2026-04-24
Apr 23 12:00Z → Apr 24
- OpenAI GPT-5.5 (Apr 23): First OAI flagship explicitly framed as 'agent runtime' not chat model. 88.7% SWE-bench, 60% fewer hallucinations vs 5.4. Three variants ($5/$30/M tokens). Released 6 weeks after 5.4 — model cadence now faster than enterprise eval cycles.
- ServiceNow -17.7% on record earnings miss (Apr 23): Now Assist grew 130% YoY, CEO raised AI forecast 50% to $1.5B — stock still fell 17.7% on gross margin compression (81.5% vs 82.1% expected). Fortune: 'The numbers are good, but the vibes are bad.' Salesforce, Workday, Oracle dragged with it. Market treating any SaaS weakness as AI disruption referendum.
- DeepSeek V4 Flash + Pro (Apr 24): Open-source, 1M context window, $0.14/M input tokens (35-200x cheaper than frontier). Claims near-frontier on reasoning, lags 3-6 months on knowledge. Tencent + Alibaba in talks to invest — first external funding round for DeepSeek.
- Sierra acquires Fragment (YC, French, Apr 23): 3rd acquisition in 2 months — workflow integration + European expansion. Consolidation pattern: well-funded AI sales-led leaders buying capability vs. building.
- Google Cloud $750M partner fund (Apr 22, Cloud Next '26): Embeds Google FDEs at Accenture, Capgemini, Cognizant, TCS for agentic AI deployment. SIs are the enterprise AI delivery layer — mirrors OpenAI/Accenture+Infosys+PwC pattern from run-11.
- Builder Demo Radar — Karpathy AutoResearch still viral (circulating Apr 23): 66K+ GitHub stars, Greg Isenberg framing for GTM: 'give it a goal like lower customer acquisition cost — then it runs.' Qualifies: agentic, unexpected GTM use case, Claude Code runtime, SF builder community signal.
ServiceNow's Now Assist grew 130% YoY. Stock fell 18%. The market doesn't believe the AI story even when the numbers are real. The new SaaS question isn't 'do you have AI?' It's 'does AI drive net new revenue — or just cannibalize your seats?'
Karpathy's AutoResearch ran 700 experiments in 2 days and found 20 optimizations that actually worked. No sleep. No status quo bias. No missed windows. The bottleneck in ML research was never compute — it was human availability.
228/280 ✓The same thing is true in performance marketing. We already knew what to test. The bottleneck was how fast we could run and analyze. An agent running 24/7 doesn't change what's possible — it changes how fast you get there.
222/280 ✓Greg Isenberg framed it best: 'give it a goal like lower customer acquisition cost.' The goal function IS the skill now. If your objective is clear, an agent can run 100 experiments overnight to find the path.
209/280 ✓DeepSeek V4 Flash: $0.14/M input tokens. Frontier models: $5-30/M. That's 35-200x cheaper with near-comparable performance. 'Which model does this product run on?' just became a real procurement question every AI vendor will face in 2026.