Good morning. Here's what's worth acting on today.

01 · BREAKING STORY

DeepSeek has upgraded V4.1 Flash to DeepSeek-V4.1-Flash-1017 (17 October), with model pricing unchanged, and opened its in-house Search and Fetch APIs. MarkTechPost reports they are open to every developer.

Terms as of 18 October 2026, per DeepSeek's Search & Fetch guide:

DeepSeek reports 77.0 on DeepSWE v1.1, up from the 74.2 it listed for V4.1 Flash in September. Both are its own figures.

My take: Web search for your agent now costs under a cent per call. This week, run a batch of your real queries through DeepSeek Search alongside your current provider and compare result quality, latency and cost per useful answer. If you work in Claude Code or Codex, test before 1 November while those calls are free. Treat the benchmark gain as a claim until your own tests agree.

02 · MARKET TRENDS

Vercel's October AI Gateway Production Index (15 October) puts open-weight models at 71.5% of September tokens and 23% of spend, up from one in twenty dollars in June. Spend was about 13 times December's while the average price per token fell 47%. DeepSeek's models processed 58% of gateway tokens for 6.6% of spend.

Caveats: this is Vercel Gateway traffic only, which excludes direct vendor and on-premises use. Spend is estimated at list prices, and Vercel broadened its open-weight definition, so older months compare loosely. Within those limits, the pattern is consistent: bulk work goes to cheap open-weight models, and frontier models are kept for tasks where quality pays.

How to make money from it: Sell model-routing audits to small teams. Log a week of their AI calls, move high-volume, low-stakes tasks (classification, extraction, summaries) to an open-weight model, and charge a share of the monthly savings.

Proton Mail offers what others won’t:

We don’t scan your emails. We don’t sell your data. And we don’t make you dig through settings to find basic security. Proton is built for people who want control, not compromise.

Simple, secure, and free.

Explore Proton’s benefits

03 · GROWTH HACK / DIY

Be the primary source ChatGPT goes looking for. Two studies suggest ChatGPT is citing original and official sources more and aggregators less. Ahrefs found that among the roughly 160 most-searched sites, 75.9% of ChatGPT's site: searches targeted brand sites, and Reddit fell from about 1 in 5 US answers in March to about 1 in 43 by late September. In US finance, Greenflag Digital saw first-party pages rise from 1.8% to 22.4% of top-500 citations, while pages with dates in their URLs lost nearly all theirs. (Caveats: Ahrefs sells a citation tracker; Greenflag covers US finance only and also draws on Ahrefs data.)

The play: publish your own data on stable, undated pages and keep them current.

What you'll need: a page on your own domain, one spreadsheet of first-party numbers, about a day to build and an hour or two a quarter to update. No extra cost.

Measure citations directly: LLM Pulse reports ChatGPT referral sessions in GA4 fell around 46% from 9 October, likely a tracking change. Check 10–20 of your factual questions in ChatGPT monthly and log whether you're cited.

Full playbook →

04 · AGENT LAB

The pattern. Long agent runs re-send the whole history every step. Asana's browser test agent cached only its instructions and tools, and trimmed screenshots at nearly every step, so the growing history was billed at full price each call. The fixes in OpenAI's case study: extend caching to the agent's browsing history, and let screenshots accumulate to 20 before cutting back to the latest one. On the same model (GPT-6.1 Sol) with the larger history budget, that alone cut cost 4x, from $1.97 to $0.47 per run, and each call was about 3x cheaper because 89% of input came from cache at 5% of the uncached price.

How it works. Providers cache exact prefixes: OpenAI needs the entire rendered prefix to match, and Anthropic needs 100% identical segments, images included. Keep the header fixed, append only, and prune in batches so only the prune step misses the cache. Do not over-prune: raising the history budget from 120,000 to 480,000 characters took runs producing an answer from 3 of 18 to 18 of 18.

Try it. Take one multi-step agent, run 10 fixed tasks, log cached tokens per call, then switch to threshold pruning and compare cost per completed task.

Full breakdown →

05 · TRENDING AI RESEARCH

Old Ideas, Novel Problems (Hebrew University, AI2, Microsoft, Arizona, UW; arXiv, 1 October 2026) tests how reliably AI models judge whether an idea is original. The answer: they are not very reliable.

What it says: Across six judges (GPT-5.1, 5.2 and 5.4, Claude Sonnet 4.5, Opus 4.5 and 4.6), small changes to the wording moved results a long way. For GPT-5.4, two slightly different prompts disagreed on at least 52.6% of the same idea pairs, and one wording pushed it below chance. Supplying related papers helped by at most 16 points, and only for some models. The cheapest plain prompt beat two purpose-built novelty checkers by 8 to 29 macro-F1 points at up to 32 times lower cost (the 32x is the best case, for Claude Opus 4.6). One caveat: the "not novel" labels for AI-generated ideas are assumed, not checked one by one.

Why it matters: If you score hooks, headlines or ideas with AI, the wording of your prompt can count for more than the idea itself.

How to use it:

Full breakdown →

06 · TOOL REVIEW

OpenSEO is an open-source (MIT-licensed) SEO tool that gives your AI agent keyword research, backlinks, rank tracking, site audits and AI-visibility checks over MCP. It works with Claude, ChatGPT and Codex, and the data comes from DataForSEO.

Price: $10 a month with $10 of usage included, as of 18 October 2026. A keyword search costs about $0.05, a backlink overview about $0.08 and an AI prompt check about $0.002. Self-hosting is free, but you pay DataForSEO directly.

Who it's for: solo builders who already work through an agent and treat SEO as one channel among several.

Verdict: the best-value SEO stack for an agent-first solo builder, but it's early software. It is on v0.1.5, the included credit resets monthly, and the hosted version needs an active subscription.

Full review →

07 · WHAT GOT FUNDED THIS WEEK

Miracly · AI search brand visibility · US$2M pre-seed (€1.8M), co-led by BEENEXT and Lotus. Tracks how brands appear in ChatGPT, Claude, Gemini and Perplexity, and automates the follow-up. Interesting: per e27, rival Profound raised at a US$1B valuation in February. Founders: Arvid Lithander, Bjorn Bergstrom, Shannon Cullum.

Dapta · AI sales agents for SMBs · $6M seed, led by Chingona Ventures. Agents that non-technical owners set up to answer calls, qualify leads and book appointments. Interesting: per Portafolio, it has US$4M ARR and 200,000+ registered users. Founder: Nicolás Rojas.

Sharp · First-responder coaching · $18M Series A, led by Hidden Variable. Confidential one-on-one coaching for police, fire and EMS, from academy to retirement. Interesting: it says 110+ departments in 18 states already use it. Founders: Ben Curley, Andrew Sakmar.

Opsin · AI agent security · $21M, co-led by SYN Ventures and Race Capital. Finds a company's AI agents, maps what they can access, and contains misbehaviour. Interesting: if you sell agents to businesses, expect buyers to ask these questions. Founders: James Pham, Oz Wasserman, Jeremy Mailen.

Origin · Construction robotics · $20M, co-led by Accel and Zacua Ventures. Its OG1 robot autonomously finishes interior walls. Interesting: founded in 2025, it says its robots are already on New York sites, and it targets 150 robots by end-2027. Founders: Yogesh Ghaturle, Ashish Daga, Tushar Makkar.

QUICK HITS

That's today. Hit reply and tell me what you're building. I read every email.

Dev