{ "title": "AI Builder Weekly: April 15–22, 2026", "summary": "Self-hosted productivity stacks for agents, Kimi K2.6 vs GPT-5.4, GPT-6 Spud watch, Claude Opus 4.7 + Design, Qwen 3.6 on a single GPU, RunPod economics, and the week's biggest AI stories.", "slides": [ { "segment": "Opening", "duration": "1 min", "title": "AI Builder Weekly", "bodyHtml": "

April 15–22, 2026

Self-Hosted Stacks · Kimi K2.6 · GPT-6 Watch · Claude Design · Qwen 3.6 · RunPod Math

", "notes": "Welcome slide. Dense week — Claude Opus 4.7 dropped, Claude Design launched, Kimi K2.6 went GA, Qwen 3.6 Max-Preview landed, OpenAI put CPC ads in ChatGPT, and Anthropic locked in $5B more from Amazon.", "media": { "type": "svg", "inlineSvg": "FINALDIFFWeekly Briefing · April 22, 2026BETTER BUILDERS · FEWER BARRIERS · FINAL DIFF", "caption": "FinalDiff — builders helping builders ship." } }, { "segment": "Opening", "duration": "1 min", "title": "What We're Covering", "bullets": [ { "title": "Agent Productivity Stack", "text": "Mautic + Twenty + Listmonk + CLI — the self-hosted SaaS killer" }, { "title": "One VPS To Rule Them All", "text": "Docker Compose everything on a $20/mo Hetzner box" }, { "title": "Kimi K2.6", "text": "Open-weight 1T MoE, 32B active — cheaper than GPT-5.4?" }, { "title": "GPT-6 \"Spud\" Watch", "text": "Pre-training done March 24. Still waiting. Is the free OpenRouter model it?" }, { "title": "Claude Opus 4.7 + Claude Design", "text": "SWE-bench 87.6%, prototyping tool that spooked Figma" }, { "title": "Qwen 3.6", "text": "New king for a single 32GB GPU at Q4?" }, { "title": "RunPod Math", "text": "1× GPU for Qwen 3.6 vs paying subscriptions" }, { "title": "Week in AI", "text": "OpenAI ads, Anthropic $5B AWS deal, Snap cuts 25%, and more" } ], "notes": "Dense agenda. Move fast through the tactical slides, leave room for discussion on the RunPod economics and self-hosted stack sections — those are the most actionable for our audience." }, { "segment": "Agent Productivity Stack", "duration": "4 min", "title": "Build Your Agent's Productivity Stack", "bodyHtml": "

Stop paying per-seat SaaS fees. Three open-source tools give your AI agent (or your small team) a complete CRM + marketing automation + email stack — all self-hosted, all CLI-operable.

TWENTY CRM

Open-source CRM replacing Salesforce/HubSpot. GraphQL API, full contact/deal/pipeline management. Your agent can create contacts, update deals, log activities — all via API. Apache 2.0, 25K+ GitHub stars.

MAUTIC

Marketing automation — visual workflow builder, lead scoring, landing pages, behavioral triggers. HubSpot/ActiveCampaign alternative. Needs 2–4GB RAM. Your agent builds nurture sequences and manages segments via REST API.

LISTMONK

High-performance newsletter manager. Written in Go, runs on 512MB RAM, handles millions of subscribers. No automation — just fast, reliable email sending. Agent sends campaigns via simple API calls.

", "notes": "The key insight: each tool has a clean REST or GraphQL API that an agent can call. Twenty for contact data, Mautic for automation logic, Listmonk for actual email delivery. The CLI layer (curl, httpie, or a simple Python wrapper) ties them together. Total SaaS replacement cost saved: $500-2000/month for a small team.", "media": { "type": "svg", "inlineSvg": "THE SELF-HOSTED AGENT STACKAI AGENT / CLITWENTYCRM · GraphQL APIMAUTICAutomation · REST APILISTMONKEmail · REST APIPostgreSQLReplaces: Salesforce + HubSpot + Mailchimp → saves $500–2,000/mo", "caption": "All three share a PostgreSQL backend. One database to rule them all." } }, { "segment": "Agent Productivity Stack", "duration": "3 min", "title": "One VPS — Everything in Docker Compose", "bodyHtml": "

A single Hetzner CAX21 (4 vCPU ARM, 8GB RAM, 80GB SSD) at ~€7.50/month runs the entire stack comfortably. Add Coolify or Dokku as a git-push deployment layer, Tailscale for secure access, and Cloudflare Tunnels for public endpoints.

docker compose up -d postgres twenty mautic listmonk
tailscale up --hostname=agent-stack
coolify deploy --from=github.com/you/stack

RAM budget: PostgreSQL ~1GB, Twenty ~1.5GB, Mautic ~2GB, Listmonk ~256MB, overhead ~1.5GB. Fits comfortably in 8GB. Scale up to the CAX31 (16GB, ~€14/mo) if adding n8n or Grafana.

", "bullets": [ { "title": "CLI orchestration", "text": "httpie or curl scripts wrap each tool's API. Your agent calls the CLI, CLI calls the APIs." }, { "title": "Backup strategy", "text": "pg_dump nightly to a Hetzner Storage Box (1TB for €3.81/mo). One cron job, all three apps." }, { "title": "Total cost", "text": "~€12/mo for everything. Compare to HubSpot Starter at $50/mo with 1/10th the features." } ], "notes": "Audience action item: the docker-compose.yml for this entire stack is ~80 lines. We'll share it in the FinalDiff Matrix channel after the session. Key insight: since all three tools use PostgreSQL, you can have ONE backup strategy for everything. Hetzner CAX (ARM) boxes are cheaper than x86 equivalents and all three tools run fine on ARM." }, { "segment": "Kimi K2.6", "duration": "4 min", "title": "Kimi K2.6 — Open-Weight Frontier for Agents", "bodyHtml": "

Moonshot AI shipped Kimi K2.6 on April 20 — a 1T-parameter MoE with 32B active, 256K context, native multimodality, and open weights from day one. The fastest preview-to-GA transition in the K2 series.

The headline isn't a single benchmark — it's duration, breadth, and coordination. K2.6 ran a 13-hour autonomous coding session, iterated through 12 optimization strategies, made 1,000+ tool calls, and modified 4,000+ lines of code on an 8-year-old financial matching engine — extracting a 185% throughput improvement.

Agent Swarm scales to 300 sub-agents across 4,000 coordinated steps (up from K2.5's 100/1,500). Day-0 support in vLLM, OpenRouter, Cloudflare Workers AI, Hermes Agent, and OpenCode.

", "notes": "Is it cheaper than GPT-5.4? Kimi K2.6 API: $0.60-$0.95/M input, $2.50-$4.00/M output depending on provider. GPT-5.4: ~$2.50/$10.00. So yes, roughly 3-4× cheaper on input, 2.5-4× cheaper on output. On SWE-bench Pro: 58.6 (Kimi) vs 57.7 (GPT-5.4). It's genuinely competitive at a fraction of the cost. But it's verbose — Artificial Analysis spent $920 evaluating it because it generated 160M tokens vs the 41M average.", "media": { "type": "html", "html": "

SWE-BENCH PRO

58.6%

vs GPT-5.4: 57.7%

SWE-BENCH VERIFIED

80.2%

open-weight SOTA

API COST

~3–4×

cheaper than GPT-5.4

1T params · 32B active · 256K context · 300 sub-agents · 13-hour sessions · Open weights

", "caption": "Modified MIT license — commercial use needs written auth for 100M+ MAU." } }, { "segment": "GPT-6 Watch", "duration": "3 min", "title": "GPT-6 \"Spud\" — Still Cooking", "bodyHtml": "

Pre-training completed March 24 at Stargate (Abilene, TX). Sam Altman said \"a few weeks.\" The April 14 rumor came and went. We're still waiting.

Polymarket odds have drifted from 78% to ~72% for an April 30 launch. Most defensible window is now late April to mid-May 2026. Rumored specs: 40% improvement over GPT-5.4, 2M token context. Nothing confirmed.

Meanwhile, every major competitor shipped this week: Claude Opus 4.7, Kimi K2.6, Qwen 3.6-Max. The longer Spud sits, the more the \"frontier model\" narrative gets eaten. HumanX conference last week: the dominant story was \"Claude Mania,\" not OpenAI.

", "notes": "Is GPT-6 the free model on OpenRouter? Almost certainly not — that's likely GPT-5.4 in a promotional tier, not Spud. OpenAI has given free OpenRouter access to older models before to drive developer adoption. The real GPT-6 will be a paid API product. Don't plan around free access to frontier models.", "media": { "type": "html", "html": "

Timeline So Far

Mar 24 Pre-training done at Stargate
Apr 7 Unverified source claims Apr 14 launch
Apr 10 Employees post cryptic teasers
Apr 14 Nothing happens. Silence.
Apr 22 Still waiting. Polymarket: 72%

Free model on OpenRouter ≠ GPT-6. It's likely GPT-5.4 in a promo tier.

", "caption": "Don't plan around free access to frontier models." } }, { "segment": "Claude Opus 4.7", "duration": "3 min", "title": "Claude Opus 4.7 — Shipped April 16", "bodyHtml": "

Anthropic's new flagship. SWE-bench Verified jumped to 87.6% (from 80.8%), GPQA Diamond hit 94.2%. Better at long-running agentic work, substantially better vision (2,576px), and more creative on professional tasks.

New features: xhigh effort level, task budgets (token countdown for agent loops), /ultrareview (multi-agent code review), Auto mode for Max subscribers. Hex reports: \"low-effort Opus 4.7 ≈ medium-effort Opus 4.6.\"

Caveat: New tokenizer uses up to 35% more tokens for the same text. Same $5/$25 pricing, but effective cost goes up. Thinking content hidden by default — set display: \"summarized\" to restore.

", "notes": "Less broadly capable than Mythos Preview (the model they didn't release). Cyber capabilities deliberately reduced during training. Security professionals can apply to a new Cyber Verification Program. Breaking API changes: extended thinking budgets removed, temperature/top_p/top_k setting returns 400 error. Migration needed.", "media": { "type": "html", "html": "

SWE-BENCH VERIFIED

87.6%

from 80.8% (Opus 4.6)

GPQA DIAMOND

94.2%

reasoning benchmark

CURSORBENCH

70%

from 58% (Opus 4.6)

TOKENIZER TAX

+35%

more tokens, same price

", "caption": "Same $5/$25 pricing. But replay your traffic — effective cost may rise." } }, { "segment": "Claude Design", "duration": "4 min", "title": "Claude Design — Rapid Prototyping That Spooked Figma", "bodyHtml": "

Launched April 17 under Anthropic Labs. Not an image generator — a visual prototyping assistant powered by Opus 4.7. Describe what you need, Claude builds a working first version. Refine through conversation, inline comments, direct edits, or AI-generated sliders.

Why it matters for builders: It reads your codebase and design files to build a custom design system. Every project after that uses your real components automatically. Export as PDF, PPTX, HTML, or send directly to Canva. Hand off to Claude Code for implementation.

Brilliant (the learning app) reported pages requiring 20+ prompts in other tools needed only 2 prompts in Claude Design. Figma stock dropped 7% on the announcement tweet alone. Datadog compressed week-long design cycles into single conversations.

", "notes": "Available in research preview for Pro, Max, Team, Enterprise. Separate weekly usage limits — doesn't count against chat or Claude Code quotas. Not competing with Figma for professional designers (yet). Competing for the much larger group of people who were never going to use Figma: founders, PMs, marketers. The Canva partnership is smart — build rough in Claude Design, polish in Canva.", "media": { "type": "html", "html": "

What It Does

Interactive prototypes
Pitch decks & presentations
Landing pages & marketing assets
Wireframes & mockups
Internal tools & dashboards

Export Options

PDF for sharing
PPTX for presentations
Standalone HTML
Shareable internal URL
Direct to Canva (editable)

", "caption": "\"A product that doesn't exist yet just vaporized billions.\" — @aakashgupta" } }, { "segment": "Qwen 3.6", "duration": "4 min", "title": "Qwen 3.6 — New King for a Single 32GB GPU?", "bodyHtml": "

Alibaba shipped a whole family this month. The open-source Qwen 3.6-35B-A3B (April 16, Apache 2.0) is the one that matters for self-hosters: 35B total, only 3B active per token. At Q4 quantization, it fits in ~22GB VRAM.

It scores 73.4% on SWE-bench Verified — competitive with models 10× its effective size. 1M native context via the Plus variant. The proprietary Max-Preview (April 20) topped six coding benchmarks and beat Claude 4.5 on agentic skills.

The full Qwen 3.6 lineup: Max-Preview (proprietary, frontier tier), Plus (balanced, 1M context, was free on OpenRouter), Flash (speed-first), and 35B-A3B (open-source, Apache 2.0).

", "notes": "The question from the slide title: yes, Qwen 3.6-35B-A3B at Q4 is likely the best model you can run on a single 32GB GPU right now. It beats Gemma 4 26B-MoE on coding tasks while being similarly efficient. The 3B active parameter count means inference is fast even without high-end hardware. On an RTX 3090 or 4090 (24GB), you'll want Q3 or Q4_K_S quantization — tight fit but workable.", "media": { "type": "html", "html": "

35B-A3B (OPEN)

73.4%

SWE-bench Verified
~22GB VRAM at Q4
Apache 2.0

MAX-PREVIEW (PROP.)

#1

on 6 coding benchmarks
256K context
Beats Claude 4.5

35B total · 3B active · MoE · Fits a single consumer GPU at Q4

", "caption": "For self-hosters: Qwen 3.6-35B-A3B may be the new sweet spot." } }, { "segment": "RunPod Economics", "duration": "4 min", "title": "The Math: RunPod GPU vs Paying Subscriptions", "bodyHtml": "

Running Qwen 3.6-35B-A3B at Q4 needs ~22GB VRAM. An A100 80GB on RunPod Community Cloud: ~$1.89/hr. That gives you massive headroom for the model plus inference serving.

If you run it 8 hours/day for a month: ~$454/mo. 24/7: ~$1,360/mo. Compare: Claude Max ($200/mo) + ChatGPT Pro ($100/mo) = $300/mo but with usage caps and API restrictions.

The sweet spot: RunPod on-demand for heavy coding sessions (spin up, do work, spin down), API access for lightweight tasks (Kimi K2.6 at $0.60/M input), local 9B model for simple queries. Total: $50–150/mo with no caps and full data control.

", "notes": "The real comparison isn't RunPod 24/7 vs subscriptions — nobody should run RunPod 24/7 for inference. The right pattern is: (1) Local small model (Qwen 3.5-9B, 6GB VRAM, free) for simple queries, (2) Budget API (Kimi K2.6 or MiniMax M2.5 free tier) for medium tasks, (3) RunPod A100 spin-up for heavy agentic coding sessions, (4) Claude Max or Opus API for tasks that genuinely need frontier quality. This hybrid approach costs $50-150/mo depending on intensity.", "media": { "type": "html", "html": "

RunPod A100 80GB

$1.89

/hr community

$454

/mo @ 8hr/day

Subscriptions

$200

Claude Max

$100

ChatGPT Pro

HYBRID SWEET SPOT

Local 9B + budget API + on-demand RunPod = $50–150/mo with no caps

", "caption": "Nobody should run RunPod 24/7. Spin up, do work, spin down." } }, { "segment": "Week in AI", "duration": "3 min", "title": "OpenAI Puts CPC Ads in ChatGPT", "bodyHtml": "

Confirmed April 21: OpenAI turned on cost-per-click ads inside ChatGPT. Advertisers can now set bids between $3 and $5 per click, in addition to existing impression-based (CPM) ads launched in February.

OpenAI targets $2.4B in ad revenue for 2026 and $11B for 2027. The ad pilot already topped $100M annualized in under two months. CPMs have dropped from $60 at launch to ~$25 now.

For builders: this is the clearest signal yet that free-tier AI will be ad-supported. OpenAI pledges ads won't influence answers, but they're also losing $57B in 2026. The pressure to monetize is existential. Anthropic's Claude products remain ad-free.

", "notes": "Context for the audience: OpenAI lost $8B in 2024, $25B in 2025, $57B projected in 2026. Ad revenue even at $2.4B barely dents that. They're also hiring their first advertising marketing science leader. The ad formats are for Free and Go ($8/mo) tier users only — Plus and Pro subscribers don't see ads. Yet." }, { "segment": "Week in AI", "duration": "2 min", "title": "Anthropic Locks In $5B More from Amazon", "bodyHtml": "

Anthropic announced it has secured up to 5 GW of compute from Amazon, with an additional $5B investment now and up to $20B more later. This is a massive signal on frontier-model capex.

Anthropic's revenue more than doubled to a $9B run rate by year-end 2025. The 2026 forecast was raised 20% to $18B. Claude Code alone is at $2.5B ARR.

They also started requiring government-issued photo IDs and selfies from some users to prevent access from US adversaries (China, Russia, North Korea). The auth crackdown continues.

", "notes": "The compute deal is separate from the investment. Anthropic is essentially pre-purchasing AWS capacity at scale to ensure they have the GPUs for Mythos-class model training. IPO expected later this year alongside OpenAI's. The ID verification policy is notable — this is Anthropic directly responding to the distillation attacks from Chinese labs we covered in the last briefing." }, { "segment": "Week in AI", "duration": "2 min", "title": "Snap Cuts 25%, AI Writes 65% of Code", "bodyHtml": "

Snap CEO Evan Spiegel laid off ~1,000 employees and closed 300+ open roles — roughly a quarter of planned headcount — citing \"rapid advancements in AI\" that let smaller teams do the same work.

The bombshell detail: AI now generates over 65% of Snap's new code. The restructuring expects $500M+ in annualized cost savings by H2 2026. Snap stock rose 11% on the announcement.

Connects to the Stanford AI Index finding: software developer employment for ages 22-25 dropped ~20% since 2022. The \"AI replaces junior developers\" narrative is becoming measurable.

", "notes": "This is the most concrete example yet of a major tech company explicitly attributing layoffs to AI productivity gains. Not 'we're restructuring' — specifically 'AI lets smaller teams do the same output.' Worth discussing with the audience: what does this mean for small businesses that compete for talent with companies like Snap?" }, { "segment": "Week in AI", "duration": "2 min", "title": "More Headlines This Week", "bodyHtml": "

ALLBIRDS → AI COMPANY

The sneaker company pivoted to a GPU rental business called \"NewBird AI.\" Stock up 600%. They're stripping their \"public benefit\" status. From $4B IPO to $22M market cap to GPU-as-a-Service. Peak 2026.

YOUTUBE DEEPFAKE DETECTION

Made its deepfake detection tool available to anyone at high risk of having their likeness abused. Previously limited to public officials.

OPENAI + NOVO NORDISK

Strategic partnership to accelerate drug discovery and integrate AI across global operations by end of 2026.

QUANTUM + AI

Researchers showed blending quantum computing with AI dramatically improves predictions of complex chaotic systems.

AI LAWYER SUSPENDED

Nebraska attorney suspended — 57 of 63 citations were defective, including 20 AI hallucinations. US courts imposed $145K+ in AI citation sanctions in Q1 2026.

", "notes": "The Allbirds story is the perfect absurdity of the moment. A failed shoe company pivoting to GPU rental. The AI lawyer sanctions are a cautionary tale worth mentioning — always verify AI-generated citations." }, { "segment": "Week in AI", "duration": "2 min", "title": "Stanford AI Index — The Numbers That Matter", "bodyHtml": "

The 423-page 2026 AI Index landed April 13. Key findings most coverage missed:

US lead is narrower than you think. US and Chinese models have traded the top performance position multiple times since early 2025. The gap is 2.7% as of March 2026. China leads in publications, citations, and patent grants.

Safety benchmarks are mostly empty. Only Claude Opus 4.5 reports on more than two responsible AI benchmarks. Most frontier models report nothing on fairness, security, and human agency.

Humanity's Last Exam: Best model accuracy went from 8.8% to 50%+ in one year. SWE-bench: 60% → nearly 100%. These benchmarks are expiring faster than they're being replaced.

", "notes": "The safety gap is the most underreported finding. Capability benchmarks are reported consistently. Safety benchmarks are largely absent. Only one company (Anthropic) reports on more than two responsible AI benchmarks. Worth asking the audience: should builders care about safety benchmarks, or is that someone else's problem?" }, { "segment": "Takeaways", "duration": "3 min", "title": "What To Do This Week", "bodyHtml": "

TRY CLAUDE DESIGN

If you have a Pro/Max/Team plan, Claude Design is live now. Build a prototype for your next feature idea in 10 minutes. Export to Canva or PPTX. The 2-prompt-vs-20 claim from Brilliant is real.

SPIN UP THE PRODUCTIVITY STACK

docker-compose.yml with Twenty + Mautic + Listmonk on a Hetzner CAX21. ~€12/mo replaces $500+/mo in SaaS. We'll share the compose file in Matrix.

TEST KIMI K2.6 ON OPENROUTER

Swap in moonshotai/kimi-k2.6 on your existing OpenRouter setup. Compare against whatever you're paying for now. 3-4× cheaper than GPT-5.4 at competitive quality.

RUN QWEN 3.6 LOCALLY

If you have 24GB+ VRAM: ollama run qwen3.6-35b-a3b. If not, qwen3.5-9b at Q4 (~6GB) is still the best daily-driver for laptops.

WATCH THE ADS

OpenAI CPC ads in ChatGPT are live. If you're using the free tier, you're the product now. Consider whether $20/mo for ad-free access is worth it — or switch to ad-free alternatives.

", "notes": "Action-oriented closing. Every item is something someone can do this week. The compose file for the productivity stack will be the most requested resource — have it ready to drop in Matrix right after the session." }, { "segment": "Discussion", "duration": "10 min", "title": "Breakout Discussion", "bodyHtml": "

Groups of 3–4 · Pick 2–3 questions · 10 minutes

1

Would you trust an AI agent to manage your CRM and send emails on your behalf? Where do you draw the line on automation?

2

Claude Design dropped and Figma stock tanked. Is AI prototyping a genuine threat to design tools, or is this overhyped?

3

Snap says AI writes 65% of their code and cut 25% of staff. Is this the future for every company, or a Snap-specific story?

4

RunPod vs subscriptions vs local hardware — what's your current setup and what would you change after today?

5

OpenAI put ads in ChatGPT. Anthropic is ad-free. Does this affect which platform you build on?

", "notes": "Standard FinalDiff format: breakout groups of 3-4, pick 2-3 questions, 10 minutes. Question 1 ties directly to the productivity stack slides. Question 3 is deliberately provocative — expect strong opinions. Reconvene and share highlights." } ] }