Skip to content
Source · Daily Brief

AI Daily Brief — 26 August 2025

One of the year’s loudest single days for shipped products in agent and image-generation. Google launched Nano Banana = Gemini 2.5 Flash Image. Anthropic shipped Claude for Chrome + $1.5B Bartz settlement. OpenAI quietly noticed Assistants API sunset.

Top stories

  • Google launches Gemini 2.5 Flash Image (“Nano Banana”). Officially unveiled as state-of-the-art image generation and editing model. Multi-image blending, character consistency across generations, targeted natural-language edits, Gemini world-knowledge grounding. Available via Gemini API, Google AI Studio, Vertex AI. Priced at $30/1M output tokens (~$0.039 per 1290-token image). All outputs carry invisible SynthID watermark. Debuts #1 on LMArena for both Text-to-Image and Image Editing. via Google Developers
  • Gemini 2.5 Flash Image lands on Vertex AI for enterprise. Same-day enterprise rollout. Production workloads with same pricing and capabilities (multi-image fusion, character consistency, conversational edits, SynthID). via Google Cloud
  • Nano Banana is first image model on OpenRouter. OpenRouter partners with Google to bring Gemini 2.5 Flash Image to 3M+ developers — first image-generation model offered through the API aggregator, which had only served text models among 480+ live models before. via OpenRouter
  • Anthropic launches Claude for Chrome browser extension (research preview). Sidebar extension letting Claude take autonomous browser actions (calendar, email, navigation, form-filling). Limited to 1,000 users on $100/$200/month Claude Max plan, with waitlist. Announcement unusually heavy on prompt-injection warnings: 23.6% attack success without mitigations, 11.2% with safeguards. Blocks usage on financial, adult, pirated-content sites. via Anthropic
  • Bartz v. Anthropic — $1.5B copyright class-action settlement filed. Anthropic and plaintiffs file notice of proposed class-wide settlement in N.D. California, disclosing binding term sheet signed Aug 25. Anthropic to pay at least $1.5B plus interest to class of ~500,000 authors (≈$3,000 per work) and destroy LibGen/PiLiMi pirated-book datasets. Largest publicly reported copyright recovery in US history. Follows Alsup’s June 23 ruling. via Slashdot
  • OpenAI notifies developers of Assistants API sunset. Emails customers announcing deprecation and full removal one year later, Aug 26, 2026. Path forward is Responses API (released March 2025), into which OpenAI folding all remaining Assistants features. via OpenAI Docs

Who shipped

Google shipped Gemini 2.5 Flash Image. Anthropic shipped Claude for Chrome + $1.5B settlement. OpenAI shipped a deprecation notice.

Open-source pulse

fal.ai opens Gemini 2.5 Flash Image edit endpoint — first third-party hosted endpoint alongside Google’s own, signalling rapid uptake by AI infra ecosystem on launch day. via fal.ai

Money, infra & hardware

Simon Willison published “Piloting Claude for Chrome” analysis same day — canonical outside-Anthropic reference for launch. Examines prompt-injection threat model and limits of 1,000-user research-preview rollout. via Simon Willison

Quiet corners

The Register frames Claude for Chrome with “massive warnings” — Anthropic’s own launch post is more about prompt-injection risk than capabilities. Tone-setter for press coverage on how aggressively the browser-agent category is being shipped with caveats. via The Register VentureBeat adds Anthropic explicitly positions limited beta as way to gather adversarial data before any broader rollout.

By the numbers

  • $30/1M / $0.039 / SynthID — Nano Banana pricing / per-image / watermark
  • #1 / 5 / 480+ — LMArena rank / multi-image fusion / OpenRouter models prior
  • 23.6% → 11.2% — Claude for Chrome attack success without/with mitigations
  • 1,000 / $100-$200/mo — Chrome research preview users / Max plan
  • $1.5B / 500,000 / ~$3,000 — Bartz settlement total / class size / per-work

Compiled by AI Feed’s editor from verified web sources for 26 August 2025.