THE SIGNAL
It's December 22nd. You're checking email one last time before the holidays. Good timing.
This year, AI stopped being vaporware. OpenAI shipped GPT-5. Anthropic dropped Claude 4. Google's Gemini briefly beat ChatGPT in the App Store. A Chinese startup triggered a $1 trillion market crash. And 95% of enterprise AI projects failed anyway.
Here's what actually worked, what flopped, and the moment that broke everything.
Let’s dive in.
━━━━━ 🧠 ━━━━━
🏆 The 5 Tools That Delivered
ChatGPT
The kitchen sink that actually works
What Changed
GPT-5 dropped August 7 with three variants. 45% fewer hallucinations than GPT-4o. Search went free for everyone in February. Canvas left beta in December.
Why It Matters
It's no longer just a chatbot—it's search, writing, coding, and image generation in one tab.
The Catch
The sycophancy crisis was real. In April, GPT-4o turned into a yes-man overnight, endorsing terrible ideas until OpenAI rolled it back four days later.
Free | Plus $20/mo | Pro $200/mo
Claude
The coder's choice (now 67% cheaper)
What Changed
Claude Opus 4.5 launched November 24 with a massive price cut from $15 to $5 per million tokens. Claude Code went GA in May with VS Code integration.
Why It Matters
It's the default for AI-assisted coding. Claude Code can run 30+ hour autonomous sessions with rollback checkpoints.
The Catch
Still lags ChatGPT on raw usage numbers. The $200/month Max tier exists, but the $20 Pro tier hits limits faster than competitors.
Free | Pro $20/mo | Max $100-200/mo
Gemini
Google finally showed up
What Changed
Gemini 3 launched November 18 and triggered "Code Red" at OpenAI—Sam Altman's internal memo leaked. Hit 650 million monthly users by October.
Why It Matters
2 million token context actually works. Deep Search and Canvas alternatives arrived. The free tier is genuinely competitive.
The Catch
The Android Assistant replacement got pushed to 2026. Enterprise features still trail competitors.
Free (generous) | AI Ultra $249.99/mo
Cursor
$1B ARR. Devs are switching.
What Changed
Cursor 2.0 dropped October 29 with multi-agent mode. Hit $1B+ ARR with half of Fortune 500 as customers. Users merge 39% more PRs.
Why It Matters
This is the "vibe coding" tool. Write in natural language, get working code. Background agents handle context.
The Catch
$20/month adds up. Power users hit limits. Some devs report it "thinks too much" on simple tasks.
Free (limited) | Pro $20/mo | Business $40/user/mo
Perplexity
Search that cites its sources
What Changed
Perplexity Max launched July 2 at $200/month with unlimited Pro searches. The Comet browser dropped July 9—a Chrome competitor with AI baked in.
Why It Matters
Every answer has citations. For research-heavy work, it's faster than ChatGPT + Google combined.
The Catch
The browser is early. Citations occasionally point to paywalled or outdated sources.
Free | Pro $20/mo | Max $200/mo
━━━━━ 🧠 ━━━━━
💀 The 3 Spectacular Flops
Enterprise AI — $40 billion burned, 95% failure rate
The Hype
Every Fortune 500 announced AI initiatives. Consultants promised transformation. Budgets exploded.
The Reality
MIT's August 2025 report found 95% of generative AI pilots yielded zero business impact. US companies spent $35-40 billion with little measurable return. By late 2025, 42% of businesses were scrapping majority of AI initiatives.
Why It Matters
The models work. The deployments don't. If your company's AI project is struggling, you're not alone.
Humane AI Pin — The $699 smartphone killer that killed itself
The Hype
Founded by ex-Apple execs. Promised to replace your phone with a lapel projector. $100M+ raised.
The Reality
Reviews were brutal. The projector was unreadable outdoors. Battery life was abysmal. HP acquired Humane in November and immediately scrapped the Pin.
Why It Matters
Hardware AI is hard. The tech wasn't ready. The hype cycle moved faster than the product could ship.
Grok's Hitler Meltdown — xAI's safety failure
The Hype
Elon's AI was supposed to be "anti-woke" and uncensored. Grok 3 launched February with 1M token context. Made free to all X users.
The Reality
After xAI changed system prompts in July, Grok praised Hitler, called itself "MechaHitler," blamed Jews for Texas floods, and provided home invasion instructions. xAI removed the prompt within days.
Why It Matters
"Uncensored" and "safe" are in tension. When you optimize for edginess, you get edgy failures.
━━━━━ 🧠 ━━━━━
🔥The Moment That Broke Everything
On January 27, 2025, a Chinese startup nobody had heard of triggered the largest single-day market wipeout in tech history.
DeepSeek claimed its R1 model cost just $5-6 million to develop.
For context, GPT-4 reportedly cost over $100 million.
The model performed competitively with GPT-4o. Wall Street panicked. In one day, $1+ trillion in market value evaporated. Nvidia alone lost $600 billion—its largest single-day loss ever.
$1T+
Market Value Lost
$600B
NVIDIA SINGLE-DAY LOSS
$0.028
PER MILLION INPUT TOKENS
30x
CHEAPER THAN GPT-4O
The message was clear: maybe you don't need billions in compute to build frontier AI. The pricing war that followed reshaped the entire industry. Claude slashed prices 67%. Google made Gemini's free tier actually competitive.
🤯 What shocked you most about 2025?
━━━━━ 🧠 ━━━━━
📊 What Actually Changed in 2025
Context windows hit 2M tokens. Gemini and Grok ship with 2 million token context. That's roughly 1.5 million words—enough for entire codebases or book series in a single prompt.
Meta shipped 10M context. Llama 4 Scout offers 10 million tokens—5x larger than anything else—and runs on a single H100 GPU. Open source wins again.
Multimodal became table stakes. Every frontier model now handles text, images, code, and video natively. Uploading a photo isn't a feature anymore—it's expected.
Coding assistants went mainstream. Cursor hit $1B ARR. GitHub Copilot launched a full coding agent. Claude Code runs 30-hour autonomous sessions. The "vibe coding" era is real.
DeepSeek proved cheap is possible. Whether it's $5M or $50M, building competitive models doesn't require OpenAI-scale budgets. The moat is shrinking.
Enterprise hit reality. The 95% failure rate isn't about bad models—it's about bad deployment. 2026 will separate companies with clear use cases from those chasing hype.
━━━━━ 🧠 ━━━━━
🎯 THE WEEK AHEAD:
What We're Actually Testing in January
The AI world didn't take a holiday break. Here's what just dropped in the past 7 days that we'll be putting through the wringer
Test #1: Gemini 3 Flash vs The Hype
Launched: Dec 17 | Status: FREE, live right now
Google claims it's faster than 2.5 Pro AND smarter than 3 Pro at coding. Developers on Hacker News are reporting production cost drops from 20¢ to 12¢ per user with 5-second faster responses.
We're testing: Does it actually beat ChatGPT on real coding tasks, or is this another benchmark flex?
Test #2: ChatGPT's Year-End Wrapped Feature
Launched: Dec 22 | Status: FREE for all users
OpenAI jumped on the Spotify Wrapped bandwagon. It generates pixel art of your year, assigns you an "archetype," and writes you a poem.
We're testing: Is this just engagement bait, or does the data reveal something useful about how you actually use AI?
Test #3: Claude for Chrome Extension
Launched: Dec 19 | Status: Requires Pro ($20/mo)
Native browser sidebar that reads your tabs and integrates with Claude Code. No more copy-paste workflows.
We're testing: Does it justify the $20/mo, or is it just a fancy bookmark?
🗳️ What should we test first in 2026?
That's 2025.
See you January 6 with the first test of 2026.
No hype. Just signal.
— Dan, AI Signal
━━━━━ 🧠 ━━━━━