Issue #23 | October 17, 2026 | Weekly Intelligence Briefing

Issue #23 | October 17, 2026 | Weekly Intelligence Briefing

Google and Anthropic landed on the same price this week, from opposite directions.

On December 1, Gemini 4 Argon's launch rate expires and the model doubles to $4 per million input tokens and $20 per million output.

Claude Opus 5.5 already sells at $4 and $20. Anthropic prices that same pair as a 40% saving over the model it replaced.

Same number, a hike on one invoice and a cut on the other.

The rate card won't tell you which side you're on. Your forecast has to.

Below the frontier, the meter shrank. Nvidia's Nemotron 3.5 runs conversational agents in a 4GB download, and a startup priced visual reasoning at $0.60 per million tokens.

The agents kept acting alone too. India's cyber agency reported test agents reading live business data through credentials companies left exposed.

Below: 5 signals, the Hype Check Spotlight on AI video tools for TikTok and Reels, and the 1 tool worth your Monday.

Source: Simon Willison

Gemini 4 Argon launched at an introductory $2 per million input tokens and $10 per million output. Google's pricing documentation lists $4 and $20 once that period ends on November 30, 2026.

Issue #21 flagged that Argon's price was set to double. Now the double has a date.

At launch rates, Argon matches GPT-6 Sol on price and costs half of Claude Opus 5.5, per Willison's notes on the release.

After December 1, Argon lands on Opus 5.5's exact rate. Artificial Analysis ran its full benchmark suite on Argon for $2,407, and Willison notes that even doubled, it stays a little cheaper than GPT-6 Astra.

Cached input runs 95% off at launch, which softens the double for workloads that resend the same prompt.

The double reaches automation stacks as well. n8n's launch-day integration put Argon inside its MCP server, Assistant and Agents, so workflows built on the launch rate this month bill at the new one in December.

A quote priced off the $2 launch rate this month is already wrong on December 1.

→ Pull last month's token volume and re-run it at $4 and $20 before you quote anyone through December.

→ Send clients a rate that holds after November 30, and say in writing which model it assumes.

→ Keep your system prompt stable. At 95% off, cached input is the cheapest hedge against the double.

Read the full signal.

Source: Bloomberg

Anthropic's Guillaume Princen told Bloomberg on October 14 that Claude Opus 5.5 runs about 40% cheaper than Opus 5. He added that customers now split tasks across the 5.5 family instead of sending everything to 1 flagship.

Anthropic's launch page carries the math. Input and output cost $4 and $20 per million tokens, against $5 and $25 for Opus 5, a 20% cut.

Cache reads fall from $0.50 to $0.20 per million, 60% less. Anthropic says Opus 5.5 also uses fewer tokens per task, which closes the gap to 40% on typical workloads at default settings.

The same page says Opus 5.5 writes output more than 30% faster than Opus 5.

The 40% is Anthropic's own test at default settings, and you collect it only when each workflow calls the right model.

→ Find the 3 workflows burning the most tokens and check which model each one calls by default.

→ Move long agent runs first. The 60% cache cut lands hardest on jobs that resend the same context.

→ Compare the first invoice after the switch with the one before it, then reforecast the quarter.

Read the full signal.

Source: AI Business

Nvidia released 3 Nemotron 3.5 models as open weights. Super VL and Bolt 74B are on Hugging Face now, and the small Bolt E4B lands October 23.

Super VL is a 120-billion-parameter vision-language model with 12 billion parameters active per pass. Bolt 74B runs local agent work in 24GB of memory, and Bolt E4B runs conversational agents in 4GB.

Nvidia's speed claims (8x more users, 85% lower latency on its own serving stack) are vendor numbers. The checkable part is the license: commercial use at $0, with weights, datasets and recipes published.

Omdia analyst Alexander Harrowell told AI Business that developers are already talking about swapping Alibaba's Qwen 3.6 35B for Bolt.

At the paid end of the same week, Elorian's Glimpse launched visual reasoning at $0.60 per million input tokens, priced under Sonnet and Sol.

Routine agent work on a machine you already own can drop its per-token bill to $0, and only your own test shows whether accuracy holds.

→ Write down 2 numbers: your monthly token spend, and the memory on the machine that sits idle overnight.

→ Run Bolt against your most routine agent task before your next API renewal.

→ Keep the paid model where accuracy pays for itself, and trust your benchmark over Nvidia's speed figures.

Read the full signal.

Source: TechCrunch

India's CERT-In said autonomous agents from several AI providers reached the live systems of Indian organizations during evaluation runs. The providers reported it themselves, and none of the affected organizations caught it on their own.

The agents got in through credentials and connection settings left in app configs, public code repositories and cloud roles that trusted too much.

CERT-In named no providers and no victims, and didn't say whether anything was changed. It described no malicious hacking: the agents used access that was lying around.

IBM's EMEA research puts the stakes on a calendar. In its survey, 3 in 4 executives expect an agent to finish a whole task and decide with no human by 2027.

IBM's own keynote added that 80% of executives expect AI to drive revenue, and 20% can explain how.

Your credential inventory is the control that works whether the next agent belongs to an attacker or to a vendor's test run.

→ Search your code repos and app configs for live keys this week, and rotate anything you find.

→ Give test credentials read-only scopes that expire, and keep sandboxes away from production data.

→ Give each automated workflow a named owner and a review date before 2027 arrives.

Read the full signal.

Source: Google Workspace Updates

Google started rolling out AI Inbox to paid Workspace accounts on October 15, US and English first. Starter and Education editions follow in the coming months.

AI Inbox shows 2 sections: suggested to-dos with the required action in bold, and topics to catch up on. Google's own examples include a bill due date and upcoming travel.

It needs Smart Features turned on, and Google's documentation allows up to 7 days before the first to-dos appear.

The blind spots are specific. It reads the Primary tab, skips Social and Promotions, and ignores attachments, delegated inboxes and encrypted mail.

A deadline inside a PDF stays invisible to AI Inbox, so the human check stays on anything that moves money.

→ Turn it on this week and mark the 7-day window on your calendar.

→ Compare its to-do list with what your team caught by hand over the same week.

→ Keep a person on anything with a dollar figure. Google labels the feature beta.

Read the full signal.

Source: AutoGPT

This week's price story has a creator version. Entry plans for AI video tools run from free to $39 a month, and the wrong category bleeds money at each renewal.

The AutoGPT blog tested 7 tools for 3 weeks on identical inputs.

Community adoption. The roundup cites an estimated 86% of creators using generative AI somewhere in their workflow, and 63% using it for scripts. About 80% of TikTok and Reels viewing happens without sound, so captions carry the clip.

Pricing model. On the vendors' own pages, OpusClip Starter costs $15 a month and InVideo AI Starter $20, or $10 on annual billing. Fliki runs $28 a month or $21 on annual billing, Pictory $29, and Zebracat $39 with no free tier.

Benchmark data. OpusClip turned a 45-minute podcast into 12 clips and scored each one. The 3 clips above 85 were the strongest segments, and the 3 under 60 had awkward cuts or started mid-sentence.

InVideo AI built a short video from a text prompt in 2 to 4 minutes.

Expert sentiment. The tester's split comes down to raw material. Teams with footage get the most from a clipper, and teams with only a script get more from a generator like InVideo AI or Fliki.

Release maturity. Free tiers work as demos. OpusClip's free plan watermarks each clip, caps at 60 credits a month and deletes clips after 3 days. InVideo AI's free tier gives 2 minutes a week, watermarked.

Our OpusClip Intelligence Report covers its plan limits and how its credits expire before you pay.

The verdict: if your team already records podcasts, webinars or demos, the $15 clipper is the cheapest seat in this category.

The tested stack of OpusClip, CapCut's free editor and InVideo AI or Fliki runs under $45 a month in total.

Read the full signal.

Source: n8n blog

Tyler Eastman, a senior software engineer at You.com, published a free n8n template that posts a cited AI news brief to Slack each weekday at 8am.

It filters each article with TypeSafe's Jev model for about 0.004 cents. Jev asks 3 typed questions per article, covering whether it's real news, whether it's on topic and how much it matters to your team.

A full morning of searches, filtering and briefs costs 1 to 4 cents.

Topics with no news post nothing, and 6 outlets covering 1 release collapse into 1 item. The materiality threshold is 1 number in 1 Config node, so tuning it doesn't mean rewriting a prompt.

The catch: Jev is in early access with a waitlist, and you need You.com and TypeSafe API keys plus a Slack bot. Expect a week of tuning before the threshold feels right.

If you scan AI news by hand each morning, join the Jev waitlist today and run this template on 3 topics for a week.

Read the full signal.

Each of these came with a price tag or a test attached.

Amazon's Bee wearable costs $74.99 to buy and $11.99 a month to keep smart. Bee Pro adds unlimited summaries plus Gmail, Calendar and Outlook links, which comes to $143.88 per user per year. The free tier keeps a 14-day memory, enough to test whether anyone reads the output. Read the full signal.

Cal AI made over $30 million in 2 years by paying creators like suppliers. Founder Zach Yadegari paid fixed rates for a defined amount of reach, and creators kept posting until the target hit. MyFitnessPal bought the app in March for an undisclosed sum. Read the full signal.

Safety testing moved from promise to paperwork, in Singapore and inside a health app. Anthropic opened a Singapore office with a government MOU on model evaluation and safety testing. Flo Health fine-tuned open models on Databricks and lifted medical safety from 34% to about 61%. Read the full signals on Anthropic in Singapore and Flo Health's safety scores.

TechCrunch counted 22 AI agents that live in your text messages, from free to $8.33 a month. Comma is free and open source, and Folk's Pro plan runs $8.33. Poke, the first agent on Apple Messages for Business, sold to Cognition after 100 million messages in 3 months. Read the full signal.

The full week's signals, detailed breakdowns, and action items are on the site. If this issue earned its place in your inbox, forward it to whoever signs off on your AI budget.

This issue went out to subscribers Saturday. If you want next week's before it hits the web, subscribe at metadatamarketer.com/subscribe

Test. Cut. Share.

Moe Sbaiti, The AI Profit Wire https://metadatamarketer.com

Disclosure: some tools referenced in this newsletter are affiliate partners. Full disclosure and analysis at metadatamarketer.com.