Smarter Routing, DeepSeek V4.1 Flash, and Expanded Free Access: The September 2026 Platform Upgrade

Every time you submit a prompt to ARMES, an automated routing engine analyzes your query and pairs it with the single best model in your plan. Rather than forcing you to decide whether your task requires a full-stack code architect, a creative narrative writer, a deep web researcher, or a mathematical proof engine, ARMES handles model selection autonomously.

Over the past month, the global AI frontier experienced an unprecedented wave of architectural breakthroughs:

• OpenAI launched GPT-6 Astra, introducing recurrent depth reasoning and near-saturation of formal mathematical benchmarks.
• Google released Gemini 3.8 Flash, delivering 305 tokens per second with sub-second time-to-first-token and groundbreaking automated security patch remediation.
• DeepSeek unveiled DeepSeek V4.1 Flash (September 10, 2026), debuting their first Causal Encoder-Decoder (CED) architecture with asymmetric compute (~8B prefill / ~16B decode active) and FP4 compressed KV caching, establishing SOTA results on agentic coding and formal math before retiring legacy V4-Pro and V4-Flash endpoints.

Rather than letting older models linger or allowing routers to become cluttered, we conducted an end-to-end audit of all four ARMES tiers: Ultra, Pro, Eco, and Free. We retired models that no longer earned their place, upgraded mathematical and execution engines, settled the Eco tier on an uncompromising low-energy architecture, and expanded the Free tier's daily message allowance by 2.5x.

Here is the complete architectural breakdown of what changed, why we changed it, and what it means for your daily workflows.

---

1. Ultra Tier: The Uncompromising 4-Lab Frontier Quad

Canonical Pool: x-ai/grok-4.6, openai/gpt-6-astra, anthropic/claude-opus-4.8, google/gemini-3.8-flash

The Ultra tier ($200/month) is built for professionals, researchers, and engineering teams who require the absolute peak of artificial intelligence on every turn. In reviewing previous iterations, we observed that routing routine utility tasks to smaller models diluted the Ultra experience. When you subscribe to an elite tier, you expect world-class intelligence on every single turn.

We consolidated Ultra down to a focused 4-Lab Frontier Quad where every model is an undisputed category leader:

• xAI Grok 4.6 (Default Intelligence Anchor): Scoring 61 on the Artificial Analysis Intelligence Index and holding the top position on GDPval-AA v2 (1753 Elo), Grok 4.6 serves as the primary default. It anchors interactive full-stack implementation, complex architecture design, commercial strategy, and broad professional reasoning.
• OpenAI GPT-6 Astra (Apex Autonomous Reasoning & Formal Mathematics): OpenAI's latest generational flagship replaces GPT-5.6 Sol entirely on Ultra. Astra delivers an unprecedented leap in capability: 97.6% on FrontierMath Tier 4 (saturating frontier mathematical proofs), 88.0% single-attempt binary reverse engineering on SRE-Bench (vs Sol's 55.9%), 95.9% on BenchCAD geometric code generation, and 57.9% on Terminal-Bench 4.0. Powered by recurrent depth latent reasoning and Codex harness state persistence, Astra enables deep multi-repo distributed systems refactoring while guaranteeing 0.00% sandbox evasion on ExploitGym.
• Anthropic Claude Opus 4.8 (Writing, Nuance & Human Judgment): The gold standard for nuanced prose, creative narratives, executive communications, emotionally sensitive counseling, careful code review, and high-stakes interpersonal strategy.
• Google Gemini 3.8 Flash (High-Speed Execution, Scoped Code Repair & Science): Released September 2, Gemini 3.8 Flash delivers 305 tokens per second with a 1-million-token window and sub-second time-to-first-token. It handles rapid terminal workflows (Terminal-Bench 90.8%), scoped bug fixes (DeepSWE v1.1 73.8%), automated vulnerability remediation (CWE-Bench 47.2%), quantitative finance (Vals Finance 61.4%), and biomedical research (BioMysteryBench 56.5%). All high-volume data extractions, shell automation, and transformations run through Gemini 3.8 Flash, completely eliminating the need for lesser models on Ultra.

---

2. Pro Tier: The Curated 6-Powerhouse Pool & DeepSeek V4.1 Flash

Canonical Pool: x-ai/grok-4.6, google/gemini-3.8-flash, anthropic/claude-sonnet-4.6, qwen/qwen3.8-2.4t-a95b, deepseek/deepseek-v4.1-flash, z-ai/glm-5.3-flash

Pro is ARMES's most popular tier. To ensure Pro users experience exceptional execution speed without sacrificing analytical depth, Pro v15 introduces a major upgrade to its mathematical and algorithmic core:

DeepSeek V4.1 Flash Replaces Legacy V4-Pro

Following DeepSeek's upstream architectural transition to the Causal Encoder-Decoder (CED) architecture and subsequent phaseout of V4-Pro, we integrated DeepSeek V4.1 Flash (deepseek/deepseek-v4.1-flash, released September 10, 2026) as Pro's dedicated algorithmic anchor.

V4.1 Flash represents a generational leap in architectural efficiency and reasoning:

• Asymmetric Compute Efficiency: Built on a 552B backbone (+ ~196B Engram), V4.1 Flash activates only ~8B parameters during prompt ingestion (prefill) and ~16B during token generation (decode). Paired with FP4 compressed KV caching, it delivers sub-second time-to-first-token and generates at a blistering ~200–420 tokens per second.
• Decisive Benchmark Superiority: At maximum reasoning effort, V4.1 Flash decisively outperforms the retired V4-Pro on agentic coding (DeepSWE v1.1 74.2% vs 62.7%), terminal execution (Terminal-Bench 2.1 90.6% vs 87.9%), and formal mathematics (MATH-500 97.3% pass@1).
• The Mathematical Anchor: While Gemini 3.8 Flash excels at rapid terminal scripting, it exhibits known limitations on competitive math (MathArena 35.0%). Routing competitive programming, calculus derivations, dynamic programming algorithms, and SQL query-plan optimizations to DeepSeek V4.1 Flash gives Pro subscribers near-flagship mathematical proofs and deep algorithmic rigor.

The Rest of the Pro 6-Powerhouse Pool

• Full-Stack Architecture & Default Reasoning: x-ai/grok-4.6 (AA Index 61, CursorBench 69.9% #1, GDPval 1753) anchors end-to-end full-stack features, complex backend services, database schemas, and commercial business strategy.
• Rapid Terminal Automation & Scoped Code Repair: google/gemini-3.8-flash (~305 t/s, sub-second TTFT) takes over command-line execution, shell scripts, CI/CD pipeline configs, test scaffolding, and rapid bug fixes (DeepSWE v1.1 73.8%). Routing CLI and repair tasks to Gemini 3.8 Flash eliminates terminal bottlenecks on everyday developer workflows.
• Creative Prose, Empathy & Legal Care: anthropic/claude-sonnet-4.6 provides best-in-class human warmth, brand voice, persuasive messaging, sensitive counseling, and careful contract review.
• Academic Research & Dense Specifications: qwen/qwen3.8-2.4t-a95b leads the world on academic literature synthesis (PaperBench 93.0 #1) and multi-constraint document authoring (IFBench 82.8), handling formal whitepapers, RFC-style technical specs, and multi-paper reviews across a full 1M context.
• Visual Frontend & High-Speed Extraction: z-ai/glm-5.3-flash (Code Arena Frontend #2 worldwide) generates modern React, Tailwind, and dashboard interfaces while handling rapid tabular transformations with sub-second responsiveness.

---

3. Eco Tier: The Settled 5-Model Pure MoE Pool

Canonical Pool: deepseek/deepseek-v4.1-flash, google/gemini-3.8-flash, minimax/minimax-m3, nvidia/nemotron-3-ultra-550b-a55b, z-ai/glm-5.3-flash

The Eco tier is engineered for energy-conscious subscribers who demand capable, frontier-adjacent intelligence while minimizing compute and datacenter energy consumption.

In Eco v14, we implemented the most significant architectural streamlining in the tier's history, establishing the Settled 5-Model Pure MoE Pool.

1. Consolidating DeepSeek into V4.1 Flash

Previous Eco iterations split execution across two legacy endpoints: V4-Flash-0731 (routine tasks) and V4-Pro-0813 (algorithmic tasks). With DeepSeek's upstream CED model release, DeepSeek V4.1 Flash cleanly absorbs both lanes. Operating with an agile ~8B active prefill compute footprint, it is significantly more energy-efficient than both predecessor models while delivering 74.2% on DeepSWE and 97.3% on MATH-500.

2. Removing Qwen 3.8 Max (qwen/qwen3.8-2.4t-a95b)

We took a hard look at Qwen 3.8 Max's inclusion in Eco. While Qwen 3.8 Max utilizes sparse MoE routing, it possesses a massive 2.4-trillion-parameter total footprint, activates 95 billion parameters per token, and requires multi-node cluster serving. It was a glaring compute and parameter outlier in a tier dedicated to low-energy efficiency.

Crucially, benchmark evaluations showed that other models already in the Eco pool significantly outperformed Qwen 3.8 Max in its primary engineering lanes:

• Real-World Coding (DeepSWE v1.1): DeepSeek V4.1 Flash scores 74.2% and Gemini 3.8 Flash scores 61.4%, both towering over Qwen 3.8 Max's 56.6%.
• Terminal Execution (Terminal-Bench 2.1): Gemini 3.8 Flash reaches 90.8% and V4.1 Flash reaches 90.6%, compared to Qwen 3.8 Max's 86.6% (vendor) / 81.3% (independent AA).

By removing Qwen 3.8 Max and reassigning complex software engineering, architecture planning, and multi-file refactoring to DeepSeek V4.1 Flash:

• Coding Quality Surges: Users receive a +17.6 point jump on DeepSWE v1.1 (74.2% vs 56.6%).
• Active Compute Plummets: Active parameters per token drop by ~83% (from 95B down to ~16B decode).
• Execution Speed Accelerates: Sub-second time-to-first-token with rapid ~200–420 tps generation.

The Settled 5-Model Gradient

With Qwen 3.8 Max removed, Eco v14 operates with 5 tightly disciplined, non-overlapping champions where 100% of the pool runs on sparse MoE or hybrid linear architectures activating only 8B to 55B parameters per token:

The results of this restructuring speak for themselves:

• Active Parameter Ceiling Reduced: Slashed from 95B active compute down to 55B, ensuring every response stays strictly within high-efficiency bounds.
• Pure MoE Consistency: 100% of models run on sparse MoE or hybrid linear architectures activating only 8B to 55B parameters per token.
• 100% 1M Context: Every single model in the Eco tier natively supports a 1-million-token context window.

---

4. Free Tier: Tool-Hardened Reliability

Canonical Pool: deepseek/deepseek-v4-flash-0731, tencent/hy3, z-ai/glm-5.3-flash

For our Free tier, our engineering objective was clear: ensure that every newcomer experiences fast, accurate, and completely reliable intelligence without encountering tool calling errors or hallucinations.

ARMES agents frequently utilize web search and context retrieval tools. Lighter models often fail on tool chains—either printing raw JSON syntax into chat or hallucinating missing parameters. The Free v11 pool eliminates this friction with three dedicated workhorses:

• DeepSeek V4 Flash 0731 (Conversational Workhorse): Handles everyday conversation, general writing, text summarization, and single-file utility scripts. Because V4 Flash 0731 remains independently hosted by third-party providers on OpenRouter, keeping it in Free maintains sustainable unit economics and proven conversational stability for everyday interactions.
• Tencent HY3 (Grounded Research & Retrieval): Specifically trained to acknowledge missing evidence rather than fabricate facts, HY3 achieves an extraordinary 5.4 percent hallucination rate, along with 79.1 percent on MCP-Atlas and 84.2 percent on BrowseComp. It executes all live web search queries, current events, and notes retrieval with pinpoint precision.
• Z.ai GLM-5.3-Flash (Peak Intelligence Anchor): Scoring 57 on the Artificial Analysis Intelligence Index, GLM-5.3-Flash brings near-frontier power to our Free tier. It handles multi-file software engineering (DeepSWE 63.4%), terminal debugging (Terminal-Bench 84.3%), and complex STEM reasoning.

---

5. Expanded Free Daily Limits & How We Run the Free Tier

A Note on How We Run the Free Tier: Real Privacy Costs Real Compute

If you have signed up for ARMES recently or visited our public access pages, you have likely encountered the note we share in our welcome emails and on our homepage:

"On major Big Tech platforms, 'free' accounts are subsidized by venture capital and paid for with your private data. Your prompts, code, and creative thoughts are routinely harvested to train corporate AI models. If you are not paying, your intellectual property is the product.

ARMES never trains on, profiles, or sells your data. Every query routes through enterprise Zero Data Retention (ZDR) and is not retained by providers. Because real privacy costs real compute, our Free tier provides private queries per day on capable economy models so you can experience sovereign AI risk-free."

This distinction is fundamental to the architecture of ARMES. Legacy corporate platforms can offer seemingly unlimited free access because they monetize your data on the backend. At ARMES, every single request is processed under strict enterprise Zero Data Retention contracts without data monetization or model training. Because we refuse to subsidize free access through surveillance or data harvesting, free compute represents real, uncompromised infrastructure.

Why Expanding to 25 Daily Queries Matters

When we first launched, free accounts were capped at 10 messages per day with an additional rolling weekly limit. While sufficient for a brief taste of the platform, 10 messages often ran out mid-stride when a user was debugging a stubborn script, structuring an outline, or conducting multi-turn research.

By re-architecting the Free pool around ultra-efficient, tool-hardened engines—pairing DeepSeek V4 Flash 0731, Tencent HY3, and GLM-5.3-Flash—we achieved sustainable unit economics that allow us to dramatically expand free access:

• Daily allowance increased to 25 messages per day: A 2.5x expansion in daily query volume.
• Weekly limit completely eliminated: We removed the rolling weekly cap entirely. Your allowance refreshes cleanly every 24 hours.

This expansion allows free users to complete two to three comprehensive, multi-turn problem-solving sessions every single day. Whether you are debugging code, exploring research with live web tools, or organizing your personal notes, you can experience sovereign AI without rationing prompts.

The Contrast: Unrestricted Access on Paid Plans

While 25 daily queries provide a private, reliable workspace for everyday tasks, subscribers on our paid plans—Eco ($9/mo), Pro ($19/mo), and Ultra ($200/mo)—enjoy NO daily query count limits.

Upgrading removes all daily query count restrictions, unlocking unrestricted access to our full constellation of frontier models across OpenAI, Anthropic, Google, xAI, DeepSeek, and Alibaba, alongside 40+ specialist agents, persistent workspace memory, and encrypted collaboration.

---

Summary of Live Routing Pools

---

Zero Data Retention Across Every Tier

Regardless of whether you use ARMES on Free, Eco, Pro, or Ultra, our privacy architecture remains unwavering: Zero Data Retention (ZDR).

Every model in our network processes your requests without storing them. Your conversations, proprietary source code, uploaded files, and agent configurations are never stored by model providers and never used to train machine learning models.

These updates are live right now in your ARMES workspace. Try out the new models and experience the difference in intelligence, speed, and precision.