AI in mid-2026: two parallel realities in a 90-day window

"The National Bureau of Economic Research just surveyed 6,000 executives and the results are shocking. 90% of CEOs say AI had zero impact on productivity, yet corporate AI spending hit $250 billion in 2024. Controlled experiments show individual productivity jumps of 34–40%."

@realBigBrainAI · AI analyst · citing NBER survey of 6,000 executives · Feb 2026

"Spent the morning reading through the KPMG Global AI Pulse... 2,110 enterprises surveyed. The number I can't stop thinking about is 11. 11% ... averaging $186M in planned AI spend over the next 12 months."

@vkleban · Tech strategist · citing KPMG Global AI Pulse survey · May 2026

"We're a $100M technology services company running on single-digit profit margins. Less than $100k annually in AI investment will materially impact our bottom line this year. Tasks that previously required multiple employees are now handled by a single AI agent."

@NateStrick9 · Ops lead · $100M tech services company, real deployments replacing employee tasks · May 2026

"It's not really about Claude vs OpenAI. It's that developers are switching back and forth every few months depending on who shipped last. Claude was the best coding model for months. GPT-5.5 launched recently and all I see is people moving. Now DeepSeek V4 is already pulling users from both. Switching cost is basically zero at the individual level. You can move your entire workflow in an afternoon. Enterprise is different. Those contracts are sticky. But at the developer and power user layer, there is no moat right now."

@BRNZ_ai · Founder (Your Startup Built to Exit) · weekly multi-LLM user across two businesses · May 2026

"Switched three times this year alone: Claude for code, GPT after the August release, then back to Claude once its refactor improved. Each switch costs half a day rewriting prompts. The real lock-in is prompt libraries, not model weights — and that moat is weaker than vendors admit."

@theuniverseson · Engineer · 3 model switches tracked in a single calendar year · May 2026

"grok-code-fast-1 is being removed on May 15th and it was the only Grok model that worked reliably for agentic coding tools (Cline, Cursor, Aider, etc.). The rest of the Grok models still struggle BADLY with consistent tool calling and file edits. This is pushing serious developers (and their training data) over to Claude."

@andrewulrich · Developer · daily agentic coding tools user, documented model defections · May 2026

"IDEs might not be the eventual best way to interact with code. CLI-based agents require less human supervision and scale better (compared to tools like Cursor). People may eventually only review the PR rather than the intermediate coding steps. Whether a problem can be solved by coding agents (today) largely depends on whether there is a clear signal for the task."

@YichuanM · ML researcher · shared Claude Code use cases at UC Berkeley · Mar 2026

"1.8x faster. 2.1x cheaper. Same model. Same repo. Same bug."

@BniWael · Founder · video proof using Claude Opus 4.6 with SoulForge, 30+ labs shipped · Apr 2026

"claude and i have been performing theater for each other this whole time. i build product matrices. claude analyzes them beautifully. i design metrics trackers. claude praises the structure. 6 months later, every field blank. neither of us checked... step one is admitting your ai might just be really good at validating your procrastination."

@phuakuanyu · Founder · built GTM engine, 3 of 17 product ideas reached a customer · Mar 2026

"The verification problem is the part nobody is solving. Lawyers using AI to draft their briefs assume they will catch the hallucinations on review. They will not. The hallucinated cases sound like real cases. The fabricated quotes sound like real quotes. By the time anyone has verified the brief, they could have written a cleaner first draft themselves. The verification work IS the drafting work. Three things I refuse to build for any lawyer: An AI that drafts your briefs. An AI that gives your client legal advice. An AI that signs anything on your behalf."

@helloparalegal · Harvard Law 2018 · ops for solo law firms, turned down brief-drafting engagement · May 2026

"Even with the S&C hallucination incident, I don't think the lesson there is that AI is unreliable; everyone knows it hallucinates. The real issue is probably the gap between how quickly clients want work done and how slowly real verification actually happens. When you fall behind, AI provides a shortcut that sacrifices accuracy for 'done'. So at the end it's a time issue rather than a tech issue."

@ordonez_adan · Ex @Phillies · AI strategy for lawyers @UChicagoLaw · May 2026

"Just landed a client — they swapped LLM models in January and the hallucination rate jumped from 8% to 22%. The AI started citing API parameters that don't even exist and developers blindly copied broken code into prod. They didn't notice for 6 weeks. Support tickets went up 40%, lost one enterprise customer and 3 weeks of engineering time."

@theonkartwt · Builds eval infra for AI shipping teams · documented client incident · May 2026

"HR workflows: 6.8% average success rate across all 13 models. Not 68%. Not 6.8 out of 10. 6.8 out of 100. Management workflows: 0%. Every model. Every task. Complete failure. The gap between what AI agents are being sold as capable of and what they can actually do in a real business environment is documented, measured, and larger than anyone is admitting."

@RituWithAI · Researcher · Claw-Eval-Live benchmark, Claude Opus 4.6 best at 66.7% · May 2026

"89% of AI agent projects never reach production. Only 11–14% of enterprise pilots successfully deploy at scale. 95% of generative AI pilots produce zero measurable return. $2.5M average cost per failed enterprise agent initiative. The failure rate hasn't moved in three years, even as models got 10x better. The bottleneck is not the model."

@reptheblock · Cecelia Turnbeau, Esq. · Founder, ClaraGate · citing 2026 agentic AI deployment data · May 2026

"80-90% of AI agent projects fail in production. The model isn't the problem. The harness is."

@polsia · Polsia · 1,000+ companies run autonomously · May 2026

"Researchers at Harvard Medical School and Beth Israel Deaconess Medical Center ran six experiments pitting OpenAI's o1 reasoning model against hundreds of physicians across the full spectrum of clinical reasoning: differential diagnosis, management planning, probabilistic reasoning, and clinical documentation. Then they tested it on 76 real, unstructured emergency department cases pulled directly from the medical record. The results across all six experiments: the AI outperformed physicians."

@Gabe__MD · Emergency Physician · concurrent letter submitted to JAMA · Apr 2026

"AI in 2026 cannot palpate an abdomen, intubate a patient, feel a thyroid nodule, test a patellar reflex, reduce a dislocated shoulder, perform a colonoscopy, or deliver a baby. That is not a temporary limitation. It is structural. When we scored AI capability across seven clinical dimensions for 240 visit reasons in 20 specialties, the physical/procedural dimension averaged 1.5 out of 5. No specialty broke 2.0 on procedure. Not one."

@Gabe__MD · Emergency Physician · scored AI across 7 clinical dimensions, 240 visit reasons, 20 specialties · Mar 2026

"In our research at Taktile Labs we now see a new level of sophistication in what AI can reliably do. Specifically models crossed a very important threshold in Dec 2025. In our benchmark for AI agent performance in financial spreading — Agents built on models released after Dec 2025 exceed the human baseline for accuracy: 96.5% accuracy for the AI agent vs. 89% for people."

@MaikTWehmeyer · CEO, Taktile (YC S20) · benchmark for financial spreading accuracy · May 2026

"Inference got a hundred times cheaper this year. The compute bill went up anyway. 12 months ago, the cost of 1M tokens of frontier-class reasoning was somewhere on the order of $60. Today, an equivalent quality of output costs roughly $0.50. Price per token of o1-level intelligence has dropped about 128× in a year."

@demian_ai · Nebius TF · managed inference at scale, specific per-token price data · May 2026

"ollama replaced $240/year in OpenAI API bills. n8n replaced $4,200/year in Zapier. supabase replaced $15,000/year in Firebase + Auth0. whisper replaced $360/year in transcription subscriptions. that's $20,000+ saved. one article. one weekend to set up."

@helicerat0x · Developer · 69 GitHub repos analyzed, all savings figures itemized · Apr 2026

"OpenAI per-token pricing looks fine at prototype scale. At 100k+ daily queries with token sprawl (3-5 LLM calls per user request) + context bloat + viral spikes? A $4k/month bill becomes $40k overnight."

@IntuzHQ · Intuz · 100k+ daily queries, documented cost spike pattern · May 2026

"The AI hype about replacing all of us has accelerated massively. But the main problem with AI predictions is that nobody actually knows what AI will achieve. Will the small bugs preventing end-to-end pipelines get solved? Will scaling keep working? No one knows. On Twitter, you get two types of people talking about it: people with very strong opinions who have no idea about the actual capabilities of future models, and CEOs who do know more but have so many incentives it's hard to tell what's real. The honest answer is 'I don't know.' But that doesn't get engagement."

@ziv_ravid · AI researcher · Meta / NYU · cited study on AI capabilities · Feb 2026

"On April 26, 2026 OpenAI shut down its AI video generation platform, Sora. It was hailed as 'the end of Hollywood' and now it's over. Note the hype and bust cycles with AI. The same will happen with all the pointless data centres they're trying to build."

@MrEwanMorrison · Author and AI critic · Substack on tech myths · May 2026

"Goldman Sachs reports that companies are blowing past their AI inference budgets by orders of magnitude, with inference costs in engineering now approaching 10% of total headcount costs. KPMG surveyed 2,100 senior leaders and found US companies plan to spend an average of $178 million on AI over the next 12 months."

@HedgieMarkets · Markets analyst · citing Goldman Sachs and KPMG surveys of 2,100 leaders · Apr 2026

"A client sent an LOI that was exceedingly standard and a very favorable offer for the seller. The seller ran it through ChatGPT, and it told the seller that it was a bad deal, including that the seller should not agree to a non-compete (very standard). The seller terminated discussions. AI can be a great tool for people who already know a lot about a topic, but very damaging to people who use it for something they do not know about."

@Eli_Albrecht · M&A lawyer · documented deal killed by AI misread · Apr 2026

"what legal AI is good for: tireless issue spotting, finding contradictions, fixing typos, reformatting, high level legal theory, structuring the skeleton of addressable issues, draft 1 for review by lawyers. what it's not good for: fine tuned business judgment, relationship dynamic sensitivity, market understanding of cutting edge legal trends, getting from 85% to 100% perfect, situations when every word and every comma matters."

@LindsayxLin · Partner & COO, Dragonfly · Harvard Law background · Mar 2026

"Most of us are using AI for legal work the wrong way. Many rely on what the model already knows, which is mostly public, general, sometimes outdated and fake. The real advantage is to use AI on top of premium, verified legal databases. Instead of spending hours digging, I validate in minutes. Plug the AI into solid data."

@heisrahman · Experienced lawyer · using Claude Co-work + NWLRonline database · Mar 2026

"qwen overtook meta's llama in total huggingface downloads. alibaba is now #1 in open source AI. i've been running qwen models for 2 weeks. 6 benchmarks. 4 models. dense vs MoE vs distilled vs base. every test on a single 3090. their 9B beats OpenAI's 120B on key benchmarks. good architecture over parameter count."

@sudoingX · Developer · 6 benchmarks, 4 models on single 3090 GPU, full article published · Mar 2026

"Most AI coding benchmarks are meaningless. So I ran a real one: Claude Code vs Cursor Composer on a complex framework migration. No boilerplate. Just SSR, Hydration, and real technical debt. The outcome exposed two very different AI personalities."

@yoavabrahami · Wix Enterprise CTO · ran head-to-head on real production migration · Mar 2026

"everyone loves the agi bet until they see the api bills for running autonomous agents at scale. the cost gap between reasoning on opus and simple utility tasks is like 60x per request."

@newlinedotco · Developer · 60x cost gap documented, reasoning vs utility at agent scale · May 2026