OpenAI's DevDay flagship upgrades GPT-6 Sol on agentic coding, computer use and professional work at $2/M input tokens — a fifth of Astra's standard rates, with cached input at $0.10/M. An Ultrafast variant with up to 8x faster token generation in Codex is coming within days.
Bloomberg reports OpenAI is in talks for at least $30B in a pre-IPO bridge round at a roughly $1.4 trillion valuation, with run-rate revenue up 70% since July to $40B in August. An IPO is now expected next year after Altman ruled out a 2026 debut.
A widely shared Amazon AGI paper introduces AutoGym, which generates a task description, an executable RL environment and a verifier from a small domain seed or past model trajectories. Curators highlight it as evidence that building high-quality RL environments is becoming a core AI-engineering skill.
Asks whether reasoning models can perform hidden computation without revealing it in thinking traces. Shows the answer depends on task difficulty and model size: simple computations can be hidden from CoT monitors, with formal limits on monitorability.
Identifies the PPO critic as a major source of instability in RL for LLMs, pinpointing two critic failure modes. Stabilizing the critic is presented as the key to stable LLM RL training.
Finds that chain-of-thought capacity lives mainly in a thinking model's weight component within a null space defined relative to an Instruct model. A spectral null-space swap transfers reasoning ability while cutting excessive reasoning-token cost.
Targets hybrid LLMs using linear attention (Gated DeltaNet, Kimi Delta Attention) and fixes the accuracy collapse that occurs when their fixed-size recurrent states are quantized to low precision, keeping long-context quality while cutting serving memory.
Coding agents inspect and modify their own instructions, tools and execution procedures without any downstream reward signal, removing the repeated evaluation cost that existing self-improving agent loops incur.
Together AI's latest usage breakdown shows 'Flash'-class distilled open models now dominate coding-agent traffic. Fast, cheap Chinese open models have become the default for running agents.
The discount targets high-volume coding assistants and 'coworking' agents where cost-per-token dominates — a sign that the open-weights race has shifted from raw releases to inference pricing.
A long-time pillar of the open fine-tuning and agent ecosystem integrated ChatGPT sign-in into the Nous Portal for Hermes Agent and adjacent tooling — a small but telling bridge between closed frontier accounts and open agent infrastructure.
Flourish Robots is releasing a phone-trainable home humanoid built on a six-wheel mobile base with two 1.5 kg-payload arms. Preorders for the first 50 units opened September 29 with deliveries in December — the cheapest consumer humanoid announced to date.
Bain & Company's global technology report says existing AI services would cover only $1.2-1.8 trillion of what the industry needs, while data-center spending is projected at $5-6.5 trillion by 2030, adding at least 150 GW of capacity.
The store is staffed by an AI-driven humanoid from Beijing Galaxy General Robot, demonstrating retail automation as mainland China's humanoid deployments accelerate.
Thirty robots from the first graduating class of a Hangzhou training school are being deployed into operational jobs, marking a shift from classroom demonstrations to deployed embodied AI.
CNBC reports that emerging criteria tied to real deployment and revenue rather than demos may be difficult for China's many humanoid startups to satisfy, signaling a consolidation phase for the sector.
In public beta, the Agents API exposes Codex's multi-agent harness as a managed service: sessions, orchestration, context compaction, tool search, MCP servers and code execution. New computer-use support lets agents navigate websites through an OpenAI-hosted browser.
Developers can ship sidebar destinations, interactive panels and file viewers inside ChatGPT using the same platform OpenAI builds its own features on, alongside a new Plugin Creator, a simpler submission flow and MCP Events that trigger automations when connected apps change.
The new service combines OpenAI's agent harness with AWS infrastructure so agents run in the customer's environment with inference on Bedrock, handling inference, memory and skills and integrating with Bedrock AgentCore.
Developers can start and steer tasks by voice and track delegated work via a new /agents view, with Git worktree and session-resume improvements. Codex in the cloud now offers reusable, team-approved environments that run tasks remotely from desktop, web or mobile.
The Milan-based group closed its first external round — a €20M mix of equity and debt led by Nextalia Ventures — to fund a European buy-and-build strategy across DataTech, MarTech and AdTech. It serves 1,000+ European clients with over €20M pro-forma revenues and positive EBITDA.
The largest study of its kind, presented at the EASD annual congress in Milan, showed an AI model can pick up acoustic markers of type 2 diabetes from 20-second voice recordings — positioned as a fast, non-invasive, scalable screening tool.
Samsung Electronics plus five affiliates are investing a combined $1B strategic round in Helix, the KKR-founded AI data center, power and connectivity company led by ex-AWS CEO Adam Selipsky. Total committed capital to the Helix strategy now exceeds $11B.
The round, co-led by Andreessen Horowitz and Bessemer Venture Partners, doubles EliseAI's August 2025 Series E valuation. The company automates administrative work for housing and healthcare and passed $200M ARR this summer.
An extension of its February $30M Series B with AT&T Ventures, Forestay Capital and Quadrille Capital participating brings Reco's total raised to $140M, with a valuation 'more than doubled' since February.
Boston-based Atomic, incubated by DVx Ventures, closed a $12.5M Series A led by Klass Capital with Madrona Venture Group for AI agents that automate supply-chain planning and inventory decisions.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
