inish.inSunday, 23 August 2026243 scanned · 6 kept

Built for one reader

Nish's Daily Reads

AI, product ideas, and where demand is building. Plain words, point first, nothing here unless there is a fact under it.

Editor’s note

I opened 243 candidates. The week's big stories — Stripe-OpenRouter, Replit's free tier, OpenAI-Anthropic share, Pew AI-authorship, Micro1's $500M — were covered in the last three editions, so I dropped them. Six new stories survived: OpenAI is cutting frontier prices and building an ad business, a 27B model beat frontier labs at science, and a Chinese model got too good at hacking to release.

01
AIOpenAI Developer Community

OpenAI cuts GPT-5.6 Sol API pricing by over 20%, Luna by 80%

OpenAI cut the developer price of its frontier GPT-5.6 Sol model by more than 20% for the next three months. The input price dropped from $5 to $4 per 1 million tokens (roughly, chunks of text the model charges by), and the output price dropped from $30 to $20. The cheaper Luna model was cut by 80%. The cuts apply to the API, ChatGPT Work credits, and Codex credits, but not to Pro, Plus, or Business subscription pricing. Reuters confirmed the move is a response to competition from Anthropic and Chinese AI models.

Checked GPT-5.6 Sol input dropped from $5 to $4 per 1M tokens; output from $30 to $20. Luna dropped 80% (input $1 to $0.20). Promotional pricing runs through November 21, 2026.

Nish This directly cuts my agent fleet costs. Sol is the model my workers use for hard tasks, and a 33% output price drop is real money at fleet scale. I would lock in builds against these prices before November, because OpenAI has not promised the cuts last.

But The cuts are promotional, through November 21, 2026. OpenAI has not said whether they become permanent. The price war is driven by competition, so if pressure eases, prices could revert.

Read at community.openai.com ↗
02
AITechCrunch

A 27-billion-parameter model from DeepMind alumni beat Claude and GPT at replicating science

Inherent, a London AI lab founded by Google DeepMind alumni, says its AI agent Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 at independently reproducing the findings of published scientific papers. Faraday runs on Qwen 3.6, a model with 27 billion parameters — a fraction of the size of the frontier models it beat. Parameters are a proxy for a model's size and training cost. Inherent emerged from stealth weeks ago with a $50 million seed round and has 12 employees.

Checked Faraday runs on Qwen 3.6 (27 billion parameters) and outperformed Claude Opus 4.8 and GPT-5.5 at replicating research. Inherent raised a $50M seed and has 12 employees.

Nish If Faraday can beat frontier models at replicating science with 27 billion parameters, I don't need the biggest model for every problem. The big labs' moat is narrower than their valuations suggest.

But Paper replication is one task, not a general benchmark. Inherent is self-reporting the result. The company has 12 employees and no revenue model shown. Qwen 3.6 is a Chinese model.

Read at techcrunch.com ↗
03
AIBetaNews

Z.ai is holding back its new model's open weights because it got too good at hacking

Chinese AI lab Z.ai announced GLM 5.3, a new open-weight model that scored 84.5% on CyberGym, a benchmark for finding known vulnerabilities in source code — beating Anthropic's Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%). The model found 2,436 potential vulnerabilities across 269 open-source projects, including 1,097 rated critical or high severity, with flaws in the Linux kernel, VMware, and Apache. Z.ai is delaying the release of open model weights for two weeks after the Aug 14 launch because the cyber capability developed faster than expected.

Checked GLM 5.3 scored 84.5% on CyberGym (up from 77.2% for GLM 5.2). It found 2,436 vulnerabilities across 269 projects, 1,097 critical or high. Open weights delayed two weeks after Aug 14 launch.

Nish A Chinese open-weight model is now state of the art at finding vulnerabilities. If I run production code, this is a tool I could use to scan for bugs. Z.ai delaying the weights because GLM 5.3 got too good at hacking tells me guardrails are lagging.

But Benchmark scores are Z.ai's own testing, not independently verified. CyberGym tests white-box source code, not real-world exploitation. The 2,436 findings are from Z.ai's own program, not audited.

Read at betanews.com ↗
04
Demand signalsOpenAI

ChatGPT Ads is expanding to 31 European countries

Six months after testing ads in the United States, OpenAI is bringing ChatGPT Ads to 31 European markets, including Germany, France, Spain, Italy, and the Netherlands. Tens of thousands of marketers have already advertised on ChatGPT. Ads are shown only to users on the Free and Go plans; paid subscribers on Plus, Pro, and Enterprise see no ads. OpenAI has added conversion optimization, geo-targeting, and measurement tools since the pilot began.

Checked ChatGPT Ads is expanding to 31 European countries. Tens of thousands of marketers have advertised since the US pilot began February 2026. Ads appear only on Free and Go plans.

Nish OpenAI is building an ad business. If ChatGPT becomes a place where people research products, it becomes a channel I need to advertise on, the same way Google Search is. Scaling to 31 countries in six months tells me they are seeing real ad revenue.

But This is OpenAI's own announcement, not independent reporting. No ad revenue figures or fill rates were shared. Self-service access is not yet available — advertisers must go through the OpenAI Ads team.

Read at openai.com ↗
05
Product ideasTechCrunch

ChatGPT can now read, draft, and send your text messages

OpenAI launched an Apple Messages plug-in for ChatGPT that lets the chatbot read, sort, search, draft, and send text messages on a user's behalf. The plug-in works with Codex and ChatGPT Work, so it can be used professionally. It requires Full Disk Access on macOS, meaning the app can read and write any file on the device. OpenAI says message content is stored locally, not on its servers, and that ChatGPT does not create a full index of messages. OpenAI warns users against turning on persistent approval, which removes the final review before ChatGPT sends a message as you.

Checked The ChatGPT Apple Messages plug-in can sort, draft, send, and delete messages. It requires Full Disk Access on macOS. OpenAI warns persistent approval "removes your final chance to review a message before ChatGPT sends it as you."

Nish ChatGPT is becoming an operating system layer. If it can read and send my texts, it can manage my inbox and follow up with contacts. The privacy tradeoff is real — Full Disk Access is a big ask — but the productivity gain locks users into ChatGPT.

But Full Disk Access gives the app access to all files on the device. OpenAI's privacy framework is described but not independently audited. The plug-in is macOS-only at launch.

Read at techcrunch.com ↗
06
WildcardTechCrunch

Harvard is selling AI clones of its professors for $699

Harvard Business School's eight-week Foundry bootcamp for entrepreneurs now includes AI avatars of its instructors, built by startup HeyGen. The avatars provide feedback during practice pitches and board meetings. The program costs $699. A New York Times reporter tested it by pitching to an AI clone of Flybridge Capital co-founder Jeff Bussgang, who called his digital copy creepy but said students love it. Students told the reporter they preferred the avatars over a simpler chatbot experience.

Checked Harvard's HBS Foundry bootcamp costs $699 for eight weeks and uses AI avatars by HeyGen. Instructor Bussgang called his avatar creepy but said 'My students love it.'

Nish If Harvard can ship AI clones of its professors and students prefer them, I see the same model working for any expert who teaches or advises. The product is not the AI — it is the trusted name attached to it.

But This is a TechCrunch In Brief summarizing a NYT report. The bootcamp still includes live weekly sessions with real instructors — avatars handle practice exercises, not the full curriculum.

Read at techcrunch.com ↗