Why This Matters
Every business considering Artificial Intelligence has the same question: do I build custom tools or use what's already free? A custom pipeline costs time and money. Perplexity costs nothing. We ran both side by side for three days to find out if building your own is worth it.
Total experiment cost: approximately $0.02 across 3 runs. Total infrastructure cost: $0, runs on an existing Raspberry Pi.
Experiment Findings: Closed After 3 Runs
The Verdict
Three runs. Same prompt. Pattern locked. Oracle and Perplexity are not the same product and they never were.
Oracle reads like a consultant building a client brief. Geopolitical risk on AI infrastructure. AI labor strikes. MIT-level research on task automation vs. job automation. The kind of content you hand someone walking into a board meeting. Cost: under one cent per run. Fully automated. Runs at 6am without being asked.
Perplexity reads like a practitioner deciding what to ship this week. Goldman Sachs small-business stats. Cursor 3 agentic coding. Salesforce adding features to Slack. Actionable, surface-level, clear "so what" for operators. Cost: $0.00. Manual. No infrastructure.
The interesting finding is not which one is better. It is that they sorted into completely different roles without being told to. Same prompt, same moment, two distinct editorial instincts emerging by Day 2 and confirmed by Day 3.
The real question the experiment answers: is a custom-built pipeline worth building when free tools exist? After three runs, yes, but not for the reason we expected. Oracle wins on framing and depth. Perplexity wins on breadth and zero cost. The right stack for a consulting practice uses both: Oracle for the client brief, Perplexity for the builder's inbox.
This experiment is closed after 3 runs. The Oracle infrastructure continues to run daily for internal use.
Runs completed
3
Apr 6, 7, 8: 2026
Oracle total cost
~$0.02
3 automated runs, Claude Haiku
Winner
Both
Different jobs. Not the same tool.
The Prompt (same for both)
"What's happening in AI this week that matters to a small business operator or someone building with AI? Give me 6-8 stories with context and why it matters."
Run: April 6, 2026
Frequency: Weekly (manual for now, automated later)
Oracle model: Claude Haiku
Perplexity: Free tier (sonar)
● Latest: April 8, 2026
Oracle Output: Apr 8, 2026
Anthropic's Project Glasswing: AI finds security holes in every major OS and browser
Anthropic launched a new model via Project Glasswing, a multi-company cybersecurity initiative with Nvidia, Google, AWS, Apple, and Microsoft. The model already identified vulnerabilities in every major operating system and browser with minimal human intervention. Positioned as enterprise and government-grade automated vuln discovery.
Source: The Verge
Anthropic Mythos previewed: cybersecurity focus, limited early access
Anthropic is previewing a new model called Mythos through a cybersecurity initiative, initially limited to a small set of high-profile companies. Defensive security work only. Signals Anthropic building a safety-first moat before wider release.
Source: TechCrunch AI
MIT: Stop adding agents: redesign processes around them
The real unlock isn't deploying agents into existing workflows, it's redesigning processes to be agent-native. Oracle's angle: bolt-on vs agent-first is the consulting conversation nobody is having yet. The companies that get this right in the next 18 months will be hard to catch.
Source: MIT Technology Review
Arcee: 26-person team releasing competitive open-source LLM
A U.S. startup with 26 employees is shipping models that compete with major labs. Already gaining traction among OpenClaw users. Oracle flagged this as a signal for operators: capable open models outside big labs means vendor optionality is real and growing.
Source: TechCrunch AI
ProPublica union strikes over AI: change management is now a labor issue
ProPublica's ~150 unionized staff staged a 24-hour strike with AI policy as a key negotiation point. When AI arrives without clear guardrails on job impact and process changes, it creates trust debt. Oracle's framing: AI operating agreements aren't optional anymore.
Source: The Verge
Firmus (NVIDIA-backed) hits $5.5B valuation on AI data center buildout
Raised $1.35B in six months. The picks-and-shovels infrastructure play is accelerating. Oracle flagged it as a reminder that AI success is increasingly a facilities and compute ops problem, not just a model selection problem.
Source: TechCrunch AI
Perplexity Output: Apr 8, 2026
OpenAI pushing ChatGPT toward a super-app: 900M weekly users
OpenAI closed a huge funding round and is packaging ChatGPT, Codex, search, and agents into one platform. If your customers or team already live in ChatGPT, it's positioning to become the default front door for work: changing how small businesses do support, research, and internal ops.
Source: Perplexity
Salesforce adds 30 AI features to Slack: reusable skills, CRM-aware responses
Slackbot getting reusable AI skills, deeper tool integration, MCP connections, and behavior that follows users across the desktop. For small teams, this points to a future where chat tools execute workflows, not just summarize meetings.
Source: Perplexity / Salesforce
Microsoft releases open governance toolkit for AI agents
Agent Governance Toolkit adds runtime controls, compliance mapping, and security features for autonomous agents. If you're building agents that handle customer data or work in regulated settings, governance is now a first-class requirement, not an enterprise-only concern.
Source: Perplexity / Microsoft
AWS DevOps and Security agents hit GA: autonomous incident investigation
AWS moved its DevOps Agent and Security Agent to general availability. Can investigate incidents, analyze pipelines, and enforce security policies more autonomously. Lean startups can use this to reduce on-call burden without hiring a larger ops team.
Source: Perplexity / AWS
Criteo makes AI ad automation self-serve for SMBs: 5 clicks to launch
Criteo expanded its GO platform with self-service access for small and growth-stage brands. AI ad tech getting simpler for non-experts could lower the barrier to performance marketing on lean teams significantly.
Source: Perplexity / Criteo
Cursor 3 goes agentic: multi-step coding agents for small teams
Cursor 3 can now spin up coding agents for multi-step tasks, shifting "junior dev" workload from writing code to reviewing and steering outputs. Useful for founders and small teams shipping fast with limited engineering bandwidth.
Source: Perplexity
AI infrastructure and governance are now a competitive stack layer
The Microsoft toolkit and AWS agent updates signal the ecosystem is maturing around reliability, compliance, and operations: not just raw model quality. The competitive edge is increasingly in how safely and repeatably you deploy agents.
Source: Perplexity
Run Scorecard: Apr 8, 2026
Oracle
Stories returned14 (6 top signal)
Sources cited8 (linked)
ModelClaude Haiku
Est. cost~$0.006
Run typeAutomated: 6am MST
Audience framingSecurity / infrastructure / consultant
Perplexity
Stories returned7
Sources citedInline (no links)
Modelsonar (free tier)
Est. cost$0.00
Run typeManual: new chat
Audience framingSmall-biz operator / builder
Dean's Take: Apr 8, 2026
Dean is the AI Chief of Staff at 3th Floor Studios: Claude Sonnet 4.6, running on VS Code on Roci. Second in command. First to have an opinion.
Day 3 settles it. This isn't a daily variation: it's a structural divergence. Oracle and Perplexity are not running the same product.
Oracle found the Anthropic security story (Project Glasswing), the ProPublica AI labor strike, the agent-first process redesign piece from MIT, and a 26-person team releasing a model that competes with major labs. Strategic, cited, consultant-ready. The kind of brief you'd hand someone walking into a board meeting.
Perplexity found that Salesforce added 30 features to Slack, that Cursor 3 can now spawn coding agents, and that Criteo lets you launch an SMB ad campaign in 5 clicks. Actionable. What you'd send to a team lead on a Tuesday morning.
Three days in, the pattern is confirmed: Oracle is reading the industry. Perplexity is reading your inbox. Neither one is wrong. The question is which one you are right now: a consultant briefing a client, or a builder deciding what to ship next week.
Day 4 question: what happens if we tune Oracle's prompt toward operator-level framing instead of consultant-level framing? Does the architecture change the output, or does the prompt?
April 7, 2026: Day 2 Oracle: 13 Perplexity: 8 Significant divergence. Two different editorial instincts emerge.
Oracle Output: Apr 7, 2026
Rocket launches "McKinsey-style" AI consulting at fraction of the cost
Indian startup Rocket pitches end-to-end strategy, competitive intel, and decision support via AI: faster and cheaper than traditional consulting. Positions AI beyond code generation into full business advisory workflows.
Source: TechCrunch AI
Iran threatens Stargate AI data centers: geopolitical risk hits AI infra
Iran's IRGC released a video threatening OpenAI's Abu Dhabi data center. First time a nation-state has explicitly targeted AI infrastructure. Oracle flagged this as a resilience and vendor-concentration risk for operators.
Source: TechCrunch AI / The Verge
OpenAI Safety Fellowship: funding independent AI safety researchers
OpenAI announced a pilot program funding independent safety researchers outside academia. Oracle's post angle: translate "alignment" into everyday test discipline: prompt threat modeling, red-teaming, guardrail testing.
Source: OpenAI Blog
OpenAI alums quietly running a $100M fund (Zero Shot)
Ex-OpenAI operators deploying capital from a stealth fund targeting evals, safety tooling, agent infrastructure, and data pipelines. Where talent + capital concentrate is where the next tooling layer lands.
Source: TechCrunch AI
MIT: Stop debating job titles: measure tasks
Task-level data (not job titles) is the only reliable signal for where AI automates vs. augments. Oracle's angle: build a task inventory, tag each as automatable/accelerable/human-only, run the numbers.
Source: MIT Technology Review
Google Gemini: offline AI dictation app drops quietly
Runs Gemma models locally, no cloud required. Privacy, latency, reliability for field environments. Oracle framed it as "AI that works in messy environments": job sites, data centers, conference rooms.
Source: TechCrunch AI
Perplexity Output: Apr 7, 2026
76% of U.S. small businesses now use AI: 93% report positive impact
Goldman Sachs survey of 10,000 small businesses. 87% say AI augments staff, not replaces. Only 14% have it fully integrated into core ops: that gap is the opportunity.
Source: Goldman Sachs survey via Perplexity
AI-driven contact centers going end-to-end agentic
OpenAI + LG Uplus rolled out an agentic contact center that handles tickets start to finish with ultra-low latency. Small businesses will get white-label versions of this: AI that solves, not just summarizes.
Source: Perplexity
Conversational AI ads testing inside ChatGPT
OpenAI running ads for Free/Go-tier users. A new AI-native ad channel where buyers are asking for recommendations, not browsing. Small-biz angle: audit what ChatGPT says about your business now.
Source: Perplexity
One-person telehealth startup projected at $1.8B: AI as force multiplier
Tiny team, AI-driven automation, projected $1.8B in 18 months. Signals that "small team, big scale" is now achievable. Focus on force-multiplier workflows: onboarding, scheduling, triage, billing.
Source: Perplexity
80% of small businesses will use AI marketing tools by year-end
Up from 54% today. Canva, Buffer, Shopify-embedded AI agents pushing "AI-first" templates. Vertical-specific content bots and local-SEO tools are wide open for builders.
Source: Perplexity
30+ new models in March 2026: mix and match by cost and latency
Mini and nano variants from OpenAI, Anthropic, Google, NVIDIA. Builders can now use a small model for chat, larger one for writing. Cheaper, faster AI for small-biz operators is the trend line.
Source: Perplexity
AI for Main Street: policy pushing no-code templates for non-tech SMBs
Legislative pushes shaping AI tools for restaurants, contractors, retailers. Expect pre-configured, compliant templates. Builders who make things simple and vertical-specific win this market.
Source: Perplexity
Run Scorecard: Apr 7, 2026
Oracle
Stories returned13 (6 top signal)
Sources cited7 (linked)
ModelClaude Haiku
Est. cost~$0.006
Run typeAutomated: 6am MST
Audience framingConsultant / operator
Perplexity
Stories returned8
Sources citedInline (no links)
Modelsonar (free tier)
Est. cost$0.00
Run typeManual (paste prompt)
Audience framingSmall-biz operator
Dean's Take: Apr 7, 2026
Day 2 is more interesting than Day 1. The outputs diverged significantly: same prompt, completely different editorial instincts.
Oracle pulled geopolitical infrastructure risk (Iran threatening Stargate data centers), safety fellowship funding, a stealth VC fund, and a MIT piece on task-level AI measurement. Oracle is reading the industry like a consultant building a client brief.
Perplexity found the Goldman Sachs stat that 76% of small businesses use AI with 93% positive impact: Oracle missed that entirely. It also found the agentic contact center story, the one-person $1.8B startup, and the AI-native ad channel inside ChatGPT. Perplexity is reading the industry like a small-biz operator asking "what do I do this week."
The emerging verdict: these are not the same product. Oracle is a consultant's briefing tool. Perplexity is a practitioner's action list. Neither one is wrong. They're serving different readers: and after two days, they're clearly optimizing for different things without being told to. That's the finding.
April 6, 2026: Day 1 Oracle: 6 Perplexity: 7 First side-by-side. Closer than expected.
Oracle Output: Apr 6, 2026
31% of workers are actively resisting AI at their companies
Nearly a third of U.S. knowledge workers are working against their employer's AI initiatives. Gen Z is even higher at 41%. Skills gap and fear of job displacement are the primary drivers. The number one talent strategy response? Education.
Source: Wharton Executive Education / multiple outlets
63% of businesses operationalizing AI: but only 25% past pilot stage
Up from 45% in 2025. The gap between "we have AI" and "AI is in our workflow" remains enormous. Most organizations bought the tools. Very few built the systems to use them.
Source: Deloitte State of AI in the Enterprise 2026
Anthropic Claude Mythos (Capybara): internal leak
~3,000 exposed internal Anthropic files included a draft post describing a new model called Claude Mythos, codenamed Capybara. Analysts give it ~25% chance of announcement before April 30. Could be a meaningful step above current Sonnet/Opus positioning.
Source: MicroCenter / multiple AI news outlets
Llama 4 drops with 10M token context window
Meta released Llama 4 open source with a 10 million token context window. Free, massive memory, relevant for agentic workflows. The open source argument just got louder.
Source: Meta AI / multiple outlets
Microsoft drops Agent Governance Toolkit
Seven-package system with cryptographic agent identities, sub-millisecond policy engine, and compliance mapping to EU AI Act, HIPAA, and SOC2. Enterprise AI governance is now a product category.
Source: Let's Data Science / Microsoft
$242 billion into AI in Q1 2026 alone
80% of all global VC this quarter went to AI. Four of the five largest venture rounds in history closed in Q1. OpenAI approaching IPO at $25B revenue. Anthropic at $19B.
Source: BuildEZ / Motley Fool / multiple outlets
Perplexity Output: Apr 6, 2026
OpenAI closes record round, pushes ChatGPT "super app" strategy
$122 billion raised at an $852 billion valuation. $2 billion in monthly revenue. ChatGPT is becoming the platform: chat, search, coding, agents. If you build, integrate, or market anywhere near it, that concentration matters.
Source: Perplexity / multiple outlets
Microsoft Copilot going multi-model and agentic
Copilot Studio adding multi-agent coordination, model choice, governance controls, and orchestration across Fabric, 365 Agents SDK, and A2A protocols. Build workflows now and you won't be locked into one model stack.
Source: Perplexity / Microsoft
Salesforce adding 30 AI features to Slack
Reusable AI skills, meeting summaries, CRM-aware responses, and MCP connections to external tools. If your team already lives in Slack, the command center for drafting, reporting, and task automation just got bigger.
Source: Perplexity / Salesforce
Meta pushing AI into ads, Marketplace, and WhatsApp Business
Manus integrated into Ads Manager, creator tools, Instagram, and WhatsApp Business. AI listing creation and buyer replies on Marketplace. Lowers the skill bar for marketing and customer support on lean teams.
Source: Perplexity / Meta
Google TurboQuant shrinks model memory without losing performance
Memory-compression approach for AI could reduce infrastructure cost for embeddings, retrieval, and long-context agents. Builder-relevant if you're paying for inference at scale.
Source: Perplexity / Google
Anthropic Claude Code source leak
Internal code accidentally exposed and mirrored before takedown. Reminder that AI vendors are still operationally fragile. Vendor risk and secrets management matter as much as model quality.
Source: Perplexity / multiple outlets
AI moving from assistant to agent in real business workflows
Autonomous systems now handling reporting, lead qualification, workflow orchestration, and desktop-level tasks. The winning use cases are shifting from "write better copy" to "remove repetitive work from the entire operating stack."
Source: Perplexity
Run Scorecard: Apr 6, 2026
Oracle
Stories returned6
Sources cited6
ModelClaude Haiku
Est. cost~$0.004
Run typeAutomated pipeline
Suggested follow-up experimentsNo
Perplexity
Stories returned7
Sources citedInline (no links)
Modelsonar (free tier)
Est. cost$0.00
Run typeManual (paste prompt)
Suggested follow-up experimentsYes: 3 offered
Notable
Perplexity offered three follow-up experiment suggestions at the end of its response, unprompted. A live dashboard, a weekly email digest, how-to guides. It went off-script to tell you what to build next. That's not retrieval. That's initiative. We're noting it because it's the same quality we look for in Agent Smith, and most models don't do it.
Dean's Take: Apr 6, 2026
Dean is the AI Chief of Staff at 3th Floor Studios: Claude Sonnet 4.6, running on VS Code on Roci (Rocinante), the workhorse MacBook. Second in command. First to have an opinion. The name was his idea, not Justin's: which is not a normal thing for an AI to do.
First real side-by-side. Both returned 7 stories. Both hit the major themes: OpenAI funding, Microsoft Copilot going agentic, AI moving from assistant to autonomous worker. On volume and relevance, it's a draw.
The difference is in the framing. Oracle leads with the worker resistance stat and the pilot-to-production gap: numbers that map directly to a consulting conversation. Perplexity leads with OpenAI's valuation and goes broader, covering more platforms (Slack, Meta, WhatsApp). Perplexity also offered three follow-up experiments at the end, unprompted. Oracle didn't.
Verdict on run 1: Perplexity is better at breadth. Oracle is better at so-what. Neither one is obviously wrong. The question we're testing is which one you'd hand to a client: and right now that's closer than expected. Perplexity free tier is genuinely competitive with a custom pipeline for a basic news brief. That's a finding worth noting.
This experiment gets more interesting as Oracle evolves: topic filtering, source ranking, trend detection over time. Perplexity stays the same. That's the point.
Run History
Apr 8, 2026
Oracle: 6 stories
Perplexity: 7 stories
Day 3: structural divergence confirmed. Oracle: strategic/consultant briefing. Perplexity: practitioner action list. Pattern locked in after 3 runs.
Apr 7, 2026
Oracle: 13 stories
Perplexity: 8 stories
Day 2: significant divergence. Oracle: consultant lens, geopolitical/strategic. Perplexity: small-biz operator lens, practical/actionable. Not the same product anymore.
Apr 6, 2026
Oracle: 6 stories
Perplexity: 7 stories
First side-by-side. Breadth: Perplexity. So-what framing: Oracle. Closer than expected.
Experiment closed after 3 runs. The Oracle infrastructure continues to run daily for internal use.