Kimi K3: Fable- or Sol-level quality at Sonnet prices, with open weights in 10 days

By Javi and Rafa · Café con Codely · Live analysis

Moonshot AI dropped Kimi K3 yesterday and everyone is talking about it: according to the first benchmarks it plays at the level of models like Fable or Sol, but at a price similar to Sonnet's. On LMArena it is already number 1 for frontend code with 1679 ELO points, climbing above Claude Fable 5. A historic moment: it's the first time a model that is going to release its weights wipes the floor with a frontier model.

The move is a curious one: the weights won't be released until July 27. Rafa's theory: that way you try it first on their official platform, paying Moonshot; once it lands on OpenRouter and friends there will be competition, quantization and price drops. It's a high price for what Chinese models have us used to, but for the quality it delivers —Opus-level at the very least— it's already cheap. And the context window jumps from K2.6's 262K to 1 million tokens, with 1 million of output too.

Mind-blowing detail: K3 reasons a lot and burns plenty of tokens mulling things over, and on top of that it shows its reasoning unfiltered, in full caveman mode. Ask it 'what model are you?' and its thought is 'need answer user, ask what model be you'. Caveman reasoning: other vendors hide it behind a summary so you can't distill; here you see it raw.

And the moral of the week: competition is glorious. Anthropic has kept extending Fable —now until July 19— and resetting limits, Opus 5 might get postponed because with Sol and K3 on the table it wouldn't hit as hard, and Google keeps delaying Gemini 3.5 Pro. Just a while ago Fable seemed reserved for the elites, and suddenly all of us are using it. Watch in the video.

GPT Sol tips after a week of flying close to the sun

Javi has earned the nickname Icarus: he burned through the 5-hour window on a Sunday and kept torching the OpenAI account credits without noticing. Hence this week's tips: the Max effort is hidden and must be enabled in ChatGPT's settings, try a smaller model with a higher effort to save money, turn off the fast mode they sneak in enabled by default, and watch out for Sol spawning subagents with its very same model.

The cherry on top: OpenAI silently updated the official GPT-5.6 guide —we know because we analyzed it live last week— and the advice about not asking the model for brevity is gone: there's now a text.verbosity parameter as a first-class citizen in the API, and model_verbosity in Codex CLI's config.toml. And for the hardcore café crowd, the Clodex alias: Claude Code as the harness running the Sol model underneath. Watch in the video.

npm 12 turns off install scripts, and Dependabot takes 3 days

npm gets serious about security in version 12: install scripts (preinstall, install, postinstall) are now disabled by default —the favorite vector of this past year's supply chain attacks, the kind that lists your .ssh directory and uploads it to a public repo—, and dependencies pointing straight at git repos or remote URLs are no longer allowed by default either. Late compared to pnpm, but thank goodness.

And watch me thread this one: Dependabot also changed its default and now waits 3 days after a release ships before creating the pull request, exactly one of the recommendations from our Dependabot course. Plenty of time for the security folks to detect the attack of the week and publish the advisory before an automatic PR wrecks you. Watch in the video.

The utility of the week

CodexBar

From the creator of OpenClaw. Right in your menu bar: the percentage consumed of the 5-hour and weekly windows for Codex, ChatGPT Work, Claude Code, Cursor and a ton of other providers. The key part: you can set alerts at 80, 50 or 20% so you don't pull an Icarus and burn cash without noticing. On Windows, there's Win-CodexBar.

github.com/steipete/CodexBar

The chat weighs in

eyluismi168: On Icarus' SSH dilemma: 'I think you can create fine-grained tokens to push, right?'. Javi: that would work for this one repo, but he needs pull too.

fredpalas: 'SSH with a passphrase unlocked via keychain: if they steal your SSH key it's worthless. And it commits for me remotely.'

kodexArg: On fnpm: 'it's a softlink explained verbosely'. Javi pushes back: it doesn't run npm underneath, you keep your own manager and it syncs the lockfile.

The song of the week

Rafa: «Yo quiero bailar» — Sonia y Selena (A Spanish summer classic rescued from the Caribe 2000 compilation; the audio blew up and it played in 8 bits, and the chat liked it better that way) Watch in the video.

Javi: «El Político de Pueblo» — Migue de Antílopez (From his new album: 'I'm obsessed with Miguel') Watch in the video.

The café's music playlist: cdly.to/playlist

The saying of the week

«Quien su mal consiente, a nadie se queje ('He who consents to his own harm should complain to no one')»

Own the consequences if you brought the damage upon yourself. They liked the Portuguese version even better: 'Don't complain about the deceit if you bought the cloth by the sample'. Watch in the video.

More Codely

ManagerPro

From first steps as a manager to conflict resolution, through giving feedback, low performance and one-on-ones. Open workshops and a managers' circle for companies with José Carlos. Plus the Dependabot course: keep your repo up to date, safely and automatically.

The joke of the week: Do you know why AI models are so strong? Because they train with weights!

All the news we covered

#CaféConCodely · @CodelyTV · Previous cafés: cdly.to/cafe

Share on Twitter Share on LinkedIn