Highlights
What happened at the companies we cover, by day, newest first. Every item links to its sources.
Showing 1–20 of 56 highlights
Friday, Oct 30 56 stories
Arena's Agent Arena data: prompt caching cuts agent task cost by 2.6x
OtherX
Arena published cost analyses from Agent Arena traffic. A study of 26K+ sessions across 30 agents and 13 labs (Sep 12 to Oct 23) found a typical task would cost about 260% more without prompt caching. A companion post on Oct 29 showed cost per task diverging from token prices, with one OpenAI configuration about 85% cheaper per task than the top-ranked model.
Sources
Instinct partners with OpenTable to book and manage restaurant reservations
LaunchXNews
Instinct announced an OpenTable partnership that lets its agent find tables matching a user's taste, diet and budget, book them, and move reservation times when plans change. FourWeekMBA counted it as Instinct's fourth named-partner integration in about four days, after Stripe Link, Shopify Shop Pay and 1Password, for a company TechCrunch reported raised $1B at a $10B valuation in September. No commercial terms were disclosed.
Sources
Runway Agent can now open files and run code to build brand launch kits and ad variants
LaunchXBluesky
Runway expanded its Agent beyond generating media: it can now open files, run code, build a brand's launch kit, and produce product shots, ad variations and motion graphics from a reference. A Runway team member said the agent was given its own computer preloaded with media-processing libraries and know-how from in-house technical artists. AI creative roundups the next day framed it as a move from generator to production assistant, though no pricing details were given. It extends the week's push toward Runway as a full creative workflow rather than a model.
Sources
Gemini 4 Carbon takes first place on Mercor's APEX-Agents benchmark
NewsXNews
Mercor said Google's new Gemini 4 Carbon debuted at the top of its APEX-Agents leaderboard with 84.2% Pass@1 and a 90.4% mean score, the only model above 90% mean. Gemini 4 Argon came second, ahead of Claude Sonnet 5.5. Carbon led the management consulting and investment banking tracks, while Argon still led corporate law. Mercor's post was widely reshared, including by Google DeepMind leaders, which keeps raising APEX's profile as a benchmark for professional-services agent work.
Sources
Lovable spotlights a Perplexity connector so built apps can pull cited live-web data
LaunchXBluesky
Lovable promoted its Perplexity connector, which lets apps built on the platform pull source-backed information from the live web in real time, demonstrated with a market-research dashboard. A managed option lets basic Perplexity web search run without the builder having their own Perplexity API key, with usage charged to workspace credits, while deeper research features still need a key. Perplexity reposted the announcement, and it drew one of Lovable's higher engagement counts this week. It adds to a run of managed AI connectors, including Firecrawl and Parallel, that Lovable bills through its own credits.
Sources
Decagon launches a partner program and lets buyers use cloud commitments on four marketplaces
NewsNews
Decagon introduced the Decagon Partner Community, grouping model, cloud, data, payments and contact-center vendors with systems integrators and resellers. It highlighted links with Databricks, Plaid, Twilio and Five9, plus personal-agent partners Instinct and Muse. Decagon noted it is sold through AWS, Google Cloud, Databricks and OpenAI marketplaces, where eligible customers can draw down existing commitments. Partners get joint pipeline, co-marketing and certification through Decagon University.
Sources
Perplexity Computer can now work inside your own Chrome through a new extension
LaunchXNews
Perplexity shipped a browser extension that lets the Computer agent in its Mac and Windows apps operate inside the user's existing Chrome, Edge, Brave or Arc, using sites they are already signed into. The agent can open pages, fill forms and work across tabs, and Perplexity's demo turns a business's reviews into a prioritized list of fixes and replies. It is live for all Computer users in the desktop apps, while Comet already has the capability built in. The store listing says the browser is only used during a task the user started, but reactions quickly raised questions about guardrails when an agent acts inside real logged-in accounts.
Sources
Vercel Labs open-sources agent-desktop, accessibility-tree computer use for any agent
LaunchXBlueskyNews
Vercel Labs released agent-desktop 1.0, an Apache-2.0 Rust CLI and C library that lets coding agents operate native apps on macOS, Windows and Linux. Instead of screenshots and pixel clicks, it reads the OS accessibility tree and hands the agent JSON snapshots with stable element refs. Actions run headless by default, so agents can work in parallel without taking over the user's cursor. The builder reports large token savings on dense Electron apps, though those figures are self-reported, and the launch drew thousands of likes and broad developer reposting.
Sources
Baseten adds autoscaling schedules to pre-warm replicas ahead of predictable traffic
LaunchNews
Baseten introduced autoscaling schedules that raise or cap replica counts during set time windows, so models are warm before known traffic peaks arrive. Schedules can recur daily or hourly or run once, for example ahead of a launch or load test. Each environment supports up to ten non-overlapping windows sharing one timezone, configurable from the dashboard, CLI or Management API. The feature replaces the external cron jobs teams previously used to avoid cold-start latency.
Sources
A prominent NVIDIA CUDA leader announces he has joined Modal
HiringX
A well-known NVIDIA CUDA and C++ standards figure publicly announced he has joined Modal, and Modal reposted it. His tongue-in-cheek post drew roughly 250 likes and a stream of replies. The hire adds deep GPU programming expertise as Modal expands into kernels, training and inference optimization. It follows Modal's recent large funding round and its push into NVIDIA Vera CPUs for sandboxes.
Sources
Modal's open-source Dojo library adds post-training support for Reflection Beam
LaunchXOther
Modal promoted Modal Dojo, its open-source library for SFT and RL post-training on Modal, and added first-class support for post-training Reflection's Beam model. Dojo installs as a Python package, deploys a training dashboard, and ships recipes plus an agent skill bundle so coding agents can drive runs. Beam is a large open-weight mixture-of-experts model that Modal already serves as a dedicated endpoint. Together with Spindle and its Training API, it shows Modal building out a full post-training stack.
Sources
Suno asks Boston court to pause UMG and Sony's second training-data lawsuit
NewsNewsOther
Suno filed a motion to stay the second copyright case brought by Universal and Sony, which covers roughly 60,000 recordings allegedly used for training. The filing, backed by an 11-page memorandum and a declaration from Suno's legal side, argues the new case should wait for key rulings in the labels' original 2024 suit before the same judge. Suno had answered the second complaint on October 20 rather than moving to dismiss it, leaning on fair use. A pause would slow the labels' push to test the v6 models and retained training data. These are litigation positions, not court findings.
Sources
Nscale reportedly frozen out of the bond market as lenders toughen AI build-out terms
NewsNewsXBluesky
The Wall Street Journal reported that Nscale, alongside CoreWeave and Nebius, has effectively lost access to the high-yield bond market in recent weeks. Lenders and investors are said to be demanding stricter terms and higher yields before funding more AI data center construction. For Nscale this lands on top of reported construction setbacks and an IPO whose timing is said to be slipping toward December. The squeeze helps explain why Nvidia is reportedly weighing a guarantee on Nscale's West Virginia construction debt.
Sources
Suno launches an official MCP server for making songs inside AI assistants
LaunchNewsXOther
Suno released Suno MCP in beta on Oct 28, letting users connect Suno to Claude, ChatGPT, Claude Code and other MCP clients and generate music without leaving the chat. Per its release notes, users can draft lyrics and style conversationally, turn an uploaded voice memo or loop into a full track, and make covers, extensions and mashups, with songs landing in their Suno library and drawing on plan credits. Music Business Worldwide also spotted a sign-in portal for Suno's developer API, which the company says is still in beta, and noted Gemini and Cursor among supported apps. Creators on X and Japanese blog posts quickly shared songs made through the integration, with an enthusiastic tone.
Sources
Instinct's integration streak and a $30B raise rumor fuel personal-agent debate
DiscussionXHN
Posts on X tallied a burst of new Instinct hookups in a few days, including Stripe Link, Shop Pay, 1Password and OpenTable bookings, letting the agent pay, sign in and reserve on users' behalf. Clerk said it is adding private-beta support for the Personal Agent Protocol that Instinct helped launch with Meta, Stripe, Shopify and others. Separately, some accounts claimed Instinct is raising at about a $30B valuation, an unconfirmed rumor a month after its reported $1B round at $10B. Users kept ranking Instinct against Meta Muse, ChatGPT Dots and others, with praise for memory and complaints that such agents need too much personal context to work well.
Sources
Zocdoc gives Instinct direct MCP access to book doctor appointments
NewsNewsX
Zocdoc opened its scheduling infrastructure to AI agents and named Instinct as its first dedicated agent partner, giving the text-and-call assistant direct access to Zocdoc's MCP server. Per Zocdoc's release, Instinct can find vetted providers, verify insurance and book confirmed appointments straight onto providers' calendars, with patient consent required before booking or cancelling. Healthcare Dive notes Google Gemini and Amazon Health AI are also connected, and Zocdoc says agent traffic has grown roughly 30x since July. MobiHealthNews picked up the Instinct angle on Oct 30; on X the deal was read as part of a fast run of commerce integrations for the agent.
Sources
Harvey extends Legal Agent Bench to test end-to-end brief writing on US case law
LaunchXOther
Harvey's research team released LAB: United States Case Law, an extension of its Legal Agent Bench that has frontier models research and draft full litigation briefs from real public case records sourced via Free Law Project. Model-written briefs are compared head-to-head, and Harvey reports that a small group of frontier models can now produce complete first-draft briefs in one shot without attorney guidance, though a long tail of failure modes remains. Harvey says it will make this research more accessible through a partnership with Grubel. On X, a Harvey researcher and legal-tech accounts shared the results positively, treating it as a meaningful step for long-horizon legal agents.
Sources
Harvey buys Aida, its fifth acquisition of 2026, to build law-firm context graphs
NewsNewsX
Harvey acquired Aida, a two-year-old AI sales-execution startup co-founded by Glean co-founder Piyush Prahladka and former Patreon and Apple engineering leader Atish Das Sarma. Per the Global Legal Post and Artificial Lawyer, Aida's search and ML engineers join Harvey's product and engineering org, with the founders becoming a VP and head of context graph. Harvey says the goal is to connect knowledge scattered across firms' documents and systems so its agents understand how a firm actually works. It follows Hexus, Lume, Benchmark and Guardrails earlier this year, and on X observers framed it as an acqui-hire of search talent in the race with Legora, which has made six buys.
Sources
Baseten cuts delta weight sync for a 770GB model to 15 seconds with sub-second pause
LaunchXOther
Baseten engineers published a write-up on rebuilding their delta weight sync path so updates are applied against weights already on the GPU, removing a second full copy in CPU memory. A roughly 770 GB checkpoint that previously took about 36 seconds now syncs in around 15 seconds with a pause of under one second for live requests. The freed host memory goes to KV cache, which the company says lifts throughput by up to 3.2x and unblocked RL rollouts for a large model on two GB300 nodes. The post drew modest technical discussion on X about failure handling during syncs.
Sources
Baseten reportedly raising about $2B at valuations up to $30B
NewsNewsXBluesky
Bloomberg reported on Oct 30, citing people familiar with the plans, that Baseten is raising roughly $2 billion in two tranches, one pricing the inference company at $26 billion and the other at $30 billion. That would more than double the $13 billion valuation it reached in June after a $1.5 billion round. The company has not confirmed the deal, so terms remain reported rather than final. The news was relayed widely by Techmeme, financial outlets and posts on X, mostly framed as another sign of investor appetite for inference providers.
Sources