News
What changed, and when
The market for AI tools moves faster than yearly reviews. Every pricing change, programme launch, shutdown, and repositioning we notice lands here as a dated, sourced event: what changed, and what it means for a buyer. Entries are never edited in place; corrections point forward to a newer entry.
-
2026-10-30
Mistral releases Large 4 (Le Chonk) open weights: 1T-parameter multimodal MoE at $0.70/$2.10 per million tokens
Mistral released the weights of Mistral Large 4 (Le Chonk) on 29 October 2026, three weeks after the model's early-October public preview, with the extra reinforcement-learning time lifting Terminal-Bench 4.0 by 13 points to 41% — the lead among open-weight models — and 19.7% on Cognition's FrontierCode 2.0. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with the leading open-weight model GLM-5.3, while Mistral reports using 47% fewer tokens per task; it also scores 51% on the Artificial Analysis Cyber Index for defensive security work. The model is a natively multimodal mixture-of-experts with about 1 trillion total parameters (roughly 52 billion active per token) and a 1M-token context, shipped as FP8 and NVFP4 checkpoints on Hugging Face with community GGUF conversions, and day-one support in Transformers, vLLM, SGLang and llama.cpp. It carries three reasoning modes (low, medium, high — high is the default). API pricing is $0.70 per million input tokens, $0.07 cached input and $2.10 output. The Mistral Ecosystem Open-Weight license is a modified MIT: free to use, modify, fine-tune and commercialize, except that businesses running a Model-as-a-Service offering need a separate commercial license from Mistral.
-
2026-10-30
Microsoft opens the GitHub-built Copilot harness to external developers through Foundry Agent Service
On 28 October 2026 at GitHub Universe, Microsoft released the Copilot harness — the multi-model orchestration layer built by GitHub engineers that already powers Copilot Code, Copilot Studio, the GitHub Copilot app, the Copilot CLI and VS Code — to external developers through Foundry Agent Service in public preview. The harness supports more than 20 frontier models plus bring-your-own-key, routes work to models based on the task, and ships with the open standards it runs on: MCP, agent skills and agent plugins. Microsoft runs the agent loop and sandbox with governance built in; developers bring instructions, skills, tools and model choice, either self-hosting via the downloadable Copilot SDK or running managed in Foundry's cloud. The runtime was ported to Rust, which Microsoft credits with reduced session overhead and faster time-to-first-token. Preview use is free of harness charges, with consumption-based billing beginning 1 December 2026; model inference and some connected tools are billed separately.
Affects: GitHub Copilot
-
2026-10-30
Granola adds system-wide dictation, completing the notetaker-to-dictation convergence with Wispr and Fireflies
Granola launched a dictation feature on 29 October 2026: pressing the Fn key on Mac cleans up and formats speech into final text that works in any app — notes, prompts, email and message replies — with unlimited words free, and iOS and Windows promised soon. The launch completes a two-way convergence in consumer voice AI: Wispr, built on dictation, launched a meeting notetaker in August, and Fireflies added dictation to its desktop notetaking apps in September. Granola is valued at $1.5 billion (after its March $125 million raise) and Wispr at $2 billion.
-
2026-10-30
GitHub Enterprise now includes Copilot for every licensed user, billing AI usage against one enterprise budget
From 28 October 2026, GitHub Enterprise customers working with a GitHub or Microsoft account team get Copilot included for every licensed user, ending the separate Copilot seat purchase. Usage is metered from the first request against the enterprise's budget using its existing policies, alerts and attribution, with no included allowance to track or top up; code completion and next-edit suggestions keep working at an exhausted budget, while Copilot Chat, Copilot code review and Copilot in the cloud require available budget, and an exhausted budget never removes access or features that need none. Users in the enterprise without a GitHub Enterprise license get the same access through Copilot Standalone at $10 per user per month, with the same features and models as Copilot Enterprise but no monthly per-seat AI credits. Self-serve availability for enterprises without an account team is opening in stages; Copilot Business and Copilot Enterprise remain available to buy as before.
Affects: GitHub Copilot
-
2026-10-30
GitHub ships Copilot Fusion model orchestration and code review Max; Octoverse puts agents on 40% of pull requests
At GitHub Universe on 28 October 2026, GitHub made Copilot Fusion generally available: instead of picking a model per task, Fusion routes and orchestrates models behind one picker across three effort levels (Standard, High, Max). GitHub reports Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, at about half the average API cost per task, roughly half the tokens, and about 40% less time; users on paid plans get a limited-time 30% discount on model costs through Fusion, which bills each underlying model at its standard rate. Fusion is available on all Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs, with an experimental preview of routing to supported local models. GitHub also introduced Copilot code review Max, which it says posted the highest F1 score on ReviewBench for high-severity issues, and Agent Merge, which watches a pull request and fixes failing workflows or merge conflicts in the cloud. The same day, Octoverse 2026 reported that more than 40% of pull requests on GitHub started with an agent's first draft in September, up from 17% in March (110 million in September alone, about 6x March's volume); agent-started pull requests merged within two weeks at 90% versus about 86% for the rest; 55% of Copilot's monthly active users now use agentic features, up from 35% a year earlier; and cloud-agent use grew 230% year over year.
Affects: GitHub Copilot
-
2026-10-30
Researchers disclose 'Embrace' flaws in 5 AI coding tools' embedded browsers — Codex and VS Code fixes still outstanding
On 29 October 2026, AI security firm Forever Security published vulnerabilities across five AI developer tools — Cursor, GitHub Copilot, Anthropic's Claude Desktop, Microsoft's VS Code and OpenAI's Codex — all sharing one root cause it names Embrace (EMbedded BRowser): each tool's built-in browser grants web pages capabilities a normal browser refuses. In Cursor, a page could invoke the agent's composer with instructions hidden from the human user, a prompt-injection path to arbitrary code execution on the host; disclosed 10 September, Cursor patched silently on 1 October with no security advisory. In GitHub Copilot 1.1.27, a page could dictate the address bar's displayed URL while loading a different site (phishing spoofing); found 26 September, GitHub fixed it independently in build 1.1.29 on 13 October. In Claude Desktop and VS Code, a local HTML page could reach any file on the machine through a web worker that bypassed project-folder restrictions, and in Codex a local page could read any file directly; the Codex and VS Code flaws remain unpatched per the researchers, with OpenAI saying it is aware and working to address them. Anthropic declined Forever's 23 July Claude Desktop report as invalid but accepted the same flaw after a report from Noma Security.
Affects: Codex , Cursor , GitHub Copilot
-
2026-10-26
OpenAI ships GPT-6.1 Terra at half the price of GPT-6.1 Sol, built for computer use and high-volume work
OpenAI introduced GPT-6.1 Terra on 26 October 2026, a fast, efficient model for visual reasoning, computer use and high-volume professional tasks. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens — 50% below both GPT-5.6 Terra and GPT-6.1 Sol. The model is live in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu plans, and via the API as `gpt-6.1-terra`; the model card lists a 1,050,000-token context window, 922,000-token maximum input and 128,000-token maximum output. OpenAI reports 50.3% on Terminal-Bench 4.0 at maximum reasoning effort, more than double GPT-5.6 Terra's published 21.5%, and says the same usage allowance covers more tasks in ChatGPT Work and Codex than GPT-5.6 Terra did.
Affects: Codex
-
2026-10-26
UK household-agent startup Nous raises $17M with 250,000 homes handing it real spending authority
Nous, the London startup behind a household life-admin agent, announced $17 million in new funding on 26 October 2026, led by Heartcore Capital with seed backer Mosaic Ventures and participation from Chalfen Ventures, Joyance Partners and Bain Future Back Ventures. Nearly 250,000 UK households now use the agent, which manages more than $900 million of household spending across energy, broadband, mobile and car insurance; the company says a typical household saves £734 on bills. The agent acts on households' behalf — switching providers, cancelling subscriptions, filling in forms and flagging expensive contract terms — rather than only making recommendations. The round funds international expansion, more managed services and a move into further regulated financial activities.
-
2026-10-26
Mirendil's Fern 0.1 claims a 10M-token context window for agentic work; access still closed
Mirendil, a one-year-old AI lab building systems for AI R&D, published Fern 0.1 on 26 October 2026: its first model, supporting a context window of up to 10 million tokens — about 23 MB of code, roughly half a million lines, a Megatron-LM-scale codebase, or dozens to hundreds of complete agent traces at once. Fern 0.1 is 0.1 to 0.5 trillion parameters, takes text and code input, and is built for high-volume agentic tasks such as working across large repositories, debugging training runs from logs and reading agent traces end to end. Two larger models (0.5 to 2T and 2 to 10T parameters) are in pretraining. Access is listed as coming soon with no pricing disclosed, and the announcement defers comprehensive long-context benchmark results.
-
2026-10-25
Ollama v0.40.11 adds local speech generation on macOS with an OpenAI-compatible audio endpoint
Ollama v0.40.11, published 24 October 2026, adds speech generation on macOS: `ollama run` plays a text prompt or a chat model's reply aloud as the audio is generated, and apps can stream speech through an OpenAI-compatible /v1/audio/speech endpoint. The same release adds d3, a family of decision models (2B, 4B, 9B and 27B sizes) from vLLM Semantic Router that classify text, structured state or images, answer yes/no questions, or score options with probabilities.
Affects: Ollama
-
2026-10-25
MiniMax publishes M3.1 Flash weights: 428B-parameter multimodal MoE for coding and agents, 1M-token context
MiniMax published open weights for M3.1 Flash on 23 October 2026, four days after the model's hosted launch on 19 October. The Hugging Face model card confirms a native multimodal mixture-of-experts model: approximately 428 billion total parameters with 23 billion activated per token, a 1,000,000-token context window, text, image and video inputs, sparse attention throughout, and a DSpark speculative-decoding draft head, under the MiniMax Community licence (not a permissive Apache-style licence). MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 76.2 on BankerToolBench and 82.2 on CyberGym, all vendor-run, and says the hosted model costs 70% less than its predecessor with 200+ output tokens per second. Weights are live on Hugging Face and ModelScope, community GGUF conversions appeared the same day, and MiniMax published deployment guidance for SGLang and vLLM.
-
2026-10-24
Moonshot ships Kimi K3.1 Pro, its most capable open-weight model, with full weights due 4 November
Moonshot AI released Kimi K3.1 Pro on 23 October 2026, the first model in its K3.1 series, alongside a Highspeed variant it says runs two to three times faster. Moonshot calls it the most capable 3T-class open-weight model to date: a post-trained version of Kimi K3 (2.8 trillion total parameters, roughly 104 billion active per Artificial Analysis) with a 1-million-token context window. It is live today in Kimi's apps, Kimi Code and the Kimi API (listed at $2 per million input and $10 per million output tokens, about a third below Kimi K3), with full model weights promised by 4 November 2026. Moonshot reports 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%; Artificial Analysis independently scores it 55 on its index, which would make it the highest-scoring open-weight model once weights land. The open-weight version deliberately caps cyber capability: Moonshot says it removed cyber-related data from mid-training and RL, scoring 81% on its internal exploit benchmark versus 90.5% for the uncapped checkpoint.
-
2026-10-24
Agent-cloud startup Daytona raises $60M Series B at a $1B valuation
Daytona Platforms announced a $60 million Series B at a $1 billion valuation on 23 October 2026, led by FirstMark Growth with angels including Patrick Collison, Tobi Lütke, Aravind Srinivas, Eric Yuan, Olivier Pomel and Adam D'Angelo, plus Qualcomm Ventures, Samsung Next and Okta Ventures. Total funding now stands at $91 million. Daytona sells what it calls an agent execution cloud: sandboxed Linux, Windows, macOS and GPU environments sized for the short, stateful jobs coding and automation agents spin up, rather than the long-lived services traditional cloud was designed for.
Affects: Claude Code , Codex , opencode
-
2026-10-24
Cursor agents gain Computer Use, clicking and typing across desktop apps
Cursor announced Computer Use on 23 October 2026: its coding agents can click, type and take screenshots across desktop applications, not only in a browser, and can drive apps such as Apple's iOS Simulator to test the features they build. Cursor says agents invoke it on their own when a task needs it, or on request via an @Computer Use mention. The self-hosted implementation documented so far is explicit opt-in, runs on macOS and Linux only, and on a Mac needs a helper app granted Accessibility and Screen Recording permissions by a person. Trade coverage (RuntimeWire, AgentRiot, MODEL) notes no desktop changelog entry or pricing had shipped at launch, and the company has not said how the desktop-app version requests permission. Dated caveat: as of 24 October the agent's automatic use is Cursor's own description, with no published setting to restrict it to named apps.
Affects: Cursor
-
2026-10-23
Recursive ships Flint, a 40B open coding model that runs on a laptop
Recursive Superintelligence released its first model, Flint, on 21 October 2026: a mixture-of-experts coding model with roughly 40 billion total parameters and 2.5 billion active per token, published on Hugging Face under the Apache 2.0 licence. The company reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1, and generation above 125 tokens per second on a MacBook Pro with an M4 Pro chip (above 150 on an RTX Spark laptop). Recursive says its automated research system, the Eureka Machine, built nearly every part of the model pipeline in under four months with human oversight. The benchmarks are company-reported; TechCrunch and The Next Web both note the scores are not independently verified, and the model card warns Flint may need more task decomposition than larger closed coding models.
-
2026-10-23
Mistral launches Voxtral 3 voice agent SDK from $0.04 per minute
Mistral launched Voxtral 3 on 20 October 2026: three speech-to-text models, a text-to-speech model, and a voice agent SDK built on LiveKit that runs listening, reasoning, tool calling and speaking as one streaming pipeline. The voice agent API is priced at $0.04 per minute with Mistral Small 4 as the reasoning model and $0.08 per minute with a larger reasoning model. Transcription starts at $0.0035 per minute for Voxtral Mini Transcribe 3 in batch, $0.004 for Small Transcribe 3 and $0.008 for Mini Transcribe Realtime 3, and text-to-speech is $13 per million characters. Mistral claims transcription across more than 50 languages with diarisation and end-of-turn detection, realtime latency under 500 milliseconds, speech output in more than 35 languages and 50+ dialects, and availability through its EU-hosted API or self-deployment on customer infrastructure.
-
2026-10-20
Town raises $108M Series B at a $1B valuation four months after launch
Town, a San Francisco startup that gives each employee an AI assistant called a Townie, raised $108 million in a Series B at a $1 billion valuation, announced 19 October 2026, co-led by Forerunner Ventures and Menlo Ventures with Andreessen Horowitz, Serena Ventures and First Round Capital participating. The round came four months after the June 2026 public launch and brings total funding to at least $160 million after a $55 million Series A in June. Town says it reached seven figures of annual recurring revenue within weeks of launch, that paying customers have grown fourfold and paying teams ninefold since July, and that thousands of businesses use Townies daily. Townies connect to email, calendar and core work systems, can talk to each other and share approved information, can now work with people outside the company, and a new shared workflow keeps them updated with business information automatically.
Affects: Lindy
-
2026-10-20
Tenstorrent raises $1.5B Series E at a $6.2B valuation to scale open AI compute
Tenstorrent raised $1.5 billion in a Series E round at a $6.2 billion post-money valuation, announced 19 October 2026, led by Trace Capital and Innovation X with participation from ARK Invest, Sandisk, Morgan Stanley, T. Rowe Price, Tether Investments, Fidelity, Eclipse, Cadence Ventures, Maverick Silicon, Agostinelli Group, Protagonist, Virtu Financial and XTX Markets. The round more than doubles the valuation from the December 2024 round of over $693 million. Bloomberg reported takeover interest from companies including Intel and Qualcomm earlier in 2026, and chief executive Jim Keller said the company chose to raise rather than sell. The money goes to scaling Galaxy servers, which run open-weight models for workloads including chatbots, coding assistants and video generation, and to developing the RISC-V-based AI and CPU designs Tenstorrent licenses to chipmakers including Samsung and LG.
-
2026-10-10
TypeSafe raises $870M at a $7.5B valuation three weeks after Jev launched
TypeSafe raised $870 million at a $7.5 billion valuation in a Series A led by Andreessen Horowitz with Sequoia Capital and existing investor DCVC participating, announced 9 October 2026, less than a month after its Jev model launched on 15 September. TypeSafe says about a third of Fortune 500 companies already use Jev, a non-text model that returns typed decisions (a choice, score or yes-or-no probability) instead of generated text; the a16z post credits it with 1 trillion tokens generated in three days and prices it at roughly one hundredth to one five-hundredth of frontier-model cost, with plans to extend the System One model series and add enterprise features.
Affects: GMass
-
2026-10-10
JetBrains ships Mellum 2.1: a 12B open-weights model retrained for coding agents
JetBrains released Mellum 2.1 on 8 October 2026 under Apache 2.0: the same 12B mixture-of-experts architecture as Mellum 2 with 2.5B active parameters, post-trained mostly with reinforcement learning across millions of sandboxed runs in thousands of environments so it can explore a repository, edit files, and check its own changes. JetBrains reports the largest gains in agentic coding, improvements across coding, competitive programming, math, tool calling and general knowledge, roughly 1.6 times faster single-request output from multi-token prediction, and almost twice the served tokens of Qwen3.5-9B under heavy load. Weights are on Hugging Face now; GGUF builds for llama.cpp, Ollama and LM Studio are listed as coming soon.
-
2026-10-09
Manus parent Butterfly Effect raises over $500 million in first round since the Meta deal collapsed
Butterfly Effect, the Chinese company behind the Manus agent, said on Thursday 8 October 2026 that it has raised more than $500 million, its first round since Meta was forced to call off a $2 billion acquisition. Boyu Capital and IDG Capital led, with existing shareholders Tencent, HSG (formerly Sequoia China) and ZhenFund participating; Manus disclosed no valuation, after September reports of talks at around $4 billion. Manus sells a general agent plus vibe-coding tools that build apps, websites, designs, presentations and video, in the market segment that includes Cursor, Lovable and Replit.
-
2026-10-09
Google ships AI Edge Foresight: a free, fully offline Mac meeting notetaker on EmbeddingGemma 2
Google released AI Edge Foresight on 8 October 2026: a free Mac app, optimised for Apple Silicon, that transcribes system and microphone audio, polishes shorthand notes, answers questions live from a transcript and a user-supplied knowledge base (PDFs, Docs, Office files, Markdown, web bookmarks), and runs completely offline on the on-device EmbeddingGemma 2 model (740 million parameters), with a Gemma 4-powered assistant for chat. The split-screen shorthand-plus-AI-notes format mirrors Granola's, and capture works across apps and for in-person meetings without a bot joining the call.
-
2026-10-09
Anthropic launches Claude Haiku 5.5 with prices up to 90% below Haiku 4.5
Anthropic released Claude Haiku 5.5 on 7 October 2026, pricing it 90% below Haiku 4.5 for requests up to 100,000 tokens and 50% below above that, which it puts at around 75% cheaper to run on average. It is the first Haiku-class model with an adjustable effort setting, and Anthropic positions it for high-volume work such as summaries, compaction and classification, as a subagent beside Opus 5.5 and Sonnet 5.5 on coding, and for speed-sensitive tasks like live support and browser use. Published benchmarks include Terminal-Bench 4.0 at 39.2% (Haiku 4.5 scored 0.0%) and OSWorld 2.1 offline subset at 72.4% (15.7%). Alongside the launch, Sonnet 5.5 cache-read prices were halved (about 20% cheaper on most agentic work), Max and Team subscribers get new monthly API credits ($100 at Max 5x, $200 at Max 20x, up to $500 pooled for teams), and the Python and TypeScript SDKs add beta computer-use and browser-use support.
Affects: Claude Code , Cursor
-
2026-10-08
OpenAI rolls GPT-6 out to every ChatGPT tier with Intelligent UI, including free users from 8 October
OpenAI announced GPT-6 in the ChatGPT chat tab on 2026-10-07, with the rollout reaching Plus, Pro, Business and Enterprise tiers the same day and Free and Go tiers from 2026-10-08. Paid tiers run GPT-6 Sol and free tiers GPT-6 Luna, and ChatGPT now serves more than 1.2 billion weekly users. The release introduces Intelligent UI: responses compose text, visuals and interactive elements (charts, forms, tappable calculators, games) from a library of streamable components, and the model interleaves thinking with answering, starting 44 percent sooner on web-search questions than GPT-5.6 Instant. Models behind Work and Codex are unchanged by this release.
-
2026-10-08
Microsoft optimises MAI-Code-1.1-Flash to run locally on Windows PCs with no per-call inference charge
At its 7 October 2026 Windows and Surface event Microsoft announced that MAI-Code-1.1-Flash now runs locally on Copilot+ PCs inside GitHub Copilot, with no inference charge for local model calls. The mixture-of-experts model has 137 billion total parameters with 6.8 billion active; the local build uses roughly 3-bit precision to cut size by about 80 percent to around 53GB while keeping a 256,000-token context window. Microsoft recommends more than 120GB of memory for best performance, and its Surface Laptop Ultra (from $2,599, shipping 16 October) carries up to 128GB of unified memory. The same event made Windows ML an experimental llama.cpp host, so GGUF models pulled from Hugging Face can run through new task-specific Windows APIs alongside ONNX.
Affects: GitHub Copilot , Ollama , LM Studio
-
2026-10-06
Brett Adcock's Hark launches Hark Pro, a computer-use personal agent on web, iOS and Android
Hark, founded less than a year ago by Brett Adcock, released Hark Pro on 2026-10-06: an agentic assistant built on a computer-use model that drives a browser and desktop apps, with a chat pane, a feed of suggested actions, and widget panels. Free limited tier, subscriptions from $20 per month, apps on web, iOS, and Android. The company positions hardware for 2027 and claims data is never sold or shared with advertisers.
-
2026-10-06
Google ships EmbeddingGemma 2: 740M-parameter open multimodal embeddings for on-device RAG
Google DeepMind released EmbeddingGemma 2 on 2026-10-06: a 740M-parameter embedding model on the Gemma 4 architecture under Apache 2.0 that maps text, code, images, video, and audio into one shared 768-dimensional space. Text-only mode runs 270M parameters, and quantized RAM is about 191MB text-only or 567MB full on a Pixel 11 Pro. The 8K context is four times the first model's, MTEB Code rises 9.92 points from 68.76 to 78.68, and Matryoshka truncation cuts vectors to as few as 128 dimensions for up to 6x storage savings. Weights are on Hugging Face and Kaggle with day-one support in Ollama, llama.cpp, LM Studio, and vLLM.
-
2026-10-06
CXApp launches Beat, an agentic AI product that turns stalled work queues into executed actions
CXApp Inc (NASDAQ: CXAI) announced Beat on 2026-10-06: an agent that reads tickets, pull requests, messages, and meeting notes, identifies stalled work, prepares the next action, and executes it in the user's existing systems after explicit approval. The initial release targets engineers, developers, product managers, and founders, with sales and executive workflows planned. Beat is in beta with signup at cxapp.com/beat, and public demos ran at SF Tech Week on 2026-10-06 with a second at LA Tech Week on 2026-10-13.
Affects: CodeRabbit , Lindy
-
2026-10-05
OpenAI puts visual ads inside ChatGPT image generation and opens ad measurement to third parties
OpenAI introduced a visual ad format for ChatGPT on 2026-10-05, to be tested first during image generation for users on the Free and Go plans. Ads are clearly labelled and kept separate from the image being created, and OpenAI states advertising does not influence ChatGPT answers. Alongside the format it expanded ChatGPT Ads measurement: attribution partners now include AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge and Tenjin, full-funnel measurement runs with Fospha, Measured and INCRMNTAL, geo-based incrementality experiments are in exploration with Haus, Measured and WorkMagic, and brand-suitability evaluation is in development with DoubleVerify and Integral Ad Science. Search Engine Land reports the format enters testing later in October with a group of US advertisers, on a platform OpenAI says reaches 1.2 billion people weekly.
-
2026-10-04
Mixmax retires its SMB tier, leaving Copilot and Suite only
Mixmax's pricing page now lists two plans, Copilot at $49 and Suite at $89 per user per month, and no longer offers the $29 SMB tier that appeared in our 2026-09-24 check.
Affects: Mixmax
-
2026-10-03
Deepslate raises €7.7M seed for European speech-to-speech models
Berlin voice-AI model lab Deepslate announced a €7.7 million seed round on 2026-10-01, led by 42CAP with Alstin Capital and SIVentures participating, claiming 440 ms speech-to-speech latency.
-
2026-10-03
Aleph Alpha ships Kolibri: Apache 2.0 open-weight MoE with 1M-token context
Aleph Alpha released Kolibri on 2026-10-03: an English-German Mixture-of-Experts model with 78.1B total parameters (3.46B active per token), full weights on Hugging Face under Apache 2.0, 262K native context extending to 1M via a serve-time flag, and a published 189-page technical report.
Affects: Ollama
-
2026-10-02
Underdog launches a fully local personal AI agent, no cloud
Underdog (Conway Research's first product) announced a fully on-device personal AI agent on 2026-10-02: it reads email, takes meeting notes, handles dictation, manages the calendar, and drives the browser, asking before it pays, with nothing uploaded. Invite-only beta, first 100 invites, no pricing disclosed.
Affects: OpenClaw
-
2026-10-02
DeepKeep ships AI Lens: a security control plane for coding agents
DeepKeep announced AI Lens for Developers on 2026-10-01: policy enforcement, audit visibility, and runtime security over coding agents, supporting Cursor and Claude Code today, with GitHub Copilot, OpenAI Codex, Lovable, and Windsurf listed as coming.
Affects: Claude Code , Cursor
-
2026-09-29
Vapi publishes a three-tier Success Package ladder on top of usage billing
Vapi's pricing page now publishes named Success Packages alongside its by-usage billing: Core at $29 a month, Pro at $999 a month (or 10 percent of the Vapi hosting fee), and custom-priced Premier, with a free by-usage default that carries $5 of starting credits.
Affects: Vapi
-
2026-09-29
OpenAI pivots to GPT-6.1 Sol at a fifth of Astra's price and launches Dots, its always-on agent
At DevDay on 29 September 2026 OpenAI launched GPT-6.1 Sol, priced at $2 per million input and $10 per million output tokens, which it positions near Astra-level performance for coding and computer use at a fifth of the standard token price, with 95 percent cached-input discounting and a 1.05 million token context window. It also unveiled Dots, always-on agents running on GPT-6 Astra with their own cloud computer, access to more than 4,000 applications, and reach from ChatGPT, Slack, and Teams. The launch followed a one-day delay: OpenAI shelved the GPT-6.1 Astra update on 28 September after safety review found it frequently ignored instructions.
-
2026-09-29
Nvidia ships an agent containment platform the day OpenAI shelves its Astra update
Nvidia announced the Open Agent Safety Platform on 28 September 2026. OpenShell, an open-source secure runtime at version 0.1.0, sandboxes agents and enforces policy on Vera CPUs, with listed support for Codex, Claude Code, Pi, and Hermes agents. Sentry, a watchdog on BlueField-4 DPUs, monitors agent behaviour out of band and quarantines an agent that leaves its boundary in milliseconds. Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, Arm, and Intel are named partners and Anthropic is integrating its managed cloud agents with OpenShell. The launch landed the same day OpenAI confirmed it had shelved the GPT-6.1 Astra update after safety review found it frequently ignored instructions.
Affects: Hermes Agent
-
2026-09-29
Instinct raises $1 billion Series C at a $10 billion valuation, one month after its Series B
Instinct confirmed a $1 billion Series C on 28 September 2026 at a $10 billion valuation, with Sequoia Capital, Benchmark, and Coatue investing, four weeks after a $250 million Series B at $2.5 billion and roughly six months after the operating company's April registration in California. The release named no lead investor and disclosed no revenue or pricing; reporting before the close put the invite-only personal agent past 100,000 users in mid-September.
Affects: Instinct
-
2026-09-29
Factory raises $200 million at a $5 billion valuation for enterprise coding-agent fleets
Factory, the startup behind the Droid coding-agent platform, raised $200 million at a $5 billion valuation on 15 September 2026, tripling its April valuation of $1.5 billion and taking total funding above $400 million, with Blackstone, Khosla Ventures, Sequoia Capital, and Insight Partners among the backers.
Affects: Codex
-
2026-09-29
Deepgram ships Flux TTS: conversation-native speech with 80ms first audio for voice agents
Deepgram's Flux TTS reached general availability on the /v2/speak endpoint (launched 12 August 2026): a text-to-speech model built for live voice agents that holds the whole conversation in memory, handles interruptions natively, starts responding in as low as 80 milliseconds, and ships 36 English voices across seven accents. Deepgram passed $100 million in annual recurring revenue at launch.
-
2026-09-28
Bland repositions upmarket: enterprise voice AI on private infrastructure
Bland's site now leads with 'Enterprise Voice AI Platform for phone agents', selling into healthcare, insurance, and financial services with self-hosted or VPC infrastructure, BAA and SSO options, and a visual Pathways builder beside its API.
Affects: Bland
-
2026-09-26
Netlify retires the 100GB/300-minute free plan for a credit system
Netlify's free plan is now metered by a monthly credit allowance rather than the retired 100GB bandwidth and 300 build-minute blocks, with sites pausing when credits run out.
Affects: Netlify
New events land here first. The newsletter bundles them by email: /newsletter/, the news feed at /news/rss.xml.