Introducing the Eco Plan: Serious AI, $9/mo, a Fraction of the Energy

A few months ago we shipped Eco Mode — a routing option that sends your request to highly capable, energy-efficient AI models instead of the most powerful (and most expensive) frontier ones. We built it as a feature. It turned out to be a philosophy.

The response told us something: a lot of people don't want the most powerful AI in the world for every message. They want strong, dependable intelligence that doesn't burn through their budget — or a data center — to draft an email. So today we're making efficiency a first-class citizen on ARMES.

Meet the Eco plan: $9/month, with a 3-day free trial. And at its center, a new flagship agent: ARMES GPT ECO.

What sparse models actually are

When a dense frontier model answers you, the full network fires — often hundreds of billions of parameters, whether you asked for a business plan or a birthday message. That's powerful, and for most everyday tasks, wasteful.

The models behind Eco use sparse Mixture-of-Experts (MoE) architectures. They carry enormous knowledge — hundreds of billions, even a trillion total parameters — but route each request through only the expert sub-networks relevant to the task. A 671-billion-parameter model might activate just 37 billion for your answer. Across the Eco pool, most models fire roughly 3–6% of their total parameters per response — and none more than a tenth.

That is not a compromise. It's engineering. Less compute per answer means less energy per answer and dramatically lower cost — without meaningfully sacrificing quality on everyday work.

And these are not budget models. The launch pool includes models that go toe-to-toe with Claude Opus-class systems on real coding benchmarks (SWE-Bench Verified), excel at PhD-level science questions (GPQA Diamond), and lead open-model leaderboards for frontend generation and grounded research. We introduce each of them below — and the live pool is always visible, with Eco badges, on our models page.

ARMES GPT ECO: the agent built around efficiency

Every ARMES plan has a flagship agent, and Eco now has its own. ARMES GPT ECO works exactly like you'd expect an ARMES agent to work: you talk, it thinks, and behind the scenes intelligent routing picks the best model for each request. The difference is the pool it draws from — every candidate is an efficient, sparse-architecture model.

Prefer to choose yourself? Eco includes manual model selection from the efficient pool, right in the model picker.

The Eco plan also includes:

• Eco-tier agents in the ARMES Agent Library, alongside the free agents
• MCP integration — connect ARMES to Cursor, Claude Code, OpenClaw, and more
• Unlimited notes, folders, and prompt templates
• Meaningfully higher rate limits and usage budgets than Free
• The same private inference as every other plan (more on that below)

At $9/month, it's the lowest-cost way to run serious AI daily. And because every response consumes a fraction of the compute, your included budget stretches roughly 10–30x further than frontier routing. If frontier models give you ~100 quality conversations in a cycle, the efficient pool can give you 1,000+.

Meet the launch pool

So what's actually behind the curtain? At launch, ARMES GPT ECO — and Eco Mode on every Pro and Ultra agent — draws from eight models. Each earned its spot by being genuinely excellent at something, not by being cheap. A few words on what we're leaning into with each:

• DeepSeek V4 Pro — The heavy thinker. A 1.6-trillion-parameter MoE that activates roughly 3% of itself per response, it delivers elite mathematics, science, and algorithmic coding — including an 80.6% on SWE-Bench Verified, right alongside the frontier's best coding models.
• DeepSeek V4 Flash — The efficiency champion. Just 13 billion active parameters out of 284 billion, making it the lightest, fastest, most energy-frugal model in the pool. It handles the everyday flow — drafts, edits, summaries, reformatting — at a fraction of a fraction of frontier cost.
• Google Gemini 3.5 Flash — The sprinter. The fastest capable model in the pool and its strongest tool user, which matters on ARMES, where any message might mean searching your notes, reading files, or hitting the web. Also a careful, well-calibrated analyst for financial and legal material.
• Meta Llama 4 Maverick — The human voice. Distilled from Meta's much larger Behemoth, it's the pool's standout for creative writing, marketing copy, and warm, natural conversation — the model you want drafting the message that has to sound like you.
• MiniMax M3 — The straight shooter. The lowest hallucination rate in the pool, and — rarer still — a model that says "I don't know" instead of guessing, which ARMES turns into a grounded web or context search. Also the pool's most precise instruction-follower for exact-format deliverables.
• Moonshot Kimi K2.6 — The researcher. A trillion-parameter MoE with the strongest grounded-search and multi-source synthesis performance in the pool. When you ask for a deep dive with citations, this is the kind of model you want holding the flashlight. Excellent multilingual nuance, too.
• NVIDIA Nemotron 3 Ultra — The dependable generalist. A hybrid Mamba-Transformer design whose efficiency actually improves on very long inputs, with the best 1-million-token retention in the pool. Fast, well-calibrated, strong with tools — the model that's never the wrong answer.
• Z.ai GLM 5.2 — The designer-scholar. State-of-the-art one-shot frontend and interface generation (it outscores several frontier models on SWE-bench Pro), paired with the deepest knowledge profile in the pool for philosophy, obscure domains, and strategic judgment.

In Eco Mode, routing weighs each request against these strengths automatically. Prefer to drive? Every one of these models is manually selectable from the picker. And this is a living roster: as stronger and more efficient models ship, the pool evolves with them. The current lineup — on Eco and every other plan — is always on our models page.

Already on Pro or Ultra? You've had this all along

Here's the part worth re-iterating, because it's easy to miss: every Pro and Ultra agent already has Eco Mode in its mode selector, right beside Auto.

Nothing about that changes today. If you're on Pro or Ultra, you don't need the Eco plan — you have the efficient pool and the frontier pool, and you can switch between them per chat:

• Auto Mode for the work that demands the frontier: complex strategic analysis, PhD-level reasoning, nuanced legal questions, huge documents.
• Eco Mode for everything else: drafts, brainstorms, code reviews, summaries, daily conversations.

One app. Two strategies. Zero waste. The Eco plan simply makes that second strategy available as a plan of its own, for people who want efficient AI as their whole subscription — not an add-on to a bigger one.

An honest word about trade-offs

We built our reputation on not overselling privacy, and we're not going to oversell efficiency either.

Eco is not a downgrade — it's a different optimization target. Frontier routing optimizes for absolute best intelligence. Eco optimizes for the best intelligence per dollar and per watt. For the vast majority of everyday AI tasks, you won't notice a meaningful quality difference. For your very hardest work — deep reasoning, complex analysis, very large documents — frontier models on Pro and Ultra remain the better choice, and we'll tell you so.

If that honesty costs us a few upgrades, fine. It's the same reason we publish our rate limits and our architecture.

Same privacy. Always.

Every model in the efficient pool runs through the same zero-data-retention infrastructure as everything else on ARMES. Your conversations are processed and immediately forgotten — never stored, never profiled, never monetized. The privacy guarantee is identical regardless of plan, mode, or model.

We think of it as one principle applied twice: sovereignty over your data, and sovereignty over your resources. You shouldn't have to hand over your privacy to use frontier AI, and you shouldn't have to burn frontier-scale compute to answer everyday questions.

The bigger picture

The Mixture-of-Experts revolution proved that activating 37 billion parameters out of 671 billion is enough to deliver exceptional results for most tasks. We believe AI platforms have a responsibility to offer that choice — and to reward you for taking it.

Use the full power of the frontier when it matters. Run efficient the rest of the time. Either way, stay private.

Try the Eco plan with a 3-day free trial — cancel anytime before day 3 and you're never charged. Or, if you're on Pro or Ultra, open any chat and flip the mode selector to Eco. It's been waiting for you.

Start Eco → · How Eco works → · The live model pool →

Joseph Founder, ARMES