Anthropic launches Claude Science and a new model

In partnership with

The best prompt engineers aren't typing. They're talking.

Power users figured this out early: speaking a prompt gives you 10x more context in half the time. You include the edge cases, the examples, the tone you want — because talking is fast enough that you don't skip them.

Wispr Flow captures everything you say and turns it into clean, structured text for any AI tool. Speak messy. Get polished input. Paste into ChatGPT, Claude, Cursor, or wherever you work.

89% of messages sent with zero edits. 4x faster than typing. Works system-wide on Mac, Windows, and iPhone.

Start flowing free

In today's issue

01 · Business

Anthropic launches Claude Science workbench for computational researchers

Anthropic launched Claude Science, a workbench connecting 60+ scientific databases and pre-built research tools for genomics, protein structure, and chemistry. The product runs on Claude Opus 4.8 (no special model) and comes with $30,000 in free credits for 50 academic projects—a bet that workflow beats raw capability for winning over researchers.

The details:

No new model or gating: Claude Science uses the same Claude Opus 4.8 available to all subscribers, with no biology fine-tune or specialized weights.
Multi-agent architecture lets researchers spawn parallel sub-assistants for sequence analysis, structural prediction, and fact-checking, while preserving full reproducibility (code, environment, message history).
Early wins include UCSF Brain Tumor Center compressing glioma germline analysis workflows and Allen Institute's Jérôme Lecoq building computational review pipelines; Novo Nordisk is a launch customer.

Why it matters: Anthropic is positioning Claude Science as the research equivalent of Claude Code—not a moonshot model, but a workflow layer that makes existing AI dramatically more useful for scientists. With OpenAI's GPT-Rosalind gated to enterprise and Google bundling AlphaFold into Gemini, the real competition is over who owns the scientist's daily operating environment, not model leaderboard points. Anthropic's move to keep it accessible and non-gated (plus bankroll 50 projects) is a smarter long-game play for embedding in labs.

Read the full story →

02 · Business

Base44 launches its own AI model to defend $100M ARR vibe-coding business

Base44, the Wix-owned vibe-coding platform that crossed $100M ARR, just launched Base1, its own LLM trained on tens of millions of real user interactions. The move signals that applied-AI startups now see owning their models as essential to compete against frontier labs that are building their own app-generation feedback loops.

The details:

Base1 is designed to replace Anthropic's Opus for app-generation workloads, giving Base44 direct control over latency, cost, and inference spend without relying on third-party APIs.
Lovable, a Swedish rival that still uses external LLMs, reached $500M ARR this month—five times Base44's current run rate—showing the stakes of the race.
Wix acquired Base44 for $80 million when it was just six months old with eight employees; the parent company is now cutting 20% of its workforce, making the unit's margin improvements critical.

Why it matters: Base44's move is less about AI research and more about defensibility through data and cost. As frontier models get commoditized and inference becomes a real margin drag, owning your own tuned stack stops being a nice-to-have and starts being table stakes. The risk is real—Lovable's five-to-one revenue advantage shows that a better model can't substitute for better product and distribution—but the economics of enterprise deployment increasingly favor startups that control their own inference.

Read the full story →

03 · Tools

X launches hosted MCP server, opening its API to Claude, Cursor, and Grok Build

X launched a hosted Model Context Protocol server on Monday, letting AI tools like Claude, Cursor, and Grok Build tap into X's API using a user's own account permissions. The move eliminates days of engineering work to wire AI agents into the network and positions X as a real-time data source for AI systems to query.

The details:

Compatible with any MCP-compliant app at launch; X joins GitHub, Slack, Stripe, and Salesforce in operating an official MCP endpoint.
API pricing remains $0.015 per published post and $0.20 per post with links—set earlier this year to curb spam at scale.
The hosted server exposes existing capabilities (search, post retrieval, user lookup, trend analysis) but cuts integration time from days to minutes.

Why it matters: X is betting that making its data frictionless for AI agents will cement its role as a real-time information layer for autonomous systems—the same play Reddit and Stack Overflow made with training data, except X is monetizing tool access instead. As MCP becomes the standard integration surface across enterprise platforms, the lock-in shifts from individual APIs to the agent runtime itself, which favors networks with fresh, queryable data and strong abuse controls.

Read the full story →

AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

● ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat

● Generate images & video with Sora, Veo, Ideogram

● Compare any two models side by side

● From $8.99/mo · 80+ models, all included

Try AI Box →

aibox.ai

Trusted by 3,000+ teams

04 · Models

Anthropic ships Claude Sonnet 5 at $2 per million input tokens

Anthropic launched Claude Sonnet 5 at $2 per million input tokens, undercutting Opus 4.8 while scoring 63.2% on agentic coding—within 6 points of its flagship. The model becomes the default for all free and Pro users starting Tuesday, signaling that agent-capable AI is now table stakes across every tier, not a premium feature.

The details:

Promotional pricing ($2 input, $10 output) runs through August 31, then rises to $3 input—still cheaper than Opus 4.8, GPT-5.5, and Gemini 3.1 Pro.
On knowledge work benchmarks, Sonnet 5 slightly outperforms Opus 4.8, suggesting cost-per-task efficiency now matters more than raw reasoning rank.
Early testers report fewer mid-task stalls in multi-step workflows; Zapier's two-part Salesforce automation completed where previous models would abandon halfway through.

Why it matters: Agentic capability has commoditized faster than anyone expected. Three labs (Anthropic, OpenAI, Google) all launched cheaper mid-tier agents in the past month, and the message is identical: automation without handholding is no longer a flagship differentiator. The real competition shifts to reliability, refusal quality, and cost-per-completed-task. That's a win for builders shipping at scale.

Read the full story →

05 · Models

Google ships Nano Banana 2 Lite and Gemini Omni Flash to developers

Google DeepMind shipped Nano Banana 2 Lite, a text-to-image model that generates in 4 seconds for $0.034 per 1K images, and Gemini Omni Flash, a video generation model priced at $0.10 per second of output. Both target high-volume developer workflows where speed and cost matter more than maximum quality.

The details:

Nano Banana 2 Lite replaces the original Nano Banana as the recommended tier, maintaining prompt adherence and character consistency despite aggressive speed-cost optimization.
Omni Flash matches Veo 3.1 Fast pricing while adding conversational editing, multimodal input mixing (text, image, video), and text-to-action synchronization for on-screen graphics.
Nano Banana 2 Lite ships across nine products simultaneously: AI Studio, Gemini API, AI Mode in Search, Gemini app, NotebookLM, Google Photos, Stitch, Google Flow, and Google Ads.

Why it matters: Google is stacking a speed-optimized pipeline for consumer and enterprise builders—draft fast with Nano Banana 2 Lite, refine with Omni Flash, all at commodity pricing. The model family now spans four quality tiers, letting developers pick the right tradeoff instead of overpaying for flagship capability. Expect this pricing discipline to pressure competitors on the cost-per-token and cost-per-frame fronts.

Read the full story →