One-Line Summary

Lower-cost automation arrives across the stack: Anthropic pushes agentic work into a cheaper default model, Google accelerates budget image generation, and Etched signals demand for inference hardware.

Big Tech

Anthropic releases Claude Sonnet 5 as default across plans

Claude Sonnet 5 is Anthropic’s new mid-tier AI model designed to plan steps, use tools like a browser and terminal, and complete tasks on its own. It is available today across all plans—default for Free and Pro, and available to Max, Team, and Enterprise—and also in Claude Code and on the Claude Platform, with introductory pricing of $2 per million input tokens and $10 per million output tokens through Aug 31, 2026, then $3 and $15; developers can call it as “claude-sonnet-5” via the API. ¹

Anthropic says Sonnet 5 narrows the gap with Opus 4.8 while pricing lower, and is a substantial improvement over Sonnet 4.6 in reasoning, tool use, coding, and knowledge work. Internal evaluations (e.g., BrowseComp, OSWorld-Verified) show a strict improvement over 4.6, and early testers report that it completes multi-step workflows end to end and self-checks output more reliably. ¹

Pre-deployment safety reviews find lower hallucination and sycophancy than 4.6, stronger refusals to malicious prompts, and substantially lower cybersecurity capability than Anthropic’s top models; in one assessment on Firefox vulnerabilities it never produced a full working exploit. Sonnet 5 ships with real-time cyber safeguards enabled by default and participates in Anthropic’s Cyber Verification Program; Anthropic recommends Opus 4.8 for cybersecurity work that needs reduced guardrails. ¹

To smooth adoption, Sonnet 5 uses an updated tokenizer that may map the same input to roughly 1.0–1.35× more tokens, with introductory pricing set so the switch is roughly cost-neutral; Anthropic also raised rate limits across Chat, Cowork, Claude Code, and the Claude Platform to support higher-effort use. ¹

Industry & Biz

Etched reports $5B valuation and $1B in AI chip orders

Etched, a startup building AI inference chips and bundled systems, reports $1 billion in contract orders and a total of $800 million raised, including an unannounced $500 million round in December at a $5 billion post-money valuation; TSMC manufactured its chip earlier this year. ²

The company is testing its first “frontier inference clusters” with customers—systems that pair the chips with custom racks and software—to run frontier models faster, cheaper, and with better power efficiency. Backers include VentureTech Alliance, Jane Street, Hudson River Trading, Two Sigma, Ribbit Capital, and angels such as Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Arthur Mensch, Scott Wu, plus Stanley Druckenmiller and Peter Thiel. ²

Investor appetite for inference hardware remains strong: Cerebras had 2026’s first breakout IPO, Groq raised $650 million, hyperscalers are designing in-house chips, and OpenAI unveiled its first custom chip built by Broadcom—context that underscores why Etched’s pitch on inference cost and throughput is drawing attention. ²

New Tools

Google launches Nano Banana 2 Lite for faster, cheaper images

Nano Banana 2 Lite is Google’s latest in-house image/video generator variant tuned for low latency and high volume; TechCrunch reports it can produce images in four seconds and is priced at $0.034 per 1,000 images. It’s available via Google AI Studio, the Gemini API, and the Gemini Enterprise Agent Platform, and replaces “Nano Banana” as the legacy model. ³

Google also announces a wider release of Gemini Omni Flash at $0.10 per second of video output and demos Omni Product Studio, which turns static images into “cinematic e-commerce videos.” Despite pushback over “AI slop” and creative-industry concerns—including a $75 million A24 partnership that drew criticism—companies continue investing in generative media for marketing and ads. ³

Community Pulse

Hacker News (768↑) — Commenters weigh capability against real-world token costs and urge teams to test on their own workloads. ⁴

"Specific task based benchmarks don't reflect a lot of day to day agentic use cases in my experience. If you are working on a series of discrete tasks and can clear context after each one and move to the next, you might get that sort of efficiency from Opus low effort. I often find that when working through a real problem, iterating and discovering, context length can creep up, and that is where opus tends to get expensive." — Hacker News ⁴

"Correct. Albeit the nuance here is that a more capable model might solve problems more efficiently and faster, possibly saving you tokens. As with any new model, you won't know the real impact until you start using it for your workload." — Hacker News ⁴

Hacker News (270↑) — Users share practical mockup wins alongside discomfort with staging and skepticism toward public benchmarks and commercial readiness. ⁵

"Using it for "staging" shitty rentals is pretty gross, but I used Nano Banana to make some mockups for a bath room remodel I'm doing and it worked pretty great." — Hacker News ⁵

"Public AI benchmarks are fucking stupid and can be completely ignored. Too gamed. However, your advice is right. KREA2 and Ideogram4. The latter being less commercially usable without what I would surely assume is a pain in the ass." — Hacker News ⁵

What This Means for You

If you rely on AI to not just draft but actually finish work, Sonnet 5’s stronger follow-through at a lower price invites you to run longer, multi-step workflows (e.g., research → draft → QA check) without jumping to a top-tier model. Because it’s the new default across plans and available via the Claude API, you can trial it quickly on a real task and compare outputs and time-to-complete against your current setup. ¹

Mind costs as you switch: Sonnet 5’s tokenizer can count 1.0–1.35× more tokens for the same text, but the introductory price is set to be roughly cost-neutral and Anthropic raised rate limits to support higher “effort” settings. Track cost per completed task (not just per token) and adjust effort levels to balance speed, accuracy, and spend. ¹

For marketers and designers, Nano Banana 2 Lite’s four-second generations and $0.034 per 1,000 images are built for rapid iteration on ad or mockup variants, while Omni Flash’s $0.10/second pricing makes short e-commerce clips testable without a big budget. Evaluate output quality and brand fit against your current tool before adding it to your production toolkit. ³

On budgets, Etched’s $1B in orders and $5B valuation are a signal that cutting inference cost is a top priority. When your cloud or platform introduces new instance types or alternative accelerators, request the performance and pricing details for your typical workload before making commitments. ²

Action Items

1. Try Claude Sonnet 5 on a real workflow: In Claude, run a 30–60 minute multi-step task end-to-end (research → draft → self-check) and note completion quality and time.
2. Track cost per completed task: Compare Sonnet 5 against your current model on the same task, logging tokens and minutes to compute cost-per-result.
3. Generate image variants with Nano Banana 2 Lite: In Google AI Studio, create 20–50 ad or product mockups and assess speed, quality, and brand fit.
4. Review Anthropic’s cyber safeguards: If your work touches security topics, confirm Sonnet 5’s default safeguards fit your policy; use Opus 4.8 when reduced guardrails are necessary.
5. Ask your vendor about inference options: Request information on upcoming accelerator choices and pricing for your typical prompt sizes and latency targets.