51 stories
A Wagtail maintainer ran 1B tokens of GLM 5.3 Flash in September for $68 and kept it as the default for 95% of tasks. The costs, the setup and what breaks.
Only Kolibri's weights are out. Beam and Mistral Large 4 stay API previews until late October, and Kolibri's 4-bit build runs at 50 tok/s on a 64GB Mac.
DeepSeek Harness runs any provider's model, Claude Code only Claude's. The lock-in sits in plugins, hooks, sessions and config, and those do not move.
On Cursor's own benchmark, Opus 5.5 beats every Grok 4.7 setting for less, and Composer 2.5 is the cheap editor. Prices, plan burn and the break-even.
Commit on green, keep every caller with its change, and stack small pull requests. The diff-size numbers behind it, and the hooks that enforce it.
Memory is what an agent noted and a plugin replays. Documentation is what you wrote, reviewed and committed. How each loads, and what to write down.
ds4 runs big MoE models on 96 GB Macs, Magnitude tunes kernels to your device, Strata puts one model on a gaming PC. Every speed claim is self-reported.
fast-jev-compaction prunes in a second, not 30. At its default threshold it keeps almost no tool results, and a scorer that always says no cuts as much.
The agent writes the film as code, headless Chromium captures each frame, FFmpeg encodes. LaunchVideo spends 100k tokens a film; a 14-agent reel cost $30.46.
A durable harness commits every step of an agent run to storage, so a crash resumes mid-turn. What Pi Durable shipped on 1 October, and what is only claimed.
GitHub was the most added tool on Stackness in September 2026. Supabase and Cloudflare went from no profiles to 6 and 5. The table and an npm check.
Sonnet 5.5 costs half of Opus 5.5 per token, but cache reads cost $0.20 on both. Which model plans, which one edits, and what one metered run saves.
A frozen eval panel is the only rerunnable way to tell whether a hosted model got worse. What livenerf measures on Opus 5.5, and how to size your own.
You can get Jev-style typed decisions from an ordinary LLM by forcing one output token and reading its logprobs. Where that matches Jev, and where it loses.
An ADE is an app built around tasks that coding agents run, not files you edit. What JetBrains Air shipped, and whether an ADE replaces your IDE.
Facts every session needs go in AGENTS.md, procedures go in skills, and rules that must hold go in hooks. How to share one handbook across six agents.
Opus 5.5 sends its between-tool notes as empty thinking blocks and ends some turns early. What breaks in an agent harness, and the documented fixes.
To run a System One model locally, pick a runtime too: Ollaya serves ten open model families, laya-mlx runs Laya on Apple silicon, and llama.cpp needs a shim.
Showing 36 of 51 stories
Load more storiesGet the top AI news and tools straight to your inbox every week.
Weekly · AI news
Join 15,000+ AI enthusiasts
