

Tidal won’t pay royalties on AI-generated music, but isn’t banning it outright — The Verge — Starting July 15, Tidal will label 100% AI-generated tracks and demonetize them immediately; human-AI collaborations are treated differently.

Google’s NotebookLM can sum up your research in a TikTok-style clip — The Verge — NotebookLM gains 60-second AI-generated vertical video summaries of uploaded research, rolling out to Ultra and Pro subscribers.

Libby will filter out AI content, kind of — The Verge — OverDrive’s new CEO signals the Libby ebook app will surface AI-content filtering tools as libraries grapple with AI-generated titles flooding catalogs.

Generated by claude-sonnet-4-6 on 2026-06-30T10:00:00Z