Founder & CEO @ Continuous. Ex: @uber, CloudKitchens, CEO @ Invisible, Head of AI Agent Infra @perplexity_ai
Palo Alto, CA
Joined November 2009
- Tweets873
- Following266
- Followers1.1K
- Likes239
Jev would have been an immediate game-changer a year ago. Unfortunately, a lot of harnesses has been moving from custom decision space to code-based, so its applicability is now more limited.
It doesn’t solve all the problems LLM can but it’s definitely useful for many things.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Lol, this is so funny because it's so true.
This is brilliant. opusfived.dev
My favorite workflow now is to brainstorm with both Fable and Sol. The one who has a better idea will earn the right to implement it.
Anthropic is dropping Fable access mid-session 2 days before the announcement date. What are they doing?
5.6 Sol is fantastic. I always start a new model with an audit of my codebase. I know exactly what I want to see. Sol is the 3rd model (after Opus 4.8 and Fable) that I say, ok, you can work independently on my codebase.
Competition is great. Nice to see OpenAI is back in the race.
Minh Pham retweeted
I'll run a free Deep Research report on your brand or on a competitor. Here's why.
We just launched new capabilities for Archive's Deep Research product, and the fastest way to explain it is to show you your own data.
How it works: you start with the question, pick the report, and Archive asks only for the inputs that report needs. It analyzes your brand's content and your competitors' content, along with the comments on both, then turns everything into an actionable report in minutes. Trend shifts, campaign performance, competitor share of voice, creator partnerships. Open the finished report in Claude or ChatGPT and keep digging from there.
Comment REPORT with the brand or competitor you want analyzed. I'll DM you to grab the details and we'll send the report over.
This might be a stupid prediction, let’s see. With o3, OpenAI realized that people are not ready to wait longer for a better answer. With Fable, Anthropic will realize that there’s a limit to how much people are willing to pay for a coding agent.
Before Fable, I never cared about cost, just the max subscription and it’s mostly enough. Now I burn through the weekly limit within a day and need to spend additional few thousands per day for usage credit. This is more expensive than hiring a full-time engineer.
OpenAI has a huge opportunity to make a comeback here with GPT 5.6. Make $2K plan/month with fast GPT 5.6 Sol and they will have a winner.
Minh Pham retweeted
YOUR CREATOR PROGRAM JUST GOT A LOT SMARTER.
Today we're launching Archive MCP, and it changes how your team interacts with your creator data.
@app_archive is where brands like Crocs, L'Oréal, DoorDash, and UNIQLO capture and measure their creator content: the posts about your brand, the creators behind them, and the earned media value (EMV) they generate.
With Archive MCP, you can now connect Archive to the AI tools your team already uses and ask about your creator program the way you'd ask a teammate.
A few things you can ask:
→ "Show me Instagram posts with EMV over $500 this month."
→ "Which creators posted about us this week with over 100,000 views?"
→ "Summarize this week's top posts by EMV so I can share them with the team."
The possibilities are endless, and the answers come back in seconds.
Anyone on your team can get it themselves, the moment they need it, without logging into Archive or waiting on a report.
It's read-only, self-serve, and included on every 2026 plan.
Some people think this is a small update to Claude in Slack app, but I think this will be a bigger deal than people realize. Cowork doesn't have PMF for knowledge work the way Claude Code does for coding. This will fix it.
Enjoyed using Fable 5 so far, it’s definitely smarter and a good thought partner. Coding is too slow though, I still think Opus 4.8 ultracode is a better overall experience. A good combination might be using Fable for building specs and Opus for implementation.
This is the first time that using the frontier model doesn’t improve the experience for me. Up until now, I just grab Opus for everything.
Fable 5 is like a slow staff eng. Opus 4.8 is like a lazy senior eng. Opus 4.8 ultracode is like a good senior eng.
This is a great take, and on the same day Anthropic released Fable 5 no less. This matches my thought process as well. While most of the valley is deep in despair, I’m incredibly excited about the future.
2 surprising things about Fable 5: one is that the price is not too expensive (only double Opus and slightly less than double GPT 5.5). Two is that it’s the first new Claude model that perform worse on computer use. Is computer use dead?
I stopped caring about coding eval results because I don’t feel like they match reality anymore. I’m glad to see FrontierCode, it seems match what I see in the wild really well, from the jump in Opus 4.5 back in November to current performance on hard tasks like diamond.
It's finally out!!! @METR_Evals found that more than half of SWEBench results is unmergeable slop. FrontierCode represents over 1000+ hours of maintainer validated software engineering work most frontier models cannot yet solve, much less solve with high quality.
Cog had IOI Gold medalists and top code maintainers Look At The Data — FrontierCode includes 3000+ rubrics covering code quality and anticheat reward hacking plaguing other benchmarks.
FC Diamond is so hard that Opus 4.8 scores 13.8%.
Three eras of AI coding : Three eras of benchmarks
2021 • Autocomplete : HumanEval
2023 • Passing Tests: SWEBench, TerminalBench
2026 • Maintainable Code: FrontierCode
to me the most beautiful chart when I requested a special historical run into all extant old models, the data was finding that the easiest third of FC tasks (in FC Extended) were rapidlly and suddenly solved over late 2025 - Opus almost doubled from a 41% pass rate to 74% in 4 months.
This describes the "WTF happened in Dec 2025" vibe shift that a lot of folks from @dhh to @karpathy have called out: it is the difference between getting 95% success in 2 rerolls vs 6, making it finally feasible to go up the next layer of abstraction in agentic coding, eg @GeoffreyHuntley's ralph loops or @bcherny's /goals or @steipete's "loops that prompt your agents" without fearing too much that things go off the rails.
My guess: as AI accelerates from here, each FrontierCode tier will saturate in sequence, hopefully ~annually. I've already asked the team to prepare FrontierCode 2027....
The old mountains will be destroyed. Their rubble becomes regolith. And from that regolith, the next model forest grows. Circle of life.
Congrats @lateinteraction and @a1zhang. I have been bullish on RLM for a while, it’s very nice to see it got adopted in Claude Code.
People still don’t realize how early we are. Half the “AI-native” stuff shipping today is a chatbox glued onto a 2020 product. The whole computer is about to get rewritten. Magic Pointer + Googlebook is the first real preview.
