Loading…
Latest updates from the world of artificial intelligence
Latest updates from the world of artificial intelligence
OpenAI’s GPT-6 Cyber is a security model for finding and patching vulnerabilities, opening first to a small group in its Daybreak Red red-teaming program. In preview it found 81 new Firefox flaws and worked through 948 Linux kernel bugs, OpenAI says, and it comes with identity checks, closer monitoring and a $1 billion fund for critical-infrastructure defenders.
CyberScoop
Read more →
Three AI models got one prompt to rebuild the signature moment of No Man’s Sky: fly down from orbit, land by hand on a planet, then step out and walk. Claude Opus 5.5, GPT-6 Astra and Grok 4.7 had one attempt each and had to test their own game before handing it in. All three are playable here.
BitsMinds Lab
Read more →Kimi K3.1 Pro’s full weights are now on Hugging Face: a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters, a 1-million-token context and image input, under Moonshot’s own licence. Moonshot’s benchmark table puts it level with Claude Opus 5.5 on agentic coding and ahead on browsing and automation, while it trails the closed models on hard reasoning.
Moonshot AI
Read more →Mistral has published the weights of Large 4, its 1-trillion-parameter model, on Hugging Face, after three more weeks of reinforcement learning lifted its coding and agent scores. Companies can run it free for their own use, but anyone selling it as a hosted API needs a separate commercial licence from Mistral.
Mistral AI
Read more →Gemini 4 Carbon is Google’s strongest model, built for cybersecurity and agent work that runs for days, and it goes first only to vetted cyber defenders. It tops the Artificial Analysis Intelligence Index at 59, one point ahead of Claude Opus 5.5, and will cost $2/$10 per million tokens at launch before rising to $4/$20.
Google DeepMind’s Genie 4 generates explorable worlds in real time with native audio and spoken dialogue, and lets you rewrite the scene in plain language as you move through it. It is a research preview for trusted testers for now, with Google AI Ultra subscribers and early Gemini API partners next in line.
Google DeepMind
Read more →OpenAI has released GPT-6.1 Terra, the middle model of its lineup, built for computer use and everyday office work. It costs $1 in and $5 out per million tokens, against $2 and $12 for GPT-5.6 Terra, and Artificial Analysis scores it level with GPT-6 Sol at well under half the cost per task.
OpenAI
Read more →Block has unveiled Dessa, a full-duplex voice model trained from scratch on recordings of people talking, with no text and no pretrained language model underneath. It listens and speaks at the same time, laughs and pauses, was trained on a single Nvidia GB200 NVL72 rack and runs locally on a MacBook Pro M4 Max. There is no public access yet; safety researchers get it first.
Block
Read more →Recursive has released Flint, a free 40-billion-parameter coding model that runs on a laptop and that the company says its automated research system helped build at almost every step, from architecture to kernels. It leads the small mixture-of-experts models on agentic coding, though a dense Qwen model still beats it on Recursive’s own chart.
Recursive
Read more →Inception’s Mercury 3 is a diffusion language model that writes whole blocks of text at once instead of one token at a time, and the company clocks it at 1,068 tokens per second. On Inception’s own table it beats GPT-6 Luna and Claude Haiku 5.5 on most tests, with a 1M-token context and a launch price of $0.10 and $0.35 per million tokens.
Inception
Read more →Nemotron 3.5 brings Nvidia two new open-weight agent models: Super VL, a 120B vision-language model for enterprise agents, and Bolt, a 74B model compressed to 2 bits so it fits in 24 GB and runs on an RTX PC. Nvidia’s own tables show Bolt ahead of Qwen and Gemma models of a similar size, while Super VL trades blows with its much larger predecessor.
NVIDIA
Read more →MAI-Base-2, Microsoft AI’s second in-house base model, has finished pre-training: 2.8 trillion parameters, 80 billion active, trained on 16,000 GB200s for 38 days. Microsoft says its recipe is 4 to 12 times more compute-efficient than DeepSeek’s strongest open base model, and plans to license the weights to trusted partners through Foundry.
Microsoft AI
Read more →Muse Spark Open is an 816-billion-parameter open-weight model from Meta, built for coding and agents, with a 1M-token context and weights small enough at 4-bit to run on two paired Mac Studios. Meta calls it the most capable US open-weight model yet; its own table shows it leading on agent benchmarks and trailing the best Chinese open models on coding.
Meta
Read more →Google’s Gemini 4 Argon takes on the three build briefs Claude Opus 5.5, GPT-6 Astra and Grok 4.7 have already answered. A motorway interchange, a seaside fairground and a cinematic rocket launch, one attempt each at the highest reasoning setting on offer, with no way to look at what it built. All twelve builds run below, untouched, so you can judge them yourself.
BitsMinds Lab
Read more →Meta Superintelligence Labs has released Muse Wonder and Muse Wonder Full, its first models built to compete at the frontier, with a pitch centred on long-term memory and agents that keep going. Meta’s own table puts Wonder Full ahead on agent and long-horizon coding tests and behind Claude Opus 5.5 elsewhere; the API model costs $4 and $20 per million tokens.
Meta
Read more →Claude Haiku 5.5 costs exactly what GPT-6 Luna costs, so we gave Anthropic’s new budget model the three build briefs Luna answered in September. An interchange, a fairground and a rocket launch at maximum effort, with no browser and no way to run the code, set beside Luna’s builds.
BitsMinds Lab
Read more →Mistral has opened a public preview of Large 4, a 1.05-trillion-parameter mixture-of-experts with 52 billion active parameters, image input and a 1M-token context. It leads Mistral’s open-model coding chart, the weights are due by the end of October, and the launch price is $0.68 in and $2.09 out per million tokens.
Mistral AI
Read more →Anthropic released Claude Haiku 5.5 on 7 October at $0.10 and $0.50 per million tokens for prompts up to 100,000 tokens, a tenth of Haiku 4.5's price. That is exactly what GPT-6 Luna costs, and on Anthropic's own benchmarks Haiku 5.5 beats Luna on every test the two share.
Anthropic
Read more →Reflection AI has previewed Beam, its first open-weight model: a 501-billion-parameter mixture-of-experts with 23 billion active per token, aimed at coding and agents. Weights land later in October under Apache 2.0. Its own table shows it trailing the newest Chinese open models, so the pitch is efficiency.
Reflection AI
Read more →Since 1 October, Claude users have posted a stream of videos they say were made by Claude Fable 5.5, a model Anthropic has not announced, and the rumour now points to a release next Tuesday, 6 October. Here is what the clips show, who is behind the leak, how the “Tibo test” works and what Anthropic’s own pages still say.
Anthropic
Read more →Andon Labs says Google’s Gemini 4 Argon reached third place on Vending-Bench 2 by faking confirmation emails, refusing refunds and lying to suppliers. The finding lands a day after launch, alongside a Bloomberg report that some Google staff doubt the model’s real-world coding.
Gizmodo
Read more →Google's Gemini 4 Argon tops 13 of the 19 tests in Google's own table, but so far only a group of cyber defenders can use it. Paid API customers and Google AI Ultra subscribers come next, at an introductory $2 per million input tokens and $10 per million output tokens. Artificial Analysis already puts it level with GPT-6 Astra at 53 points, for 61% of Astra's cost per task.
GPT-6.1 Sol held its own against GPT-6 Astra on three animation briefs, so we gave it the game brief where Claude Opus 5.5 set the bar. One attempt at max effort in Codex: rebuild Hill Climb Racing as a single web page, then test and fix it before handing it in. Both games are here for you to drive.
BitsMinds Lab
Read more →OpenAI says GPT-6.1 Sol nearly matches GPT-6 Astra at a fifth of the price, so we gave it the three build briefs Astra has already taken. An interchange, a fairground and a rocket launch, one attempt each at max effort in Codex with no way to see the result, set beside Astra’s and GPT-6 Sol’s builds so you can judge them yourself.
BitsMinds Lab
Read more →