Dagens Vibes — 9. oktober 2026

Fra X-feedet

Gratis sikkerhedsscanning — med en reproducer

Anthropics OSS Scanner sender tilmeldte open source-projekter fejlrapporter med reproducer og mulige patches. OpenSSL og PostgreSQL beskriver brugbare fund; flaskehalsen flytter fra at finde fejl til at få dem vurderet og rettet.

Anthropic@AnthropicAI

To secure open-source software, we’re launching OSS Scanner. We’ll use our frontier models to periodically scan opted-in open-source projects for vulnerabilities, at no cost. Our reports will provide a proof-of-concept, explanation, and suggested fix. https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source

♥ 534 · ↻ 54 · 💬 34 · 🔖 165

https://x.com/AnthropicAI/status/2108302543977906649

Step 5: en times agentarbejde for $2,26

Step 5 Preview kan prøves gratis i Cline; open weights er lovet til 15. oktober, men er ikke udgivet endnu. Det bedste vibe check er en sneglobus: en times byggeri med visuel selvkontrol og en konkret API-regning — mere interessant end endnu en benchmarktabel.

Cline@cline

Step 5 Preview is now free in Cline. On DeepSWE it scores ahead of Kimi K3 and GLM-5.3, making it one of the strongest open weights coding models available. https://t.co/GWRbfaJyqY

StepFun@StepFun_ai

Introducing Step 5 Preview: Advancing the Pareto Frontier. Step 5 Preview is our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance. - 600B total / 27B active MoE, with 1M context + Vision - Substantially lower task cost at comparable intelligence - Broad software engineering capabilities with sustained execution over long horizons Try Step 5 Preview: https://platform.stepfun.ai Model page: https://www.stepfun.com/step-5-preview Open weights on Oct 15.

♥ 344 · ↻ 9 · 💬 9 · 🔖 60

https://x.com/cline/status/2108249633789263917

leo 🐾@synthwavedd

step 5 preview from @stepfun_ai spent an hour building me this cool little snow globe and iterated against the small details, constantly checking its work! it zoomed in on its own tiny cottages to check nothing was clipping, & even made a histogram to verify the snow falls from the very top the best part? the api bill was $2.26, for about an hour of work! that's the pareto frontier at just $1 in/$2.70 out per million tokens step 5 preview is now on openrouter and available in kilo code, cline, hermes agent, and opencode so you can keep your workflow and just switch models. more integrations to come :3 try step 5 preview on openrouter - http://openrouter.ai/stepfun/step-5-preview

StepFun@StepFun_ai

Step 5 Preview now exists where you build. Now live on @OpenRouter. A week of free access is rolling out today across @opencode, @cline , @NousResearch, @kilocode, and more. Flagship-tier intelligence for agentic and professional work, at substantially lower task cost. Keep your workflow. Switch your model. Try Step 5 Preview on OpenRouter: https://openrouter.ai/stepfun/step-5-preview

♥ 126 · ↻ 0 · 💬 5 · 🔖 13

https://x.com/synthwavedd/status/2108271207992930765

Le Chonk møder en rigtig opgave

Mistral Large 4 har fået et mindre flatterende førstehåndstjek: Sudo su rapporterer 57 minutter og 233k tokens uden en skrevet fil. Én opgave, ikke en dom over modellen — men en nyttig modvægt til lanceringen; vægte er lovet inden månedens udgang, licensen er stadig ikke angivet i launch-posten.

Sudo su@sudoingX

mistral large 4 failed miserably, europe's 1t model thought for 57 minutes, burned 233k tokens and never wrote a single file qwen 3.8 flash next is a 180b chinese open model and it built the whole thing on my dgx spark in 2.5 hours, eu 1t vs china 180b, that's the gap, let that sink in glm 5.3 flash and deepseek v4.1 flash built it too, all local, all open weight, all chinese and before anyone says turn thinking off, high is the default and its only setting besides off, i ran it the way it ships, same as every model before it so why release a model that can't finish a turn, forget finishing the build, it doesn't stop thinking long enough to start, 38 on artificial analysis and zero files on my desk, the eu writes the most ai rules on earth and shipped a model that can't write one file

Sudo su@sudoingX

mistral large 4 (le chonk) thinks too much on high, 37 minutes in and not a single file written, le chonk even typed "hmm, overthinking" to itself > 0 files in 37 minutes > 2 calls of 15 and 16 minutes, about 47k and 50k thinking tokens each > cut off twice at the output cap before it saved anything > 68 tok/s decode, $0.34 spent so far high is its only setting besides off, so this is the full chonk, the tree is still in its head

♥ 91 · ↻ 10 · 💬 19 · 🔖 10

https://x.com/sudoingX/status/2108074915316605254

Haiku vs. Luna: samme tokenpris, forskellig regning

En udvikler målte 45–50 % flere fakturerede tokens hos Haiku for samme tekst ind og ud, uden reasoning; Luna blev omtrent 30 % billigere i forsøget. Gårsdagens tokenizer-pointe får dermed endnu et konkret vibe check: prislisten er ikke regningen.

AVB@neural_avb

Big revelation: Haiku and Luna tokenizers are wildly different. Completely changes how much you spend via API. Both models are priced at $0.10/$0.50 per million input/output token. But Luna is roughly 30% cheaper. Initially I thought that the additional cost might come because Haiku thinks more. But this holds true even when reasoning is off. It is a tokenizer thing. For identical text in and identical text out, with zero reasoning tokens on both sides, Haiku is billed about 45 to 50% more tokens than Luna. Image below. Regarding performance and latency, I suggest everyone to run their own evals inside your products before trusting these PR benchmarks. Ran some internal benchmarks for office and it seems Luna does perform slightly better OR on-par with Haiku.

AVB@neural_avb

Both Haiku 5.5 and GPT-6-Luna is $0.10 / $0.50 per million tokens. This is the most exciting time to build AI into your products as a dev I have been using luna for a ton of background stuff in my projects, it is absolute gold Apparently Haiku just mogged it in every category https://t.co/wqrJM6m0xH

♥ 61 · ↻ 3 · 💬 6 · 🔖 15

https://x.com/neural_avb/status/2108105405293625824

Agentens terminal får en ordentlig indhegning

Microsofts MXC samler sandboxing på Windows, Linux og macOS bag SDK’er til blandt andet Node. Unsloth har integreret det og rapporterer under 100 ms overhead — konkret agentinfrastruktur, ikke bare en Docker-container med en flot landing page.

Unsloth AI@UnslothAI

Windows now has sandboxing! Microsoft released an open-source repo, mxc, for sandboxed code execution. We collaborated with Windows to add mxc OS level sandboxing to Unsloth which adds just <100 ms of overhead. GitHub: https://github.com/unslothai/unsloth Guide: https://unsloth.ai/docs/new/studio/sandboxing-in-unsloth https://t.co/NTGo1l4iRQ

♥ 1047 · ↻ 143 · 💬 26 · 🔖 469

https://x.com/UnslothAI/status/2108229721159053557

Lad agenten se sit eget diagram

Excalidraw har fået en officiel CLI til lokale filer og Excalidraw+. Den lille praktiske gevinst: agenten kan rendere diagrammet og kontrollere billedet, i stedet for bare at erklære sin JSON kunstnerisk vellykket.

Excalidraw@excalidraw

Excalidraw CLI released! Log in to your Excalidraw+ account or use with local files. You can easily render Excalidraw diagrams to give better output feedback to your agents. Install today. https://t.co/cl2V7n15Xe

♥ 1006 · ↻ 57 · 💬 34 · 🔖 564

https://x.com/excalidraw/status/2108249557519970369