New launches.

Promptable voice design with 2,000+ voices and 30-second consented cloning.

Gemini 3.8 Flash TTS and the cheaper Flash-Lite tier turn a text prompt into a designed voice, or clone a real one from a 30-second consented sample, across 100+ languages and hours-long two-speaker scenes. Every generated or cloned voice carries SynthID and C2PA provenance markers. Flash ranked #1 on the Hume Voice Design benchmark; Flash-Lite ranked #2 on Hume Overall Quality despite being the budget tier.

Two cheaper GPT-6 tiers below Astra, priced at half the outgoing 5.6 line.

Sol adds more reasoning headroom for everyday work; Luna is the lightweight, high-volume option built for fast responses at the lowest cost in the GPT-6 family. Both carry Astra's gains in factuality, coding, computer use and alignment down to a cheaper tier, and both are priced at half of what the equivalent GPT-5.6 models cost — Sol at $2/$10 per million tokens versus $4/$20 before. It's the second GPT-6 release in three weeks, after Astra's September 3 debut.

A realtime voice-to-voice model at under half OpenAI's asking price.

Gemini 3.8 Live holds a live spoken conversation, sees what your camera sees and can call tools and APIs mid-sentence while switching among 97 languages. It runs about $0.84/hour at standard settings ($3.50/hour with Extended Thinking) against $5.83/hour for OpenAI's GPT-Live-1 Astra — Google topping its closest rival on benchmark scores while charging less than half as much for the API.

Seven purpose-built agents — Casey, Paige, Carter and four more — ship inside Customer 360.

Salesforce introduced seven named Agentforce agents, each scoped to one business function (support, sales coaching, workforce scheduling and more) and grounded in a company's existing Customer 360 data and business rules rather than a single general-purpose bot. Early customers report billions of agentic work units delivered, pointing at a broader shift from one do-everything agent toward a roster of narrow, auditable ones.

A multi-agent orchestration system sold as a single model.

Rather than one large model, Fugu assembles and orchestrates a swappable pool of expert agents on the fly, learning how to route a task instead of following a hand-designed workflow. Fugu Ultra v2 posts the top or tied-top score on 5 of 8 benchmarks against frontier models it wasn't even trained alongside (its data cutoff predates Fable 5.1 and GPT-6 Astra), at $5 input / $30 output per million tokens; the smaller Fugu Max runs $2/$6. Both are available now through an OpenAI-compatible API.

The Codex agent harness, rented by the API call.

The Agents API is OpenAI's own Codex infrastructure — session orchestration, context compaction over long tasks, sub-agent coordination, lazy tool loading and crash recovery — opened to any developer through four primitives: agent, environment, session, event. It's live in public beta for all developers, with sandbox compute available from OpenAI or partners including Cloudflare, Vercel and Oracle, and no fee beyond the tokens and tools an agent actually uses. It competes directly with the orchestration layer LangChain, CrewAI and similar frameworks occupy: skip building an agent runtime and rent OpenAI's instead.

MIT-licensed 552B MoE model at $0.15 per million input tokens off-peak.

V4.1 Flash is a 552B-parameter mixture-of-experts model on a new causal encoder-decoder architecture, with native image understanding and a 1M-token context window. Weights are MIT-licensed on Hugging Face, and off-peak pricing of $0.15 input / $0.60 output per million tokens undercuts the V4 Flash tier it replaces by half; peak-hour traffic (01:00-04:00 and 06:00-10:00 UTC weekdays) runs $0.30 / $1.20. From September 14 it also absorbs all V4 Pro traffic at Flash pricing until a dedicated V4.1 Pro ships.

A native voice model behind ChatGPT Voice, no text pipeline in between.

GPT-Live replaces the speech-to-text-to-speech pipeline behind ChatGPT Voice with a model that hears and speaks directly, cutting response latency below 300ms and carrying tone and emotional nuance that a text intermediary strips out. It's rolling out now as the engine for ChatGPT's voice mode.

Autonomous AI workers for the enterprise, with access controls.

xAI opened Grok Bot, its autonomous computer-using agent, to enterprise customers with access, network and audit controls it lacked in its August consumer beta. Each bot runs in its own isolated cloud environment, uses a browser and applications the way a person would, has no account access by default, and can repeat a workflow after a person demonstrates and corrects it once. Grok and Cursor Enterprise customers get two weeks free and can invite their whole organization to try it.

Six fully open models, weights to training data, from 0.9B to 375B.

The Institute of Foundation Models released K2 Horizon: six Apache 2.0 models (0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B) sharing one architecture, vocabulary, training method and eval infrastructure, so results are comparable across sizes. IFM is publishing model weights and code plus training data, intermediate checkpoints, configs and logs where licenses allow — a rare level of openness among fully open releases. Day-zero support covers vLLM, SGLang and Ollama across NVIDIA, AMD and Cerebras hardware, with an intended deployment range from wearables (0.9B) up through enterprise serving (375B-A23B).

Google's third Flash release in six weeks, plus a dedicated cybersecurity variant.

Gemini 3.8 Flash is Google's latest workhorse model, improving on 3.7 Flash (released three weeks earlier) across software engineering, agentic tasks and multi-step reasoning in specialized domains; it posts 54.9% on HLE-Verified and leads on benchmarks like Vals Finance Agent V2 and Harvey's Legal Agent Benchmark. Gemini 3.8 Flash Cyber is a separate variant tuned specifically for vulnerability detection, which Google positions as frontier-level for cybersecurity work. Both are generally available now through the Gemini API, AI Studio, Antigravity, Android Studio and Gemini Enterprise, at the same introductory price as 3.7 Flash: $0.75 per million input tokens, $3.75 per million output tokens. The rapid Flash cadence — three releases in six weeks — is Google keeping pace with Anthropic and OpenAI's release tempo rather than a single headline jump in capability.

Runway's first 'Interface World Model' renders app interfaces frame by frame with no code underneath.

Solaris generates a working software interface as live video rather than running code: it renders frames at interactive speeds (under 500ms) at 720p and responds to clicks, drags and voice commands, with nothing resembling a DOM behind it. Runway calls it the first entry in a new 'Interface World Models' category — a research direction, not a shipping product. It isn't public: Runway is collecting early-access requests while it works with select partners, and has disclosed no price or launch date. Text rendering is still error-prone and there's no support for screen readers, so the system is incompatible with accessibility requirements most production software has to meet. It's a signal of where Runway is pointing its video-generation research next, not something a team can use today.

Open-weight multimodal MoE model at a tenth of GLM-5.2's price.

A 320B-parameter, 18B-active mixture-of-experts model that's natively multimodal across text, image and video, with a 1M-token context window and MIT-licensed open weights. Z.ai says it beats GLM-5.2 on its own evaluations for coding, agentic and long-context work while running at roughly one-tenth the cost — list price is $0.15/$0.50 per million tokens, discounted 50% through September 9. It's a serious open-weight option for teams building agentic coding tools who don't want to pay frontier-model rates.

A law-firm-specific edition of Gemini Enterprise with pre-built agents for contract review, diligence and regulatory monitoring.

Gemini Enterprise for Legal packages Google's Gemini Enterprise Agent Platform with legal-specific skills and connectors into the document management, e-discovery and legal research systems firms already run, aimed at executing full workflows — contract review, diligence, regulatory monitoring, privacy requests — rather than just answering questions about a document. It launched in preview on August 25 with Cleary Gottlieb, Freshfields, Weil and Williams & Connolly as named firms, and is Google's first industry-specific packaging of Gemini Enterprise; a financial-services edition shipped alongside it, with healthcare versions described as coming. For firms already inside Google Workspace, this is a narrower, more opinionated product than pointing a general assistant at a contract.

A separate ChatGPT experience for under-18 accounts, with age prediction, parental controls and a rewritten set of behavior rules.

ChatGPT for Teens auto-enrolls accounts OpenAI predicts belong to minors into a distinct experience: default-on restrictions around self-harm, romantic/sexual and relational content (the model can no longer call itself a friend or imply it has feelings), a Study Mode that walks through problems instead of answering them outright, and parent-set quiet hours. Adults wrongly flagged can confirm their age through Persona identity verification. The rollout started August 18 across Free and paid personal plans and is expected to finish within two weeks; enterprise and API products are unaffected. It follows wrongful-death lawsuits over chatbot safety and an FTC inquiry into how AI chatbots affect minors.

An open-weight coding model that jumps sharply on long-horizon tasks without retraining its base.

GLM-5.3 reuses the same 743B base model as GLM-5.2 — every reported gain comes from scaled-up post-training rather than a new base. The jump is steepest on long-horizon coding: Terminal-Bench 3.0 moves from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9. Cybersecurity capability grew faster than Z.ai says it expected as training scaled, reaching 84.5% on CyberGym, which is why the company is staging the release: API access and the GLM Coding Plan are live now, but open weights are held back roughly two weeks for safety hardening.

A 28B open-weight model distilled from Qwen3.8-Max, sized to run on a single high-end GPU or an AI PC.

Qwen3.8-27B is the open-weight, locally runnable sibling to Alibaba's frontier-scale Qwen3.8-Max, released under Apache 2.0 with a native 262K-token context window and text, image and video input. It distills the larger model's agentic coding, computer-use and vision-language gains into a dense 27.78B-parameter form that fits on high-end consumer GPUs or Ryzen AI Max hardware, which is the actual pitch: frontier-family capability without the API bill or the data-residency question. Weights are live on Hugging Face and ModelScope.

Google's fast workhorse model for coding and agents, replacing 3.6 Flash three weeks after it shipped.

Gemini 3.7 Flash is a full replacement of Gemini 3.6 Flash built from the same base via algorithmic improvements and user feedback rather than retraining from scratch, aimed at coding, web development and agent workflows. Google's own numbers put it at 65.3% on DeepSWE v1.1 versus 49.0% for 3.6 Flash. It ships at an introductory $0.75/$3.75 per million input/output tokens — half the prior launch price, good through the end of 2026 — and lands across AI Studio, Android Studio, Antigravity and the Gemini Enterprise Agent Platform, notably ahead of the still-delayed flagship Gemini 3.5 Pro.

A post-training upgrade on Grok 4.5 tuned for long-running agents, matching GPT-5.6 Sol at the same price.

Grok 4.6 keeps Grok 4.5's 500K context window and adds a new xhigh reasoning-effort level aimed at sustaining multi-step work: researching, analyzing, working across codebases and building applications without losing the thread. On the Artificial Analysis Intelligence Index it scores 61, matching GPT-5.6 Sol at max reasoning effort and trailing Fable 5 Max by one point, at $2/$6 per million tokens (doubling above 200K-token prompts). It shipped simultaneously in Cursor, Grok Build, the xAI API and GitHub Copilot, with OpenRouter, Vercel and Cloudflare also carrying it.

A cybersecurity-specialized GPT-5.6 variant, released only to vetted defenders through Daybreak Red.

GPT-5.6-Cyber is built on GPT-5.6 Sol and trained specifically for finding zero-days and building exploit chains, scoring 95% on OpenAI's Advanced Cybersecurity Completion Rate versus 1.5% for standard Sol. OpenAI used it internally to find two previously unknown Chrome V8 vulnerabilities, now patched as CVE-2026-15903. It ships only through Daybreak Red, OpenAI's applicant-vetted defender tier — identity verification, account security requirements, monitoring and approved-use attestations gate access, with Accenture, IBM, CrowdStrike and Cloudflare among the first trusted partners. There is no general-availability path; this is a controlled-access release, not a product launch in the usual sense.

Plus/Pro get a more accurate Sol; free users get unlimited Luna chats and a Think button.

OpenAI updated GPT-5.6 Sol for Plus and Pro users toward more focused answers and less reflexive agreement, reporting 68% fewer factual errors than GPT-5.5 Instant in its own evaluation. Free users move to GPT-5.6 Luna as the default with unlimited text chats, plus a Think button for harder questions. The rollout covers ChatGPT's web, mobile and desktop apps; ChatGPT Work and Codex keep their current Sol version, unchanged.

DeepSeek's production Flash refresh — same 284B-parameter MoE, retrained for a big jump in agentic coding and tool-use.

DeepSeek-V4-Flash-0731 keeps the April preview's architecture (284B total parameters, 13B active, 1M-token context, 384K max output) and re-post-trains it for agentic workflows, landing coding and tool-use scores close to Claude Opus 4.8 at a fraction of the price. A new reasoning_effort parameter (low/high/max) lets callers trade latency for deliberation per request.

Ant Group's hybrid-reasoning model matches rivals 2-3x its size, built specifically for production agent workflows.

Ling-3.0-Flash is a 124B-parameter MoE model (5.1B active) tuned for multi-agent collaboration — different agents divide labor and cross-check each other's output to cut single-model misjudgments. A cluster-level caching system cuts time-to-first-token on long inputs by 60-80%. It's free via OpenRouter and Vercel AI Gateway through 3 August, with full weights open-sourced after that window closes.