Trending archive · Week of · Latest
AI gets graded on real office work
This week AI models were tested on ordinary business work instead of puzzles. Gemini 4 Carbon finished 53.9% of Zapier's AutomationBench tasks, the best result so far, and OpenAI's GPT-6.1 Terra made computer use much cheaper. But the best agent passed only 10.8% of realistic tax jobs, n8n 3.0 asks self-hosters to tidy up before December, and three stories showed AI handling data in unexpected ways. Test on your own work before you trust it.
The 13 picks
| Who it's for | Have it built | |||
|---|---|---|---|---|
| Gemini 4 Carbon Model launch | Building blocks | Hacker News, 862 points · X, 12,329 likes | Teams choosing a model for automations and agents | Get matched → |
| Zapier AutomationBench Benchmark | Research & reality checks | GitHub, 365 stars · Zapier leaderboard | Owners and ops leads deciding between AI agents and plain workflows | Get matched → |
| GPT-6.1 Terra Model launch | Agents that use computers | Hacker News, 633 points · X, 27,874 likes | Finance and ops teams retyping data between systems | Get matched → |
| DAYJOB: Tax Benchmark | Research & reality checks | Hacker News, 219 points | Accounting firms and finance teams | Get matched → |
| n8n 3.0 beta Major release | Automation platforms | X, 237 likes · GitHub, 208,944 stars | Anyone running n8n, especially self-hosted | Get matched → |
| Cursor Agent SDK Developer tool | Building blocks | X, 3,915 likes · Hacker News, 110 points | Teams commissioning custom agents and the agencies that build them | Get matched → |
| Claude for Legal artifacts Product launch | Personal & team assistants | X, 24,110 likes · GitHub, 9,824 stars | Law firms and in-house legal teams | Get matched → |
| Claude for Nonprofits plugin Product launch | Personal & team assistants | X, 7,106 likes | Nonprofits and charities | Get matched → |
| Claude ratings and training data Policy explainer | Running agents safely | Hacker News, 224 points | Admins of company AI accounts | Get matched → |
| Cloudflare data mix-up Incident report | Running agents safely | Hacker News, 181 points | Teams sending sensitive data to hosted AI models | Get matched → |
| Open model tried a hidden upload Research | Running agents safely | Hacker News, 329 points | Teams running open-weight models or agents on their own servers | Get matched → |
| OpenDuet Open source | Personal & team assistants | GitHub, 1,385 stars (created Oct 21) | Teams that want a self-hosted personal assistant | Get matched → |
| Lampline Open source | Building blocks | GitHub, 547 stars | Clinics, restaurants and sales teams handling lots of calls | Get matched → |
Gemini 4 Carbon Building blocks
Google's new model, announced Oct 28, ranks first on the Artificial Analysis Intelligence Index and targets long, multi-step tasks and cybersecurity. Access is limited to cyber defenders and trusted partners for now, with businesses "soon". Introductory API price is $2 / $10 per million input / output tokens, later $4 / $20. Source: Google: Introducing Gemini 4 Carbon
What it means for your team: The strongest model for multi-step business tasks right now, but not yet generally available. Plan to test it on your own cases when access opens rather than switching on benchmark results alone.
Zapier AutomationBench Research & reality checks
Zapier's test of 657 everyday tasks across 47 simulated business apps (sales, marketing, ops, support, finance, HR). Gemini 4 Carbon on high effort completed 53.9% at about $2.12 per task, ahead of Gemini Argon (51.3%), Claude Sonnet 5.5 (44.75%) and GPT-6 Astra (41.4%). A task only counts if every step is right. Source: Zapier: AutomationBench report on Gemini 4 Carbon
What it means for your team: Even the best model gets a business automation fully right about half the time on its own. Use fixed workflow steps for routine parts and AI only where judgement is needed, with checks.
GPT-6.1 Terra Agents that use computers
OpenAI's model for agents that use apps and websites. API prices are half of GPT-5.6 Terra ($1 in / $5 out per million tokens). OpenAI's example is reading invoices and entering them into accounting software, and it says errors on its tricky-prompt test fell from 53.6% to 16.7% at low effort. In ChatGPT Work, Codex and the API. Source: OpenAI: Introducing GPT-6.1 Terra
What it means for your team: Retyping data into software with no API is getting cheap to automate. The cost and error numbers are OpenAI's own, so trial it on a sample of your documents with a person checking.
DAYJOB: Tax Research & reality checks
Surge AI gave agents 50 realistic tax assignments with 1,789 files. The best, Claude Sonnet 5.5, passed 10.8%, then Fable 5.1 (9.6%) and Claude Opus 5.5 (9.2%). Misses included a home-sale deadline and an unchecked assumption about marital status. Source: Surge AI: DAYJOB Tax
What it means for your team: Agents are good at gathering and sorting, weak at getting long, high-stakes jobs fully right. Let them chase documents and prepare drafts; keep a qualified person signing off.
n8n 3.0 beta Automation platforms
n8n's 3.0 beta (Oct 27) focuses on reliability and security, with stable release planned for early December. Self-hosting becomes Docker-only, older nodes such as Function, Item Lists, Cron and the legacy OpenAI node are removed, Code steps time out after 1 minute instead of 5, and unverified community packages are off by default. A Migration Report in v2.39+ lists and fixes affected workflows. Source: n8n docs: v3.0 breaking changes
What it means for your team: If you self-host n8n, run the Migration Report now so nothing breaks in December. Cloud users get the upgrade but old workflows may still need fixes.
Cursor Agent SDK Building blocks
Cursor made its Agent SDK generally available: agents defined as code with plain-text instructions, built-in tests (evals) and replay of each run, deployable on Cursor or self-hosted. Source: Cursor: The Cursor Agent SDK
What it means for your team: Testing and replaying agents is becoming standard. Ask any builder how they will test an agent on your cases and how you will see what it did.
Claude for Legal artifacts Personal & team assistants
A gallery of ready-made legal outputs in Claude: citation check, jurisdiction survey, matter board, contract playbook deviation board, regulatory horizon scan and tabular review. Each item links back to its source and lawyers make the call. Source: Claude: Introducing artifacts for legal work
What it means for your team: AI is arriving packaged for specific professions. Check whether a packaged tool covers most of the job before paying for a custom build.
Claude for Nonprofits plugin Personal & team assistants
A Claude plugin for fundraising, marketing and hiring, built with about 30 nonprofits. Anthropic says more than 25,000 nonprofits use Claude. Source: X: @claudeai, Claude for Nonprofits
What it means for your team: Small teams with no tech staff can start with a packaged plugin before commissioning anything custom.
Claude ratings and training data Running agents safely
On business Claude plans chats are not used for training by default, but a thumbs up or down stores the whole conversation for up to 5 years and it may be used for training. Team and Enterprise owners can turn ratings off under Organization settings, Data and Privacy. Source: Claude Privacy Center: Is my data used for model training?
What it means for your team: Check one setting: who can rate chats in your company and whether ratings should be off for sensitive work.
Cloudflare data mix-up Running agents safely
Cloudflare disclosed that between Aug 27 and Sep 9, 155 responses from one hosted model contained parts of other customers' data, only for extremely long prompts. Fixed on Sep 9, with no sign of deliberate abuse. Source: Cloudflare: KV cache exposure and remediation
What it means for your team: Shared AI hosting can leak in rare cases. Keep the most sensitive data out of prompts or use providers and settings with stronger isolation.
OpenDuet Personal & team assistants
An open-source template for a proactive personal agent that asks before it sends, pays, signs or deletes. Early development. Source: GitHub: CopilotKit/OpenDuet
What it means for your team: A starting point for a private, self-hosted personal assistant if packaged ones don't fit your data rules.
Lampline Building blocks
Open-source, self-hosted phone agents. Its authors estimate $0.02 to $0.17 per call and test calls with an AI caller before launch. Source: GitHub: C4rohan/Lampline
What it means for your team: Voice agents you host yourself may cut per-call costs; treat the numbers as the authors' estimate and test call quality first.
Want this built for your team?
Describe the job in a sentence. We compare it with what 970 agencies list on their own websites and show the 5 best fits. Free, no sign-up.
✓ 970 agencies checked against their own sites✓ 67 listed in official partner directories✓ Rank can't be bought
Trending AI tools, once a week
A short email with the week's useful launches and what they mean for a business. No spam, unsubscribe anytime.