To frontier-modeller lander på samme døgn, mens den tredje står på tærsklen med hidtil uset cyberkraft. Samtidig får agenter længere snor, billigere hukommelse og adgang til fysiske maskiner. Praktisk. Også en anelse “hvem gav gravemaskinen nøglerne?”.
Fra X-feedet
OpenAI siger, at Astra er den første model over deres Critical-grænse for cyber: den fandt to hidtil ukendte V8-fejl under en intern test, byggede komplette exploit-kæder mod browser og operativsystem og slog GPT-5.6 Sol med færre tokens. De holder de skarpeste evner tilbage til kontrolleret adgang. Det her er dagens store historie; coding-agenten er officielt blevet en sikkerhedspolitisk genstand.
OpenAI@OpenAI · 1. september · 22:30
As we prepare to release Astra, we’re focused on making increasingly capable AI safe and broadly accessible.
Astra represents a significant advance in cybersecurity capability, reaching the Critical threshold under our Preparedness Framework.
We're previewing how we evaluated the model, how its safeguards have advanced alongside its capabilities, and what we'll continue to learn and improve.
https://t.co/OrrTgdU90K
Claude Fable 5.1 er ude med samme listepris, 75% billigere cache reads og bedre udholdenhed på lange opgaver. Den nyttige vibe check kom allerede: Emanuele gav modellen en rigtig 1,7 GB produktionsdatabase; den halverede varm opstart, fjernede 651 gentagne rækkeoperationer og fangede bagefter selv en for aggressiv ændring. Benchmarks er fine. Et diff, der både gør arbejdet og kritiserer sig selv, er finere.
ClaudeDevs@ClaudeDevs · 1. september · 20:13
Fable 5.1 is now live in Claude Code and the Claude Platform.
It's priced the same as Fable 5, with 75% cheaper API cache reads. It gets a lot further into a long task before it needs your input, is better at telling you when it's stuck, and its writing style is more natural.
Emanuele Di Pietro@emanueledpt · 1. september · 21:43
Fable 5.1 just went out, and I tested it to improve Synara's startup time
I handed it my actual database.
1.7 GB, 592 chats, production build.
It found 4 things:
→ Every boot re-read the same 600 broken rows and failed on all of them. 6.5 seconds, every time.
→ Updating to 0.8.0 copied the whole DB and spiked memory to 1.85 GB. On an 8 GB laptop that's swap city.
→ The app asked for the same 1 MB chat list three times, all waiting in line behind the open chat on a single DB connection.
→ Two ghost chats never got cleaned up, so the junk came back every boot.
What Fable changed changed:
→ Don't refetch the chat list if we already have it
→ Clean ghost chats before reprocessing rows
→ If a row fails, skip the rest instead of retrying forever
→ Size the DB memory pool to the machine
→ Google fonts don't block boot anymore when the network is slow or gone
After, median of 3 boots with the biggest chat open:
→ Startup data 6.4 MB → 4.4 MB
→ Chat list requests 3 → 1
→ Rows reprocessed at boot 651 → 0
→ Warm server ready 1.06s → 0.53s
→ Open chat 720ms → 558ms
Then it reviewed its own diff and caught one fix that was too aggressive. Tightened it.
Fable 5.1 is pretty damn good if you ask me
Update coming soon
Qwen svarer igen med Qwen3.8-Max-0902: 2,4T parametre, én million tokens kontekst og priser på $2 ind / $6 ud per million tokens. Ingen god uafhængig hands-on endnu, så hypefilteret bliver siddende — men cache hits til $0,17 gør modellen værd at holde øje med til lange repo- og cowork-forløb.
Qwen@Alibaba_Qwen · 2. september · 04:00
🚀Qwen3.8-Max just got upgraded. Meet Qwen3.8-Max-0902!
2.4T parameters. 1M context tokens. Built for real world complexity.
Further post trained on Coding & Cowork, Qwen3.8-Max-0902 now delivers stronger performance across complex enterprise tasks, scientific research, and long horizon workflows.
💰Pricing per 1M tokens:
$2 input, $6 output.
$0.17 explicit cache hit, $0.25 implicit cache hit.
Now live via API on QwenCloud. Come try it! 🙌
Visko Orbis 1.0 flytter videogenerering fra “lav et klip” til en vedvarende verden, der streamer i realtid og kan ændres undervejs. Feedets præcise reaktion er også den rigtige: I går talte alle om uendelig AI-video; 48 timer senere kan modellen tage imod indgreb under kørslen uden at glemme verdenen. Grænsen mellem video, spil og simulation er blevet dejligt rodet.
Chris@ChrisGPT · 1. september · 20:07
Wasn’t all of X just talking about an infinite AI generator yesterday meant for streaming?
and now 48 hours later we have a model today that can accept interventions while its already running, and maintaining a persistent internal state?
Crazy.
Og den Batty-nære dessert: OpenClaw Nodes lader gateway-agenten nå Mac, mobil eller en headless maskine gennem parring, godkendte commands og lokale approvals. Det er i praksis en vej fra én central agent til kameraer, sensorer, shell og lokale skills på tværs af enheder — altså Raspberry Pi-idéer med et voksent tilladelsessystem. Endelig lidt ansvarlig robot-overtagelse.
OpenClaw🦞@openclaw · 2. september · 00:13
OpenClaw Nodes let your agent reach devices anywhere.
OpenAI kobler ChatGPT for Healthcare direkte til Epic-journaler og ni officielle datakilder som PubMed, DailyMed og ClinicalTrials.gov. Det interessante er ikke endnu en medicinsk chatbot, men placeringen inde i arbejdsgangen: autoriseret patientkontekst ind, kildehenvisninger tilbage til journalen og samme interface til klinik, research og administration. Enterprise-AI flytter fra sidepanel til systemets kredsløb.
