Transcription
Two days after GPT-5.6 launched, the benchmark arguments are already starting to settle down. Most people now know roughly where it is strong, where Fable 5 is stronger, and which model they would rather use for coding, research, or reasoning.
But that creates a more interesting problem. If GPT-5.6 is better at one part of a task, Fable is better at another, Opus is better at debugging, and an open model can handle the cheap routine work, why are we still being forced to choose one model before we even start?
That is where Abacus is taking this [music] next. On top of its existing supercomputer, it is now combining Fable 5, GPT-5.6, [music] custom routers, and more than 100 other models into agents [music] that can choose the right model for each part of a task, then execute it on an always-on machine with root access.
So, the cleanest example is the new smart coding router. You give it one request, but instead of sending the entire thing to a fixed model, the router first classifies what kind of work is required. Architectural planning and difficult multi-step engineering can go to Fable 5. Routine writing, editing, refactoring, and test generation can go to Claude Sonnet 4.6. Questions about an existing codebase can be handled by GPT-5.6 Soul, and the whole setup can also be flipped so that Fable acts as the master agent while GPT-5.6 and Rock 4.5 work underneath it as sub-agents.
That sounds like a small change, but it affects the entire workflow. A normal coding assistant gives you one answer from one model. This router can break the task apart, assign different pieces to different models, and connect the result to GitHub. It can inspect a repository, audit it for security vulnerabilities, explain the architecture, and generate a scaling plan covering the database, caching, media handling, asynchronous jobs, and resilience.
The router is also exposed through an OpenAI compatible API with streaming, function calling, and multimodal support, which means it can sit behind an actual product rather than only being used inside [music] the Abacus chat interface.
Now, Abacus calls this build your own superintelligence. I would not take that to mean you are training a brand new superintelligent model. What you are building is an orchestration layer that can behave more capably than one model used alone because it knows when to call GPT-5.6, when to call Fable, when to use Opus, and when a much cheaper open model is already enough.
And this is where the supercomputer becomes more relevant than it was at launch. The underlying machine still gives you an always-on cloud environment that can run 24/7 without cold starts, host databases and APIs, use [music] S3 compatible storage, scale from lightweight scripts to GPU-heavy workloads, connect to AWS and GitHub, accept SSH access, >> [music] >> and publish apps or services to a public URL. It can host open-source LLMs, keep agents like Hermes or Openclaw running, operate game servers, and support autonomous systems such as trading agents.
What has changed is how directly the agent can now use all of that infrastructure. Inside the Abacus agent, Max mode sends the most demanding work to Fable 5. Fable gets root access to the supercomputer, so it can install dependencies, write scripts, run them, inspect the output, fix errors, configure services, deploy the application, >> [music] >> and verify that the public version works. The user is not copying code into a terminal or returning with screenshots of error messages. The model is operating inside the environment where the software will actually run.
One demonstration pushes this pretty far. The user [music] asks for a fully explorable 3D Hogwarts-style castle inside a browser. It has to be one self-contained HTML file with first-person movement, [music] six walkable areas, procedural geometry, collision detection, flickering torches, fog, a day-night cycle, and on-screen labels.
Before writing anything, Fable asks three design questions: whether the castle should feel faithful, gothic, gloomy, or more fantastical. Whether the first scene should begin at night or dusk. And which three areas deserve the most detail. Then it writes more than 1,200 lines of code, builds the environment, [music] tests player movement, opens the app in a browser, takes screenshots, checks whether the named areas can actually be reached, and performs another visual pass without being asked.
The final version grows from six areas to eight with dark gray gothic stone, [music] amber torchlight in the halls, sickly green light in the dungeons, a layered starfield, moon glow, wet dungeon floors, WASD controls, mouse look, shift to run, and space to jump. The important part is not that an AI can generate a castle demo. We have seen plenty of generated demos. The difference is that this one can debug the environment, verify the result, and deploy it publicly from the same conversation.
The next example is an AI trading strategy lab, and this is where the [music] system looks more like software development than prompt generation. The request describes a visual platform where several agents ingest market data, read research papers and news, generate strategies, backtest [music] them, apply risk checks, vote on the results, deploy approved strategies, and monitor them live.
Fable asks about broker integrations, market data providers, research sources, [music] model choices, and strategy types. It then creates a PostgreSQL database, a fast API back-end, the agent logic, and a polished next.js front-end. The finished dashboard has demo, [music] paper, and live trading modes, real-time equity tracking, a kill switch, and a pipeline for BTC USDT, ETH [music] USDT, and SOL USDT.
Nine agents run through market data analysis, research extraction, technical indicators, news regime detection, strategy generation, >> [music] >> backtesting and robustness, ranking, execution, and monitoring. It also generates candlestick charts with Bollinger Bands and RSI, strategy cards showing expected return, maximum drawdown, win rate, [music] profit factor, and Sharpe ratio, plus approval controls, live positions, risk settings, an audit trail, and a research map connecting source material to extracted ideas and generated strategies.
Obviously, none of this means somebody should allow an AI-generated system to trade real money without independent testing. But, it does show the difference between asking a model for a trading script and asking an agent to assemble the database, back-end, interface, controls, and monitoring around that script.
Another demo is much more practical. A user asks the agent to self-host Qwen 2.5, the 0.5 billion parameter version, and build a ChatGPT style interface for it. Fable asks about model precision and recommends 4-bit quantization to reduce resource use. It also asks whether the interface should support conversation history, adjustable parameters, and multiple chat sessions.
It then produces a seven-step execution plan, verifies llama.cpp, downloads the gguf model, launches the inference server under systemd, builds a node.js proxy and front-end, configures nginx for server-sent event streaming, and tests both the local and public endpoints. The final Quen Chat interface has multiple conversations, saved history, suggested prompts, model controls, and streaming [music] responses. Because the model server and front end are running as systemd services, they automatically restart and remain available after the build session ends.
The same environment is used to build Cloud TV Studio, an always-on internet station. The user asks for an admin interface, media uploads, playlists, a public viewer page, database support, cloud storage, a heartbeat, and protection against crashes caused by missing or empty files. Fable asks whether to use PostgreSQL or SQLite, checks the upload limits, and confirms the station name.
It then creates start, stop, and restart [music] controls, server and database indicators, a media library, playlist management, continuous looping, broadcast logs, S3 monitoring, and live viewer counts. The public page shows what is playing and what comes next, while Engine X and systemd keep the station running independently.
Then there is Ring Feed, a small social platform with a live feed and one-to-one video calls. Fable builds a fast API back end, SQLite storage, WebSockets for live events, >> [music] >> and WebRTC for audio and video. Users select a username, post updates, see who is online, call another user, accept or decline incoming calls, mute the microphone, disable the camera, and hang up.
The demo runs the flow with two accounts, [music] Mufasa_001 and Lion King_002, then opens the database to show users, session tokens, messages, [music] call logs, and detailed events for invitations, acceptances, mute changes, [music] camera changes, and hang-ups.
The custom routing part is much faster to set up than the application demo suggest. [music] In the mixture of agent section, a user clicks new custom router and chooses [music] to describe the router in plain English. The instruction is one sentence. [music] Send hard coding to Fable 5, debugging to Claude Opus 4.8, and simple coding to GLM 5.2.
The platform automatically generates the categories, descriptions, routing rules, fallback behavior, and system prompt. System design, algorithm implementation, and performance optimization go to Fable. Bug diagnosis and repair go to Opus. Small scripts, basic functions, and [music] syntax questions go to GLM. Anything that does not match falls back to Fable. >> [music] >> The user names the router my_coding, saves it, and asks for a machine learning forecasting algorithm. The interface shows that the request was routed to Fable 5, which produces a gradient boosting pipeline with lag [music] features and rolling means. A basic syntax question would be sent to GLM 5.2 instead.
That cost difference is a major part of the product. Fable Max mode is powerful, but there is no reason to pay for Fable on every message rewrite or simple code question. Abacus has a cost optimized general router that can send simple Q&A to Gemini 3.1 flashlight, quick summaries to GPT 4.1 nano, creative writing to Claude Haiku 4.5, fast coding to Kimmy K2 Turbo, and lightweight reasoning to DeepSeek V4 Flash. The examples include professional message rewrites, Slack ready product summaries, homepage calls to action, meeting note extraction with owners and deadlines, technical explanations, and simple HTML pages.
For more demanding work, the performant router [music] uses GPT 5.6 for research and image generation, Opus 4.8 [music] for complex analysis, Fable 5 for deeper creative work and difficult building tasks, GPT-5 for general conversation, Qwen 3 Coder [music] for advanced coding, and E3 Pro for math and science. That configuration is used for financial forecasting with best, base, and worst-case scenarios, cinematic writing, product concepts, landing page design, code generation, and strategic analysis.
There is also an open-source coding router. DeepSeek V4 Flash handles quick commands and straightforward requests. Kimmy K 2.7 Code handles code generation and database schemas. GLM 5.2 takes on complex debugging, root cause analysis, and agentic coding, while Qwen 3 Coder can handle more advanced implementation work. The examples include installing Tailwind, React, and Vite, comparing FastAPI, Express, and Hana. Generating a PostgreSQL schema [music] for a task management app, and tracing a Next.js hydration error down to specific code-level fixes.
Open-source models do not make the compute free, but they remove per-token API pricing and give the user more control over deployment, privacy, and scaling. A company could use open models for high-volume internal work, [music] reserve Fable and GPT-5.6 for mission-critical requests, send premium users [music] to stronger models, and route free tier users to cheaper ones. Customer support, research, coding, and internal analysis could each follow different rules.
So, yeah, that is what Abacus is building on top of its supercomputer. If you enjoyed the video, make sure to like it and subscribe for more AI news and updates. Thanks for watching, and I'll catch you in the next one. >> Yeah.