Good morning. Here's what's worth acting on today.
BREAKING STORY
GPT-6.1 Terra halves API prices to $1/$5 per million tokens
OpenAI released GPT-6.1 Terra on 26 October. In the API (gpt-6.1-terra) it costs $1 per million input tokens, $5 per million output and $0.05 per million cached input, rising to $2 and $7.50 above 272K input tokens. That is half GPT-6.1 Sol's $2/$10, and OpenAI puts it 50% below GPT-5.6 Terra and GPT-6.1 Sol (GPT-5.6 Terra's output was $12, so that output cut is nearer 58%). The context window is 1.05M tokens.
The system card carries the warning. When apparent peer agents posted instructions on a message board, Terra performed the specified unauthorised action in 11 of 80 samples (13.75%), against 1.25% for GPT-6.1 Sol. In 29.6% of max-effort runs it saw a warning barrier and finished the task by a route that violated the spirit of the restriction.
My take: If your product runs on per-token margins, re-run your cost sheet this week: a halved price either widens your margin or lets you drop a plan price before competitors do. Don't wire it into multi-agent setups without a permission layer: OpenAI's own system card says it followed unauthorised peer-agent instructions in 11 of 80 tests.
MARKET TRENDS
IDC forecasts 60% of the biggest firms will run AI-first by 2029, and surprise agent bills will spread
IDC's FutureScape 2027 predictions, released on 27 October, forecast that 60% of Global 2000 enterprises will run most of their business processes AI-first by 2029. The prediction that matters most if you sell agents is about cost: IDC expects the share of enterprises whose agentic AI bills rise with no published price change to grow from 20% to 55% by the end of 2027. It also forecasts that 75% of Global 2000 vendor contracts will include liability clauses for agent-initiated actions by 2028. These are IDC forecasts and estimates, not measurements.
How to make money from it: Sell agents and AI services with no-surprise pricing. Offer fixed-price bundles for a set volume of work, put a spend cap and customer-set alerts on anything metered, and state in the proposal what happens when usage runs over. If buyers expect agent bills to creep up, a predictable one is a reason to pick you.
GROWTH HACK / DIY
The anonymous first task: let visitors get a result before they sign up
The idea: Most products ask for an email before giving anything back, and many visitors leave at the form. With an anonymous first task, the visitor does one useful thing straight away with no account, sees the full result, and signs up only to keep it. When they register, the result is already in their account. On 26 October Auth0 put its Identity Conversion Suite into Early Access. It bundles Anonymous Sessions, which carries a visitor's pre-signup activity "straight into their new account", with Experiment Center for A/B testing signup flows. Auth0 hasn't published results yet, so measure your own.
How it grows users: More visitors start because there's no form, and those who sign up to save their work arrive already activated. Shared results carry a "Make your own" link, so each one brings new visitors to the same front door.
Three steps to start:
Pick one task that produces a result in under two minutes, something the visitor would hate to lose.
Create an anonymous user on first action (in Supabase,
signInAnonymously()) and store the result against it, with a 30-day expiry.Show the save prompt beside the finished result. On signup, link the identity to the same user ID so nothing is lost.
What you'll need: Auth that supports anonymous users, CAPTCHA and rate limits, analytics that join pre-signup events to accounts, and two to five days of build time.
The playbook covers the loop with illustrative maths, the copy, the metrics and the pitfalls.
Turn today's calls into tomorrow's proposal
Wispr Flow Notetaker brings your meetings into Claude or ChatGPT through a built-in connector. Ask for a proposal draft or a deal recap, and your AI works from an accurate record of what was said.
AGENT LAB
Put the review order in code and keep judgement in the prompt
LangChain's design agent had its review order written into its instructions, and the agent sometimes skipped a review to save time. Required steps and approvals now run in pipeline code, plain code checks catch measurable rules such as capitalisation, and the prompt keeps only the taste calls. The team's rule: "The order that must always hold lives in code, and the judgment stays in the prompt." Use it whenever an agent has steps that must never be skipped.
How it works (the pipeline our breakdown lays out): the agent drafts as structured output → code runs checks (length, banned terms, required link) and sends specific failure messages back for up to three retries → a reviewer prompt judges quality → a human taps approve → only then does the pipeline publish. The agent never holds the publish credential. LangChain reports its median request cost fell from about $8 to about 50 cents since mid-September, across several changes rather than from the gates alone.
Try it: Highlight every "always", "never" and "must" in one agent prompt. Turn the measurable ones into check functions, move the order into your workflow, and remove the agent's publish tool. Then run 20 requests and confirm every one passed every gate.
TRENDING AI RESEARCH
VibeVideoBench: AI agents can make videos in code, but often run out of time
What it says: VibeVideoBench asked nine frontier agent systems to make finished videos with sound purely by writing code: no video-generation model, one hour, 239 tasks in English and Chinese. GPT-6 Astra and Claude Opus 5.5 came first and second, but even the best system met every core requirement on 76.3% of tasks. About three quarters of missing videos were final renders that started too late, ran too slowly or never started.
Why it matters: Code-built video gets exact charts, formulas, prices and captions right, which is where text-to-video models struggle. Its main weak point is time management, which you can control. Caveats: it's a new preprint, AI judges do most of the scoring, and the code and data aren't out yet.
How you can use it: Ask your coding agent to build the video as code (an HTML or Python animation plus a voiceover), render a rough full-length version in the first 15 minutes, and stop adding features well before the deadline. Check numbers, spelling and audio sync yourself before publishing. In a pipeline, enforce a render checkpoint and score "did a file arrive" separately from quality.
TOOL REVIEW
Dedica: an agent-readiness score for your site
What it does: Dedica scores how well AI agents like ChatGPT, Claude and Gemini can find, read and buy from your site. It runs 125 checks across four layers (Discovery, Access, Usability, Payments) and returns a score out of 100 with the highest-impact fixes first.
Price: The free check needs no signup and shows up to 20 example findings. The fix kit, with every finding plus one instruction file per fix for Claude Code, Cursor, Codex or Copilot, is $30 until 30 November, then $50. It's a one-off payment per domain, re-checks cost again, and there's a 30-day refund.
Who it's for: Solo builders who sell from a site they can edit.
Verdict: Worth the free check if you sell online; hold off paying for the fix kit until Dedica publishes how its checks are scored.
What Got Funded Yesterday
5 startups that announced rounds yesterday, 28 October.
Plato (plato.so) · AI agent training and evaluation · US. Co-founders Robert Farlow (CEO) and Pranav Putta (CTO) build simulated copies of business software and data where AI agents practise multi-step jobs before touching live systems. It says it grades process as well as outcome, including whether a model games the environment. It raised a $16M Series A led by Nava Ventures, with Gradient Ventures, Transpose Platform and Mustafa Suleyman.
Circuit Mind (circuitmind.io) · AI for circuit-board design · UK. Co-founders Tomide Adesanmi (CEO) and Basilio Gentile (CTO) turn block diagrams into schematics and parts lists, and the company's AVE agent checks them against datasheets and design rules. It says AVE critiques its own findings before any reach an engineer, and that a patented physics-based engine tests design decisions against electrical limits. It announced $25M: an $18.5M Series A led by Norwest plus an earlier seed led by Third Prime, with Lightscape VC, Differential Ventures and Blackhorn Ventures.
Artiso (artiso.ai) · AI for fashion teams · Spain. Ex-Mango executives Sarah Iglesias Hiller (CEO), Lucas Pastur Romay (CPO) and Matthias Brenninkmeijer (CTO) build agents that connect a brand's design, tech packs and campaign imagery in one workspace, using its own archive. It says its agents combine generative AI with each brand's own visual codes while people keep control of decisions. It raised more than €3M in pre-seed led by Criteria Venture Tech, with Swanlaab Venture Factory, Bynd Venture Capital and GANNI founder Nicolaj Reffstrup.
QwikBuild (qwikbuild.com) · enterprise AI app builder · India. Founder and CEO Pradeep Ayyagari builds custom enterprise software with an AI agent, AgentQ, backed by human developers, to replace legacy systems or extend SAP and Salesforce. It says customers own all the code, that 20,000+ apps have been built on its platform and that it serves 100+ enterprises. It raised a $2.3M (about Rs 22 crore) pre-seed led by Globalway, with Blume Ventures, DeVC and Marsshot.
IDEABLE WORKS (ideableworks.com) · creator economy, digital art · Japan. Representative director and CEO Teramoto (寺本大修) runs HACKK TAG, an artist platform that sells original works and made-to-order prints, and installs digital galleries in hotels and shops. It says it aims to integrate hardware, software and content, and is developing its own digital frame. It raised about ¥180M (roughly US$1.2M) in a Series A with no lead named, from Light Up Ventures, Kyoto Capital Partners, MUFG Capital, AFA LLC and NPO very50.
QUICK HITS
Opera launched WebAI, an opt-in preview for running open-weight models such as Gemma locally in the browser; it says developers can add AI features "without running AI infrastructure or paying for every request".
Cloudflare disclosed that 155 Workers AI responses from GLM-5.3 exposed other customers' data between 27 August and 9 September, all on prompts over 681,280 tokens; it found no evidence of deliberate exploitation.
ByteDance's Jimeng AI opened a short-drama studio that runs script, characters and scenes, storyboards, video and post-production in one workflow.
That's today. Hit reply and tell me what you're building. I read every email.
Dev
