Typeface
Text size
Sound
Serif is Literata (the reading face); Sans is Inter. Applies to article text; saved on this device.
Volume
Off by default — never plays on its own. Plays while you read this page; one press resumes after you move to the next.
Founding Editor
Aditya Marin Gasga is the founding editor of The Counter Brief and Head of Growth at Demand Nexus, its parent company, where he works on sourcing qualified pipeline across SDR, content, and paid channels. His background is in performance marketing and demand generation. He studied business administration at Northumbria University.
81% of enterprise tech leaders say AI-generated code is causing more production failures, while 92% still think their code was ready to ship. The gap between those two numbers is the whole problem.
Companies score leads on engagement, opens, clicks, page views, then wonder why the high scores don't close. Engagement measures interest, not fit, and the two are not the same thing.
Tesla, Uber, Meta, Amazon, and Walmart have all put hard caps on employee AI spending. Token-based billing turned every prompt into a variable cost, and the tools shipped without the guardrails to contain it.
Claude Code's price ranges from $20 to over $1,200 a month for identical work. The sticker price on the pricing page is the wrong number. Here is what actually determines your Claude Code cost.
OpenAI and Anthropic each stood up private-equity-backed enterprise deployment firms within a week of each other. The convergence says the hard part of enterprise AI was never the model, it's getting anyone to use it.
The US restricted Anthropic's Fable 5 for about two and a half weeks using export controls built for physical goods — exposing what all sides now admit: no established process exists for restricting a frontier AI model.
The fear is mass layoffs. The data shows something narrower: a 13 percent drop in entry-level employment in exposed fields, while the AI screening those candidates runs under a bias law almost no one enforces.
Bots now generate more web traffic than humans, and AI answers are replacing the clicks publishers were built to capture. The deal that funded the open web is quietly coming apart.
The public imagines diagnostic AI catching what doctors miss. The actual deployment win is documentation, and meanwhile half the cleared diagnostic devices were never shown to work in a clinical study.
Federal regulators ordered all six US grid operators to justify or rewrite the rules for connecting AI data centers, shifting the legal burden onto the operators and turning a year of signaling into binding deadlines.
ARPA-H is funding a push for the first FDA-authorized agentic AI for clinical care, deliberately starting with a high-risk, well-defined disease so regulators are forced to set the precedent.
NHS England is putting Microsoft Copilot in front of 505,000 staff after its own pilot reported saving 43 minutes a day. The open question is whether that time becomes patient care or just moves somewhere else.
Tesla and Figure get the headlines. The AI earning its keep on the factory floor is unglamorous: predictive maintenance that cuts downtime by up to half, and vision systems that catch the defect a human eye misses.
The intermediate companion to our RAG explainer: the four ways RAG degrades in production (chunking, retrieval, faithfulness, a stale index), the fixes that hold, and the evaluation that makes them visible.
A from-scratch explainer of RAG: how chunk, embed, store, retrieve, augment, and generate fit together, why it reduces hallucination without ending it, and where it quietly breaks, in about 13 minutes.
A from-scratch explainer of the context window: the shared token budget a model reads at once, why long context costs more and recalls less, and when retrieval beats a bigger window, in about 12 minutes.
AWS gave AI agents a way to spend money autonomously. The payment was never the hard part; authorizing, capping, and auditing a non-human spender is, and most enterprises have not built for it.
The public imagines AI trading the markets. The real deployment is in the back office, where one bank put a model in front of 200,000 employees, and the job most at risk once trained the next managing director.
LangGraph vs CrewAI vs OpenAI Agents SDK: pick by philosophy, not features. A state machine for auditable production, role-based crews for fast prototypes, handoff chains for OpenAI-native speed.
The same tool that does real work in contract review is getting attorneys sanctioned for citing cases that never existed. The line between the two is one discipline, and it defines where AI fits in legal work.
Same technology, opposite calls, one stretch. AI retail's clearest lesson: cashierless checkout works where the basket is small and fails where it is big. The deciding variable is the use case, not the capability.
Clearance was supposed to be the hard part for medical AI. It isn't. The gap between a cleared algorithm and a bedside is a billing code, and most tools lack one. Reimbursement, not accuracy, is the bottleneck.
HHS is running five years of audit reports through ChatGPT across all 50 states. A first pass over public documents is defensible; as the basis for cutting a state's funding, the model's confident errors are the catch.
The capex is real, rational, and historically unprecedented. Whether it pays comes down to one clock: the depreciation schedule on the chips, running against a revenue ramp that may be slower than the buildout assumes.
A model went from launch to government-ordered shutdown in three days. Everyone measures capability. Governability, whether a model can legally run for your users tomorrow, is the axis that just bit, and few price it.
Trained inspectors miss 15 to 30 percent of defects and wrongly reject about a third of good parts. That gap is why computer vision pays off in factories. The metric that decides ROI is not the one vendors lead with.
Washington is arguing over whether to set one national AI standard or strike down the states. For anyone shipping a product, the binding rules already exist, and they live in Brussels and a handful of state capitals.
Training-data law is resolving into a rule: training can be fair use, but how you got the material and whether the output competes with it can turn the same act into liability. Provenance is now the question.
A bipartisan run of states has passed AI laws since the federal government warned them off, and the fight over who governs AI has moved from a memo to a courtroom. The vacuum did not stop regulation; it scattered it.
Anthropic removed the models and, in the same statement, called the order wrong. The dispute: whether what the government calls a jailbreak finding can recall a model used by millions, and the process it skipped.
Three days after launch, an export-control directive ordered both models suspended for any foreign national. Because that line cannot be drawn in real time, Anthropic shut them off for everyone, worldwide.
OpenAI models and Codex can now be bought against your existing Oracle Universal Credits, no separate contract. It removes the blocker that stalls enterprise AI, and it bundles two lock-ins that reinforce each other.
Everyone is still counting GPUs. The real bottleneck in 2026 is firm electricity, and the companies that train frontier models increasingly depend on whoever can deliver the megawatts.
Gemini 3.5 Flash is GA at $1.50/$9 per million tokens (cached input $0.15): triple the Flash it replaces, 25% cheaper than 3.1 Pro, and ahead of it on coding and agentic benchmarks. What it buys.
The $1.77 trillion debut is the largest IPO in history. The detail that matters more sits in the prospectus: SpaceX now owns xAI, so the float quietly took Grok public, and that changes what you're betting on.
Agentic AI plans and executes multi-step tasks on its own: it decides what to do, calls tools, and loops until the goal is met. A clear guide to what actually ships as an agent in 2026.
The five vector stores worth considering in 2026, what each costs, how fast it is, and the one question that decides most of them: do you even need a dedicated database, or is Postgres enough?
Sonnet 4.7 shipped this morning with Python execution baked directly into the API: no separate sandbox, no harness wiring, one new boolean parameter. Same model pricing as 4.6 plus a per-execution surcharge.
The Counter Brief's pricing calculator shipped without DeepSeek, Qwen, Llama, or Mistral. A reader called it out. Here's what changed, the corrected pricing, and when each non-US model is the right pick.
Your dashboard says 90% cache hits. Your bill says otherwise. The gap is almost always three specific patterns inside the system prompt that don't show up in any log, and you can fix them in an afternoon.
A from-scratch explainer of what embeddings actually are, how they're compared, why they make modern search possible, and which model to pick in 2026, in about 15 minutes.
The Model Context Protocol went from a quiet Anthropic release to industry-standard plumbing in under two years. Here's what it is, why every major lab adopted it, and where the hype outran the facts.
Most cache audits never happen because nobody owns them. Here's the version that survives: 5 minutes a week, one CSV file, and a GitHub Action that runs it whether the owner remembers or not.
Seven AI agents tested on the same task suite: Manus, Devin, Operator, GPT Agent Mode, Gemini Workspace, Replit, Claude Computer Use. Two are genuinely useful today; the rest need a specific use case.
Six AI coding tools, one honest conclusion: there's no single winner, and the best developers pair two. Here's what each is actually for, what it costs, and which combination fits your workflow.
Google's Gemini 3.5, launched at I/O 2026, is built around a single idea: stop answering questions and start completing tasks. Here's what actually changed and why it matters.
Grok 4.3 isn't trying to top the leaderboards. It's trying to be the model you can afford to run at scale, and on that bet, it mostly delivers. Here's where it fits and where it doesn't.
A 70% versus 95% cache hit rate isn't the model. It's whether you set a cache_control breakpoint and keep the prefix before it byte-identical, with every dynamic value pushed into the user turn.
OpenAI filed confidentially for what could be the largest tech IPO ever, at a valuation near $1 trillion, while losing more than a dollar for every dollar it earns. Here's what's actually going on.
OpenAI flipped agent mode on for all Plus and Team accounts this morning. The capability is real; the practical use cases are narrower than the demo suggests.
We tested seven AI writing tools on the same set of tasks. Three are great, three are fine, and one is genuinely worse than nothing. Here's how they stack up.
Opus 4.8 lists at $5/MTok in and $25 out (a third of the prior Opus generation) with cache hits at $0.50. Here's what that does to the real per-request math for production teams.
A from-scratch explainer of how large language models actually work (tokens, attention, the inference loop, and what to make of it all) in 12 minutes.
A four-prompt workflow that turns a 4-hour blog post into a 75-minute one, without the soulless AI tells. Real prompts, real before/after.