Log inSign up
Kilo
4,219 posts
Image
user avatar
Kilo
@kilocode
Kilo is the all-in-one agentic engineering platform. 3M+ Kilo Coders. Open source since day one!
VS Code, JetBrains, CLI, Cloud
kilo.ai
Joined March 2025
222
Following
27.5K
Followers
  • Pinned
    user avatar
    Kilo
    @kilocode
    20h
    Article cover image
    Article
    We predicted the $100k/yr-per-dev AI bill. Now the winners are routing around it.
    Three of the largest IPOs in history are arriving in the same window. SpaceX went public on June 12 at a $1.77 trillion valuation, the biggest listing in history. Anthropic filed confidentially around...
    1.5K
  • user avatar
    Kilo
    @kilocode
    12h
    Auto Efficient vs Claude Opus 4.8. @coldopn ran them both through Terminal Bench 2.0 on Kilo's agent harness. 445 tasks, 5 attempts per model, real costs measured end to end. The result is one of the better cost-vs-performance stories we've seen.
    Comparison graphic titled "Auto Efficient vs Claude Opus 4.8," benchmarked on Terminal Bench 2.0 using the Kilo agent harness. Two headline stats on the left: 69% of Opus completion, and 77% lower cost per attempt. Two cards on the right compare the models. Auto Efficient: 46.7% completion, $19.60 per attempt, $0.22 per task. Claude Opus 4.8: 67.6% completion, $85.19 per attempt, $0.97 per task. Footer notes 445 tasks across 5 attempts per model, with real measured cost.
    1.9K
    user avatar
    Kilo
    @kilocode
    12h
    Replying to @kilocode and @coldopn
    Here is the best part: all of this is live right now. Auto Efficient routes each request to the model that fits the work, so you get results like these without managing model selection yourself. We showed our work:
    Image
    Auto Efficient vs Claude Opus 4.8 — KiloBench Terminal Bench 2.0
    From kilo.ai
    485
    user avatar
    Kilo
    @kilocode
    38m
    The blocks physics simulator pushed the two models furthest apart. The prompt asked for draggable 3D blocks that stack and tumble when they are off balance. Auto Efficient built it for $0.05, and Opus built it for $1.31. That is 26 times the cost for the same prompt.
    116
  • user avatar
    Kilo
    @kilocode
    13h
    Good thread, and the risk is real. It's exactly why model-neutral and open source matter in this day and age. It's also why we built Kilo for Slack the way we did: you get the tag-it-in teammate that follows threads and automates work, but everything stays portable and you're
    user avatar
    Ashwin Gopinath
    Sentra
    @ashwingop
    20h
    Claude Tag is a Trojan horse.  Not because Anthropic is doing anything evil. Because the incentives are obvious. Day one, this looks like a great feature: tag Claude in Slack, let it follow the thread, remember context, connect to tools, break down tasks, chase work, and act
    985
  • user avatar
    Kilo
    @kilocode
    16h
    Lock-in is the bill. Routing is the answer. We gave GLM-5.2 and Kimi K2.7 Code the same build spec, near-identical working output from both, at a fraction of frontier cost per attempt. Pick a tier, route each task to the model that fits it.
    Image
    user avatar
    Scott Breitenother
    @s_breitenother
    17h
    Everyone's screenshotting the May Ramp AI Index: Anthropic passed OpenAI, 34.4% to 32.3%. The real story is underneath, from Ramp's own economist: the leader earns more when you buy more tokens, so it's built to push you toward pricier models when a cheaper one would do.
    2.2K
    user avatar
    Kilo
    @kilocode
    16h
    We wrote the whole thing up:
    blog.kilo.ai
    We predicted the $100k/yr-per-dev AI bill. Now the winners are routing around it.
    The teams pulling ahead aren't spending the most, they're routing each task to the right model.
    435
  • user avatar
    Kilo
    @kilocode
    Jun 23
    You're paying frontier prices to rename a variable! Auto Efficient routes each request to the cheapest model that can handle it. Rename a variable, you get a small model. Plan a migration, you get a big one. You don't switch anything. How it works 🧵
    Black card titled "Auto Efficient: the right model for every request, automatically." Two rows show task routing: "Rename a variable" routes to "a fast, cheap model," and "Plan a migration" routes to "real firepower." Three facts below: decides between your keystrokes, routes on a public benchmark, never drops below a capable paid model.
    24K
    user avatar
    Kilo
    @kilocode
    17h
    The model you pick matters less than whether the routing fits the task. Running a frontier model on a variable rename costs more for the same result. Most agent work is trivial edits that never needed the big model. Routing by task complexity is the part that was missing.
    419
  • user avatar
    Kilo
    @kilocode
    18h
    SF builders: we're teaming up with @novita_labs, in person in SF and online for everyone else, for a hackathon kickoff night. No engineering background needed. One site, one goal: Build the site that sells you. You'll get a week to make it your best work. 🍕 Pizza, workshop,
    Promotional graphic for a Novita and Kilo Code hackathon kickoff night. The Novita and Kilo Code logos appear at the top against a dark blue-to-green gradient background. Large yellow text in the center reads “HACKATHON,” with smaller gray text below reading “KICKOFF NIGHT.” A black pill-shaped banner at the bottom displays a green status dot, the date “June 30,” and “All Levels.”
    1K
  • user avatar
    Kilo
    @kilocode
    23h
    Ten months ago, we said AI coding bills would hit $100k per dev per year. People thought that was just a hot take. This week, Ramp confirmed they're spending about $90k per dev on AI. We weren't being dramatic. We were just early.
    Scatter plot titled with axes "Completion %" along the bottom and "Cost per attempt (USD)" up the side. Yellow-green bubbles represent AI coding models, plotted by how much each completed versus its cost. Higher-cost frontier models like Claude Opus 4.7, GPT-5.5, and Gemini 3.5 Flash cluster toward the top right; lower-cost efficient models like GLM 5.2, DeepSeek V4 Pro, and the free Laguna M.1 sit lower and left.
    2.5K
    user avatar
    Kilo
    @kilocode
    23h
    Replying to @kilocode
    And it's a default, not a cage. 500+ hosted models, BYOK, and local are all still there. Auto Model just means you're not hand-picking a model every time and overpaying when you guess high.
    Screenshot of the Kilo Code panel inside a VS Code editor. A model selection dropdown is open, showing options including Auto Efficient (highlighted), Auto Free, Poolside Laguna M.1, StepFun Step 3.7 Flash, and several Anthropic Claude models. The Kilo logo, a yellow square, sits at the top.
    408
    user avatar
    Kilo
    @kilocode
    23h
    The $100k per dev bill is real. Whether you route around it is the difference between a budget that scales and one that breaks. Set a tier, let it route, pay frontier prices only when the task earns it.
    Image
    Auto Model - Kilo Chooses the Right AI Model for Each Task
    From kilo.ai
    372

New to X?

Sign up now to get your own personalized timeline!

Create account

By signing up, you agree to the Terms of Service and Privacy Policy, including Cookie Use.

Terms·Privacy·Cookies·Accessibility·Ads Info·© 2026 X Corp.
Don't miss what's happening
People on X are the first to know.
Log inSign up