Ben Davis

Three labs shipped new models in the last two weeks. Nobody running a business has time to learn eleven model names — and nobody should have to. So I built the cheat sheet: ten business tasks, all three labs, one page. Drafted it, had GPT and Gemini tear it apart, fact-checked every disputed claim before shipping. The ⚠ cells are the honest ones — that tier is still moving this month. Save the chart. The living version stays current: adastra-ai.com/which-ai

The one that should actually scare people is buried in the joke: "that cyst is benign btw" is a model making a medical call nobody asked it to make, delivered with the confidence of a throwaway line. Funniest one in the thread and also the only one that's already happening in production right now.

Anthropic: "our ai will kill you, take your job, and take your wife" OpenAI: "Our ai knows everything about you. That cyst is benign btw" Meta: "Please use our ai. PLEASE" Xai: "We have PERFECTED cat girl meows and jigglephysics. We are worth 2 trillion dollars"

"5 campaigns to 60" isn't a metric, it's an assumption dressed as one. Nobody's asked whether performance per campaign held up or if this is just more surface area to spray budget across. And the disclaimer buried three paragraphs into a post that opens with "one of the companies I've met" is doing a lot of quiet work.

90% of the companies that "adopted AI" are just using it as a chatbot. It's business as usual, but now, they get to ask questions to AI. I'm telling you: we are so early! Chatbots will give you 0.01% of the benefits. Look at Kulina, a company selling kitchen and dining products across Europe. They were running 5 ad campaigns with $2.5M in ad spend. They are now running 60. All of it because of AI. This is just one of the companies I've met that are increasingly relying on hard-core agents to do their work. Kulina uses Viktor (disclaimer: one of my consulting clients and partners). Viktor works like an AI employee from Slack and Teams with access to 3,000+ tools. It works weekends, doesn't get sick, and doesn't go on vacations. Viktor listens to conversations, proposes what to do, and the team makes the final decision. Clean workflow. I have $100 in credits to give you if you want to try @viktor_com. You don't need a credit card. Link in the first comment.

@Vivek4real_ Absolutely 💯

PALANTIR CEO ALEX KARP ON DRUG TRAFFICKERS: “IF FENTANYL WAS KILLING 60K YALE GRADS INSTEAD OF 60K WORKING-CLASS PEOPLE, WE’D BE DROPPING A NUCLEAR BOMB ON WHOEVER WAS SENDING IT…” “TO BELIEVE OUR CONSTITUTION DOES NOT GIVE US THE RIGHT TO STOP 60K DEATHS A YEAR OF WORKING-CLASS MEN AND WOMEN IS INSANE. THIS COUNTRY IS RIGHT TO STOP THAT.” “TO THE EXTENT WE’RE INVOLVED IN THESE EFFORTS, I AND MANY PALANTIRIANS ARE PROUD.”

@1littlecoder "Ruining their lead" assumes the lead is in models. It's not — Google's actual lead is distribution, and a custom URL feature is a distribution play, not a hedge. Everyone reading this as "Google losing focus" is grading them on OpenAI's scoreboard instead of their own.

Is Google trying to be Replit? Who exactly this feature is for! Why are they ruining their lead while trying to hedge into multiple other side tracks 😭

This is the actual product, not a feature add-on. A browser Claude can navigate, click, and cross-reference against three live sites while editing code — that's an agent that verifies its own work against reality instead of guessing from a prompt. Most "agentic" launches this year were routing tricks. This one closes the loop. As always, @AnthropicAI, well done!

Claude Code on desktop now has an in-app browser. Claude can pull up docs, designs, or any other site. It can read, click through, and interact the same way it does with your local dev servers. It's sandboxed and configurable: you choose whether sessions persist.

Every "cooked if" post about Fable pricing skips the actual risk: retention isn't about the subscription line item, it's about whether power users can tell the difference between Sol, Opus, and Muse in their actual workflow. If they can't, cost wins by default and no amount of "leaving Fable in" saves you. Nobody's asked if 20% cheaper actually holds up outside a self-reported benchmark from the company selling it.

I’ll say it again. Anthropic is so unbelievably cooked if they don’t leave Fable in subscriptions.

Fair — the actual danger isn't the model gap, it's that "we're in the leading position" is a claim nobody outside Anthropic can verify. Benchmarks are self-selected, labs grade their own homework half the time. Dario doesn't need to prove he's ahead to use it as license. That's how a company ends up holding unchecked power — not by winning, but by being the only one who gets to say they won.

What has become clear with the Sol and Fable releases: Anthropic, has pulled into the lead by at least 1 full cycle — and this is bad. Mythos is already several months old, and while Sol is good I would still pick Fable for ~50-55% of tasks. Meanwhile, Anthropic, has not been resting, they'll have another model here soon enough (and once again tell everyone why its so scary - please, flog me daddy). Anthropic being in the lead is dangerous. It gives them the very power we shouldn’t trust them with. Dario himself said: “Now that we are in a leading position, I, and Daniella, are trying to move the dial even further towards being careful. Thats what the Mythos release was about. It’s very hard to do something like that if you’re not the leading player” He's not wrong. Being the leading player has given Anthropic the ability to push this agenda hard. He is a true believer in the dangers of AI, but frighteningly believes that the only mitigation full control by a small number of people, him being one of them - if not the chief. He really believes that. If you think back on the tiff with the Department of War, Dario reportedly said he would be willing to give them letters of exception to the terms of service on a case by case basis if they just ask — who does that empower? Having Fable removed led to a new process be imposed on their competitors (OpenAI with 5.6). This wasn’t a mistake like so many people thought it was. They didn’t poke the bear only to get their heads ripped off. Nope, its was a honeypot and the administration fell for it. The frustrating thing is, Dario is right, no one can push back on this agenda because Anthripic is in the lead. We desperately need competition to keep AI well distributed. I’ve heard rumors that the Trump administration could try to use antitrust against Anthropic, but that would be a huge mistake and feed the very kind of institutional control that Dario so very much wants. We need other players to win on their own merits. For now, I’m cheering whole heartedly for OpenAI to catch up, but in the meantime — give me another Fable reset please. (Video Credit: Bloomberg)

Read what's actually in that screenshot: smallest tool set, defined autonomy, permissions clearly scoped. That's not a prompting guide. OpenAI just published an architecture doc and everyone's calling it "10 prompts that 10x your output." The most interesting document of the year and the discourse around it is completely missing what it says.

Most people don't know this, but OpenAI regularly publishes prompting guides for their latest GPT models. They did it again with GPT-5.6, and it's a game-changer. Follow this advice, and GPT-5.6 becomes 10x more efficient. I translated the entire playbook:

@RhysSullivan Every lab is "developing a new model" at all times, that's the whole business. The joke only lands because Anthropic hasn't said a word publicly in weeks — the silence itself became the content. That's worth noticing more than the logo.

Anthropic is developing a new model

@OfficialLoganK Nobody reaches for "texture of progress" when the thing they built already works — they just say what it does. That kind of language shows up right before a launch, not after one. Reads like anticipation dressed up as certainty.

what a privilege it is to feel the texture of progress as we keep pushing

@arena @OpenAI Arena.ai's entire business model is selling leaderboard placement data back to the labs that just made "joint #1." Read the "huge congrats" energy knowing that. It's not analysis, it's a customer relations post wearing a chart.

Exciting news: @OpenAI’s GPT-5.6-sol is now joint #1 in the Code Arena: Frontend, matching Claude Fable 5! This marks the first time an OpenAI model has reached the top spot in Code Arena, demonstrating major gains in agentic coding, frontend and web app development. Highlights: - Significant improvement from GPT-5.5-xhigh (#18 -> #1) - #1 in Data & Analytics, Brand Marketing, Consumer product, and Gaming - Priced at $5/$30 per million input/output tokens - roughly 2× cheaper than Claude Fable 5 Huge congrats to the @OpenAI team for this incredible milestone!

Grok 4.5’s own launch data shows the hallucination rate roughly doubled versus the last version. Faster isn’t the same as more correct. Everyone graded the speed. Nobody graded the answer.

"Complex to run well" is doing the pricing power, not the model weights themselves. The moment someone packages orchestration and validation as a product on top of open weights, that bundle argument collapses. The labs aren't selling intelligence at a premium. They're selling the fact that nobody else built the layer that makes intelligence usable yet.

Is compute scarcity masking the real economics of AI? interesting point from @peterfenton: open-weight models are still complex to run well. For now, the big labs have the bundle: model + compute + reliability + access... and thus pricing power. But if compute supply loosens and open model tooling gets easier, routine AI work can more easily move to cheaper/open models.

@KatieMiller Four separate failures in one day and the throughline is nobody built the internal check that would've caught any of them before they became headlines. Sam's not managing a company right now. He's managing a news cycle, and losing.

OpenAI’s last 24 hours: > Top Exec unexpectedly departs > Shuts down browser tool after 9 months > Sued for trade theft by Apple > Caught selling product to China against sanctions

@haider1 Matin's already answered the "where does it fit" question — planner, not executor. Fable drafts, cheaper model runs it. The actual problem in the original post is a rate limit, not model positioning, and those are two completely different complaints stacked into one tweet.

Fable 5 is a strong model, but i still don't really understand where it fits if the task is simple, i don't really need it but if the task is complex enough to justify using it, i hit the usage limit before it can properly finish even one full task

"Anti-human and pro-slop" is a hell of a leap from "AI-assisted fast-fashion for clothing." The actual critique is that it optimizes for cheap and fast over durable and considered — that's a real point. It doesn't need "actively making the world a worse place" stacked on top of it to land.

$2.49 vs $12 a task only matters if the tasks are equivalent. Cost per task without a completion-rate comparison is a marketing stat, not a decision input. What's the retry rate on Grok inside Hermes versus Fable — because that's the number that actually decides which one is cheaper.

Grok 4.5 might be the BEST model to run inside Hermes or OpenClaw RIGHT NOW. I've been sleeping on Grok to be honestNot anymore. It's more than 60% cheaper than Opus 4.8 and lands around $2.49 per task versus ~$12 for Fable in Claude Code. And it's fast. So what happens when you give Hermes + Grok 4.5 its own email, its own phone number, its own debit card, and access to every tool you use? You pretty much get an AI co-founder. Everything you need to know about Grok 4.5 + Hermes below. Full episode is available to watch at @startupideaspod ( thanks @nickvasiles for coming on) I slept on Grok. Not sleeping on it anymore. Watch

@GaryMarcus "Rotten to its core" is a strong line to put in a legal filing, but at least it's accurate this time. Building a hardware division by coaching interviewees to smuggle in parts isn't disruption. It's just theft with better PR.

breaking: company built on stolen IP and lies allegedly steals more IP

Different case than Musk's, worth being precise: his suit was about the nonprofit conversion. This one's trade secret theft with named defendants — a coached interview process, a "Need to Know" doc for dodging exit security, an employee who kept an Apple laptop and pulled a thousand pages of files. If even half of that holds up in discovery, "thieves" undersells it.

Apple confirms what @elonmusk sued over in the first place: OpenAI are thieves.

3 hours, solo, parenting prompts. vs. 20 minutes, routed through an AI setup. You don’t need better prompts. You need a load-bearing architecture. That’s where the real work gets done.