Sakana Fugu vs AskCodi CTO: Same Problem, Opposite Designs
Sakana Fugu shipped GA Jun 22 as a single multi-model API. AskCodi's CTO runs locally as an agent team. Same problem, opposite designs. Here's which fits when.
Sakana Fugu hit GA two days ago. A 7B "Conductor" model trained to route across a pool of frontier LLMs, served as a single OpenAI-compatible API endpoint. Drop-in. No SDK changes. The Japan-based Sakana AI team built it on the back of two ICLR 2026 papers (TRINITY and Conductor), and it's the cleanest implementation of "multi-model orchestration as a service" anyone has shipped.
I've been building the same thing at AskCodi for the opposite reason. Local-first, on your machine, with a coordinator agent that hires specialist agents to do real work across isolated git worktrees.
Same problem. Opposite designs.
This post is the honest comparison, including where Sakana wins and where it doesn't fit.
What Sakana Fugu actually is
The 30-second version:
- One API endpoint, OpenAI-compatible. If your code works with GPT, it works with Sakana Fugu.
- A 7B "Conductor" model picks which underlying LLM to call for each subtask. Trained via reinforcement learning, not hand-coded.
- Sub-agents get assigned Thinker, Worker, or Verifier roles dynamically. They coordinate in natural language under the hood.
- Two variants: balanced (everyday coding, chat, review) and Ultra (Kaggle-grade, security audits, multi-hour research).
- Released by Sakana AI on June 22, 2026, after a two-month closed beta.
The model itself is small (7B parameters), but its job is to be smart about which bigger model to ask. That's the whole bet.
What AskCodi CTO actually is
The 30-second version:
- A local desktop app. The CTO orchestrator runs on your machine, not in the cloud.
- It breaks a feature request into subtasks and assigns each to a specialist agent: backend, frontend, testing, security.
- Each specialist runs in its own isolated git worktree. Parallel work, no collisions, auto-cleanup.
- Full filesystem and shell access. MCP tool connections (GitHub, databases, your own APIs).
- Sign in with your existing Claude Pro/Max or ChatGPT Plus/Pro plan. No extra API bill.
The orchestrator is explicit. The model choice per subtask is visible. Code never leaves your machine.
The tradeoff matrix
This is the meat. Same problem, completely different answers.
| Dimension | Sakana Fugu | AskCodi CTO |
|---|---|---|
| Where it lives | Cloud API | Local desktop |
| Orchestrator | A 7B model, RL-trained | A coordinator agent with explicit logic |
| Routing visibility | Opaque (model output) | Audit trail of per-task choices |
| Code execution | API in, text out | Full filesystem, shell, MCP |
| Auth | Per-token Sakana billing | Bring your own Claude/ChatGPT plan |
| Setup | Change one base URL | Install + configure desktop app |
| Latency for one request | Fast (one round-trip) | Slower (multi-step agent loop) |
| Best for | Hosted apps, prototyping, classification | Shipping features in your own repo |
Both are correct designs. They optimize for different things.
Where Sakana wins
Be honest about this. Three things Sakana does that AskCodi does not.
Zero setup. Change one base URL in your existing OpenAI client and you're routing across a pool of models with a trained conductor making the calls. No new SDK, no new auth flow, no agent infrastructure. If you're building a hosted app, this is hard to beat.
Frontier-grade routing. The Conductor model is purpose-trained on the routing decision itself. It saw more examples of "this kind of request gets best results from this kind of model" than any handwritten router I've seen. Likely sharper than my heuristics.
Single-request latency. One round-trip to one API. AskCodi's loop is multi-step by design. For a one-shot task like "classify this support email," Sakana is the right shape.
Where AskCodi wins
The other side.
Audit trail. You can see which model the coordinator picked for which subtask, why, and what it returned. Sakana's Conductor decision is a model output, not a logged choice. Day 1 demos make this look like a non-issue. Day 2, when an agent does something weird, it's everything.
Real execution. Sakana returns text. AskCodi specialists open files, run your test suite, edit code, hit MCP tools, commit to git worktrees. If the work involves touching your repo, Sakana isn't the tool.
Bring your own subscription. No new bill. Use the Claude Pro/Max or ChatGPT plan you already pay for. If you're a heavy user, that's a real annual difference.
Local-first. Your code never leaves your machine. Sakana is the opposite by definition.
Parallel work across worktrees. Multiple specialists working on different parts of your repo at the same time, no collisions. The worktree isolation is the trick that makes it safe.
When to use which
If you're picking, here's the framework I'd use.
Use Sakana Fugu when:
- You're building a hosted application that consumes AI as a service.
- You want one API, one auth, one bill.
- The task doesn't need code execution.
- The request is short and self-contained.
Use AskCodi CTO when:
- You're shipping features in your own codebase.
- You need to see which agent did what (PR review, debugging, accountability).
- You want to use your existing AI subscription.
- Your code shouldn't leave your machine.
Use both when:
- Sakana is your inference layer for hosted, latency-sensitive operations.
- AskCodi is your shipping team for the repo you actually maintain.
These aren't mutually exclusive. They solve different problems in the same broad space.
The bigger pattern
The orchestrator wars are getting interesting. Three camps are emerging:
- Orchestration as a model (Sakana Fugu). The routing logic IS a trained model.
- Orchestration as a library (LangGraph, CrewAI, AutoGen). The routing logic is Python you write.
- Orchestration as a local coding agent (AskCodi CTO, Claude Code, Cursor). The routing logic lives in a coordinator that runs on your machine and touches your files.
Underneath all three, MCP is becoming the protocol that lets any of them call any tool without M×N integrations. That's a separate fight, and MCP appears to have won it.
Sakana's launch matters because it makes camp 1 a credible product, not just a research direction. Expect 2-3 more conductor-style models from Anthropic, OpenAI, or Google in the next 6-12 months. They have the data and the compute. The question is whether they bother.
For coding work today, I run a conductor-style API when I want hosted classification or summarization. I run AskCodi's CTO when I'm shipping features. Different tools for different jobs.
Close
The interesting thing about Sakana's launch isn't that they shipped. It's that they bet on the OPPOSITE architecture from where most of the dev-tool ecosystem is heading. Cloud-API vs local-execution. Black-box conductor vs visible specialists. Different priorities, both internally consistent.
Their bet might pay off. Ours might. Probably both win in different slices of the market.
If you want to see what local specialist agents look like in practice, Inside AskCodi's AI CTO walks through the architecture. If you want the math behind why multi-model routing saves real money, Hit Claude's Weekly Limit? breaks it down on one sprint's worth of internal numbers.
Sakana Fugu is impressive engineering. So is the entire local-agent ecosystem. The next year is going to be fun to watch.
