Changelog

Moby now starts every new OpenAI conversation on GPT-5.6 Sol, replacing GPT-5.5 as the default. There is nothing to turn on. GPT-5.6 Terra and GPT-5.6 Luna are also available in the model picker when you want a different balance of depth, speed, and cost. Default reasoning effort remains Medium.

The upgrade is most visible when the work gets complicated: longer plans, professional analysis, coding, research, and multi-step execution. There is not one universal “smarter” number, but OpenAI's published comparisons show GPT-5.6 Sol ahead of GPT-5.5 across nearly every listed benchmark.

Every GPT-5.6 model in Moby now uses programmatic tool calling by default. Instead of calling one tool, reading the entire response, and repeating that loop, Moby can write a lightweight in-memory program that coordinates a bounded set of tools, processes the intermediate results, and returns only the evidence the job needs.

That means fewer model round trips and less unnecessary context for tool-heavy work. Moby can fan out research, group and rank rows, join results, validate exact matches, and stop as soon as the answer is complete. The result is faster execution without flattening the reasoning or skipping the checks.

Grok 4.5 has replaced Grok 4.3 in Moby's model picker. It is xAI's newest frontier model for coding, agentic tasks, and knowledge work, with configurable reasoning effort for jobs that need more depth.

Moby keeps the model choice in your hands. Pick Grok 4.5 when you want the newest xAI model, or stay with GPT-5.6 Sol as the new default.

Video generation in Moby is no longer one generic tool. We broke every model out into its own tooling, so each one exposes the controls it can actually honor instead of flattening four different systems into the same prompt box.

Grok Imagine Video 1.5 is the fast still-to-motion route: one image, 1 to 15 seconds, source-ratio preservation, and 480p through 1080p. Veo 3.1 handles cinematic text generation, first-and-last-frame transitions, up to three visual references, a separate negative prompt, and optional generated audio.

Gemini Omni Flash can generate from text or references, then revise the result through a persistent interaction ID. You can change the environment or motion while protecting the product, camera, timing, and composition that already work. Seedance 2.0 takes the richest packet: up to nine images, three reference videos, three audio references, flexible ratios, exact 4 to 15 second timing or auto timing, and output up to 4K.

The workflow changed with the tools. Moby now proves one anchor clip, watches every result, classifies it as final, usable proof, targeted retry, or wrong route, and only stitches approved, compatible clips. Native generation still is not the right place for exact text, UI, captions, crossfades, or audio mixing.

Moby can now split big jobs into smaller ones and run them in parallel with sub-agents. Type @ in the composer and pick a mode. There are three, and each one solves a different problem.

Model delegation. Moby splits your task and runs it across four model providers: Google, Anthropic, OpenAI, and xAI. Each provider takes its own pass, then Moby synthesizes the best of all four into one answer. You get the strongest output from every lab instead of betting on one.

Structured sub-agents. Built for structured work at scale. Go over all my landing pages. Score every creative. Audit each campaign. Moby launches up to 50 sub-agents at a time, and when the job is bigger than that, it batches them until the work is done.

Full sub-agents. Full Moby capability inside a sub-agent. Each one can do anything Moby can do. Build a persona automation that reviews your ads and scores them against each persona, then run it on a schedule.

Most teams treat measurement as a religion: pick MTA, or MMM, or lift tests, and defend the choice. But each method is just a different way of being wrong. MTA updates constantly and only sees correlation. MMM sees the whole budget and moves weekly. A lift test is causal ground truth you can’t run everywhere.

Compass, Triple Whale’s measurement suite, runs all three: MMM, incrementality testing, and MTA. As of today it lives inside Moby, which means the thing reading the measurements can also act on them. Ask where the next dollar works hardest and it answers in marginal ROAS. Ask which channels are saturating and it shows you the curve. And when MTA, MMM, and your experiments disagree, it says so and tells you what to validate next.

The interesting part is what happens when you stop visiting the dashboard. Put the reads on a loop: a budget brief every week off the fresh MMM refresh, a monitor that catches channels as they saturate, a triangulation report that flags disagreements. Pair them with Actions and the reallocations arrive queued for your approval. Explore Compass.

Good work used to disappear into old chats. You'd ask Moby for a report, get exactly what you wanted, and three weeks later scroll through history trying to remember which conversation it lived in. The Library ends that: everything Moby makes (pages, reports, apps, media, files) sits in one place, organized across workspaces. Pin what you use, filter the rest, and stop excavating.

Conversations have a home now. The history page lists every chat with Moby, so you can find the thread from last Tuesday and pick it up mid-sentence.

The difference between a page and a tool is memory. A page shows you numbers; a tool remembers what you did with them. Apps Moby builds now hold state, so what you save in one session is waiting in the next. Create a workspace, dashboard, or app straight from the home screen, or start from pre-built apps and drafts.

The hardest part of asking an AI anything is telling it where to look. The composer now has a grammar for exactly that: type @ to point Moby at a connector or tool, / to invoke a skill, and # to aim the question at an exact data table.

When you already know where the answer lives, you shouldn’t have to hope Moby guesses right. You point. @ settles the “which source?” question before it comes up, / turns a saved workflow into two keystrokes, and # locks the query to the table you meant. The first answer comes from the right place, and you spend your words on the work instead of the directions.

The most corrosive bug in analytics is two surfaces that disagree about the same number. So we rebuilt the Summary page (the operating picture of your business) and the Pixel underneath it as one project, and it is live for every customer. The payoff: the same number, everywhere, in a Summary tile, on a dashboard, and out of Moby's mouth.

Your data shouldn't be loyal to one interface. The data platform for ecommerce is now open: the same context Moby runs on (metrics, attribution, definitions) works in Claude and any MCP client. Ask from wherever you like to work; the numbers come from the same place.

Most Slack bots wait to be summoned. Moby doesn't. Add it to a channel and it's in the room: following what's happening, keeping its own memory of that channel, and speaking up on its own when it has something worth saying, the way a colleague sitting next to you would.

The distance between “you should lower that budget” and the budget actually being lower is where most tools stop. Moby now crosses it. Around 500 actions are live across your connected platforms: budgets, bids, statuses, campaigns, audiences, discounts, emails, even ad comments. Each one is deterministic: an explicit, bounded change mapped to the exact campaign or entity before anything runs. You can read what is about to happen, and to what.

Who pulls the trigger is up to you. “Ask every time” queues each action with its reasoning, and you approve it, reject it, or argue with it. Write rules and Moby takes the obvious calls alone (“pause anything below 1× ROAS”) while checking in on the rest. The Actions Log keeps the record (what ran, when, why, in what mode) in the session, in automations, in the workspace, and across Moby. Put actions inside an automation and a morning routine like “cut the losers, scale the winners across Meta and Google” runs whether you slept in or not. Actions are the building blocks; automations make them repeatable. Read the media-buying deep dive.

Knowing which model is good at what is a full-time job, so Moby took the job. Four image engines and four video engines sit behind one prompt surface: you describe the shot, Moby picks the model. Instruction-heavy layouts with live text route to GPT Image 2. Photoreal product work goes to Nano Banana Pro, fast iteration to Nano Banana 2, stylized frames to Grok. Motion runs on Seedance 2.0, with Veo 3.1 for text-to-video and Gemini Omni Flash for video-to-video edit passes.

Video reaches 4K with multi-clip scripts: one brief becomes a sequence of shots, with negative prompts and reference routing so the subject survives every cut. Images reach 4K with up to sixteen reference inputs, which is what makes real editing possible. Composition, product, and identity hold across passes instead of morphing.

Nothing ships blind. Moby looks at every render before you do, checks it against the brief and the Brand Vault, and reports a verdict with the caveats attached. Everything below was made this way, start to finish, inside Moby.

An assistant that can only read your other tools is a spectator. This wave of connectors gives Moby hands. It reads the sheet, drafts the email, files the ticket, updates the board, ships the deploy.

That changes the size of what one instruction can carry. A single thread now moves work across surfaces: pull the campaign data, build the model in Sheets, brief the team in Slack, log the decision in Notion, open the follow-up in Linear. You ask once, and nothing gets pasted between systems by hand.

The governance stays boring on purpose. Every external write can be held for your review before it runs, and Moby reports a change as done only when the tool itself confirms it.

Generic output is what happens when the model doesn't know who you are. The Brand Vault fixes the knowing. Moby extracts your brand DNA from your live site and public sources and stores it as a structured vault: logos, color tokens, typography, voice, visual motifs, personas, and canonical asset links.

Every creative surface reads from it, so images, video, landing pages, and reports start from your system: your palette, your type, your voice. Point Moby at a brand once and everything it makes afterward answers to the same source of truth.

Memory you can't inspect is a rumor. Moby's memory is now a set of pages you can open. It runs at four levels: you, your shop, Slack, and each workspace, and every level is readable and editable in Customize Moby. Your preferences stay personal, team knowledge is shared by everyone, Slack keeps its own thread of context, and each workspace remembers its own project. What carries into a conversation is exactly what's on those pages.

If a wrong fact got stuck, delete the line. If there's something Moby should always know, write it in yourself. And if you're arriving from another assistant, import your memory from other AI providers: Moby starts already knowing what it should.

A skill is a playbook: a step-by-step workflow Moby loads and follows exactly. Two libraries now live in Customize Moby. The built-in Triple Whale library is ours to maintain, so it keeps improving underneath you. Your business's own skills are yours to write and edit, line by line.

You never pick a skill from a menu. When a request matches a workflow, the right one loads on its own. And Moby writes skills itself: it watches how you actually work and drafts playbooks from the repetition. Run a workflow a few times and it hardens into something Moby executes the same way every time, for you or for anyone on the team.

An assistant that knows ecommerce is easy to find. One that knows your business is not. Give Moby the documents that define it (brand guidelines, strategy memos, pricing, SOPs, positioning) and it stops guessing.

Context is durable. Upload once, and every chat, report, artifact, and automation is grounded in how your company actually thinks. Nobody re-attaches files. Nobody re-explains the basics at the top of a thread.

The first Moby 2 rolling launch, and the foundation for everything after it. Anything you run with Moby once can now run on a schedule: media-buying automations, recurring reports, monitoring loops. You define them in plain language, they execute on time, and every save is verified.

The word “schedule” undersells it. Automations are stateful loops: every run remembers the last one (baselines, ledgers, open questions, decisions) and picks up exactly where it left off. A run can stay on the main thread or launch sub-threads of its own, a swarm of agents working one job, fanning out in parallel and reporting back. A monitor that remembers what it flagged yesterday, escalates what changed, and hands the deep dives to sub-agents is doing something no cron job can.

True Classic already runs 100% of its Meta spend through Triple Whale. Read the launch post.

Moby dreams. After every session it replays the conversation and decides what deserved remembering: the preferences you stated, the corrections you made, the way your business actually works. Then it writes that into durable context. You never have to say "remember this," though that works too. It gets smarter while you're gone.

Memory lives at two levels, you and your shop, so the learning compounds in both directions. What you teach Moby follows you into every conversation. What your team teaches it raises the baseline for everyone.

Tell it once; it stays told.