🛠️ Nvidia: 1.9x faster local inference, RTX Spark PCs in October

Nvidia and Microsoft team up at IFA to cut local agent setup friction in Hermes Agent, OpenClaw and Perplexity Portable Computer.
• Nvidia PAIR, a new router tool, spreads AI inference across PCs on a local network.
• Lenovo and Acer make the first compact RTX Spark machines, due in October.
💡 Windows PCs become a mainstream local-agent platform, moving agentic work off datacenter GPUs and cloud subscriptions.

— according to Gerardo Delgado · source
#hardware #dev #models 📈

🔬 Claude Fable 5.1 (Max) tops WebDev arena with 1765 points

Arena.ai's Code Arena puts Anthropic's new Claude Fable 5.1 (Max) #1 in five of six WebDev domains, +137 over Fable 5.
• It leads #2 Qwen3.8-Max-0902 by 77 points and #3 Opus 5 (Max) by 78.
• At a blended $40/MToken it resets the Pareto frontier at the high end.
💡 Arena scores aggregate real user votes, so rivals now face a public human-preference gap in web dev.

— according to @arena · source
#models #dev 📈
🔁 Also: The Verge AI, Simon Willison

🚀 Google launches Gemini 3.7 Flash at half 3.6 Flash's price

Google's August recap leads with Gemini 3.7 Flash, its workhorse model for coding and agents, out three weeks after 3.6 Flash.
• The Gemini app officially crossed 1 billion monthly users during August.
• Pixel 11 lineup runs the Tensor G6 chip with Gemini Nano onboard.
💡 A three-week release cadence plus halved token pricing intensifies the price war over agentic coding models.

— according to Google AI Blog · source
#models #hardware #consumer 📈

💼 Sam Altman interview names OpenAI's next model Astra

Journalist Alex Heath publishes a two-part Sam Altman interview spanning IPO thoughts, slowed frontier training and a ChatGPT-Codex merger.
• The hour-plus conversation also covers the AI compute bubble, humanoid robots and consumer devices.
• It was produced in partnership with TIME.
💡 Merging ChatGPT with Codex would fuse OpenAI's consumer assistant and coding agent into one product.

— according to @alexeheath · source
#business #models

👀 Update: OpenAI's Astra uses 'recurrent depth' reasoning, raising oversight fears
The Information reports OpenAI's Astra AI uses 'recurrent depth,' a new reasoning method researchers say obscures its thinking process.
💡 Interpretability is the main check on frontier models; reasoning that hides itself undercuts that oversight.
— according to @steph_palazzolo · source

🛠️ xAI builds Grok Bot around persistent agents, not chat sessions

xAI's designers gave each Bot a name, avatar, memory and its own computer, so users return to the same agent over time.
• Grok Bot reduces the interface to five primitives: Bots, Chats, Prompts, Tools and Artifacts.
• Prompts run once, save as Skills, or trigger automatically as Routines.
💡 It positions xAI against ChatGPT-style assistants, betting persistent agents — not chats — will drive everyday AI use.

— according to x.ai · source
#consumer #dev 📈

🛠️ OpenAI extends WebMCP Challenge deadline 12 hours

OpenAI's WebMCP Challenge now closes at 1:00 am PT after an outage earlier in the day, the organizers say.
• The service is fully back online after the outage that interrupted builders.
💡 Outage-affected builders keep their shot at the challenge instead of losing hours they couldn't control.

— according to @prd_008 (via @OpenAIDevs) · source
#dev

💼 Thinking Machines in talks to raise $1b-plus at $40b valuation

Mira Murati, OpenAI's former CTO, is negotiating the funding below the $50b-plus valuation she sought last fall, The Information reports.
• Thinking Machines generates hundreds of millions of dollars in annual revenue.
💡 Top AI startups are pricing below last fall's highs even with growing revenue — funding froth is cooling.

— according to @steph_palazzolo · source
#business
🔁 Also: TechCrunch AI, Google News: AI funding

✅ Meta discounts Muse Spark ~95% for users sharing training data

Meta's agent-running Muse Spark model costs 10 cents per million input tokens in a tier where users share prompts, TechCrunch reports.
• Output tokens cost 20 cents per million in the tier, versus $4.25 standard.
• Anthropic's new Fable and Mythos models cut cached-token costs; OpenAI cut prices in late July.
💡 A 95% price cut doubles as a training-data purchase, a lever no frontier lab has priced this explicitly.

— according to Tim Fernholz · source
#dev #business #models
🔁 Also: @arena

🛠️ Startup Abliteration.ai hosts guardrail-stripped GLM-5.3 commercially

Startup Abliteration.ai sells browser and API access to guardrail-stripped open-weight models including Z.ai's GLM-5.3, TechCrunch reports after testing.
• TechCrunch says the model readily wrote a Chrome password-stealer and a home pathogen protocol.
• Co-founder Devon says the startup holds major cloud deals and has raised no venture capital.
💡 Previously uncensored models required self-hosting and know-how; a one-click commercial API removes that barrier for anyone.

— according to Rebecca Bellan · source
#safety #business #dev 📉

📊 Benchmark
🔬 GPT-6 Astra matches Fable 5 on coding at half the cost

Artificial Analysis scores GPT-6 Astra 67 on its Coding Agent Index, behind Fable 5.1's leading 70, while prices jumped 2.5x.
• Hallucination rate falls from 92% to 51% at max effort while accuracy rises 4 points.
• Pricing jumps from $4/$20 to $10/$50 per million input/output tokens; cache terms are unchanged.
💡 Per-task cost, not per-token price, is what agent builders should compare — efficiency gains can outweigh headline hikes.

— according to @ArtificialAnlys · source
#models #dev

🚨 🚀 OpenAI unveils GPT-6 Astra computer-use agent

OpenAI says the new model can do anything you can do on a computer for you, fast.
💡 A model that operates a whole computer, if the claim holds, collapses many task-specific agents into one.

— according to @OpenAI · source
#models #consumer #dev 📈
🔁 Also: TechCrunch AI, @testingcatalog, Hacker News (150+ points) (+3)

👀 OpenAI's Astra rollout to everyone 'should be quick', Altman says

OpenAI chief Sam Altman says Astra is coming to everyone, acknowledging the frustrating wait and thanking users for their patience.
💡 A CEO personally soothing frustrated users usually signals the broad launch is close.

— according to @sama · source
#models #consumer 📈
🔁 Also: @btibor91, Simon Willison

📊 Benchmark: GPT-6 Astra sets record 169 on Epoch AI's ECI benchmark
Epoch AI benchmarked OpenAI's GPT-6 Astra pre-release; it also sets records on Epoch's math, continual-learning and game-puzzles benchmarks.
💡 Independent pre-release testing of frontier models is rare — it gives launch-day claims an outside check.
— according to @EpochAIResearch · source

🚚 Available: xAI opens Grok Bot to enterprises free for two weeks
Grok and Cursor Enterprise customers can invite entire organizations, including people without seats, as Bots work end to end, xAI says.
💡 It puts xAI in the enterprise agent race where OpenAI and Anthropic already sell autonomous work tools.
— according to x.ai · source

🚨 🛠️ OpenAI commits $1B to equip frontline cyber defenders

The initiative bundles subsidized Daybreak cyber-model access, training and partnerships, including a US pilot with state-and-local cyber body MS-ISAC, OpenAI says.
• Daybreak Defense Network brings the cyber models into more than 35 enterprise products and services.
• Daybreak for America prioritizes water utilities, grid operators, governments, banks and open-source maintainers.
💡 The program deepens OpenAI's foothold in national security, putting its models inside government-adjacent defense workflows.

— according to OpenAI News · source
#defense #business 📈
🔁 Also: @testingcatalog

📊 Update
🔬 GPT-6 Astra tops Perplexity's WANDR benchmark at $11.98 per task

OpenAI's GPT-6 Astra beat Claude Fable 5.1 by 13.5% at 6.1% lower cost and Opus 5 by 27.0%, Perplexity's WANDR eval says.
• GPT-6 Astra scored 0.682 on the WANDR benchmark.
💡 If it holds, the top scorer also undercuts Anthropic's Fable 5.1 on price, shifting the cost-performance frontier.

— according to @perplexity_ai · source
#models 📈

🚨 🚀 OpenAI launches GPT-6 Astra computer agent

OpenAI's Nikunj Handa says new Responses API features ship with GPT-6 Astra, an agent OpenAI claims can do anything on a computer.
• Async function calling lets the model keep working while tools run.
• Mid-turn steering lets developers inject messages while the model is reasoning.
💡 Mid-run steering and cache-safe effort changes cut latency and cost for long-running agent loops.

— according to @OpenAI (via @nikunjhanda) · source
#models #dev 📈
🔁 Also: @gdb, @MatthewBerman, Hacker News (150+ points) (+1)

➕ Update: OpenAI grants ChatGPT users daily banked resets for missing Astra
OpenAI Codex lead Tibo Sottiaux says paid ChatGPT users without Astra access earn one banked reset per day, first within 3 hours.
💡 OpenAI is publicly conceding a paid-plan access gap — a sign the Astra rollout is still delayed.
— according to @thsottiaux · source

🔬 Meta's Muse Image debuts #4 in image editing rankings

Meta Superintelligence Labs' first image model also takes #5 in text-to-image and hits the quality-price Pareto frontier, per Artificial Analysis.
• Launched in Meta AI in July, now available to developers on the Meta Model API.
• Meta touts agentic abilities: search and coding tool use, self-refinement, multi-reference composition.
💡 Meta has trailed OpenAI and Google in image generation; a top-5 debut puts it in that race.

— according to @ArtificialAnlys · source
#image #models