Six New Models from Four Labs: ARMES Adds GPT-5.6, Claude Sonnet 5, Grok 4.5, and more
The model pool just got significantly deeper.
We are adding six models from four labs to the ARMES platform — the largest single model update since launch. Every one of them runs through the same zero-data-retention infrastructure as everything else on ARMES. Providers process and forget. No training on your data. No retention. No exceptions.
Here is what is new, what each model does best, and how they fit into the platform.
OpenAI: The GPT-5.6 Family (Sol, Terra, Luna)
OpenAI's newest generation is not one model — it is three, designed as durable tiers that complement each other.
GPT-5.6 Sol — Maximum capability
Sol is the ceiling. It leads on professional knowledge work (GDPval-AA #1), management consulting, finance, genomics, medicinal chemistry, health, scientific research, and long-horizon autonomous agents. When you need a polished investment memo, a board-ready presentation, a research synthesis that reconciles conflicting evidence, or a complex multi-stage agent workflow — Sol is the tool.
Key numbers: 88.8% Terminal-Bench 2.1, 64.6% SWE-Bench Pro, 72.7% DeepSWE, 53% Big Finance Bench, 92.2% BrowseComp, Intelligence Index 58.9.
GPT-5.6 Terra — The value leader
Terra sits within a few points of Sol on most evaluations while costing half as much. It exceeds GPT-5.5 across professional, health, science, and coding benchmarks. If Sol is the specialist you bring in for the final deliverable, Terra is the senior professional who handles everything else reliably and economically.
Key numbers: 87.4% Terminal-Bench 2.1, 63.4% SWE-Bench Pro, 51% Big Finance Bench, Intelligence Index 55.0.
GPT-5.6 Luna — Speed and scale
Luna costs one-fifth of Sol and nearly ties Terra on aggregate intelligence benchmarks — while running significantly faster. It is purpose-built for high-volume work: classification, extraction, summarization, drafting, reformatting, test generation, parallel subagents, and first-pass analysis that a stronger model reviews later.
Key numbers: 84.7% Terminal-Bench 2.1, 62.7% SWE-Bench Pro (exceeds GPT-5.5), Intelligence Index 51.2.
One important caveat: Luna's 36% Big Finance Bench score means it should not be your final model for finance judgment, valuation, or accounting interpretation. Terra or Sol should handle those.
Anthropic: Claude Sonnet 5
Claude Sonnet 5 is Anthropic's most agentic Sonnet yet — and it closes much of the gap to Opus 4.8 at Sonnet pricing.
Released June 30, 2026, Sonnet 5 can plan, drive browsers and terminals, and run autonomously at a level that just months ago required the larger and more expensive Opus models. On Humanity's Last Exam with tools, it scores 57.4% — nearly tying Opus 4.8's 57.9%. Its Terminal-Bench 2.1 score of 80.4% is a 13-point jump over Sonnet 4.6.
Where Sonnet 5 shines:
- Agentic coding: 63.2% SWE-Bench Pro (up from 58.1% on Sonnet 4.6)
- Browser and terminal work: 81.2% OSWorld-Verified, 80.4% Terminal-Bench
- Web research: 84.7% BrowseComp (single-agent), 86.6% multi-agent
- Business automation: 13.5% AutomationBench (Sonnet 4.6 scored 5.3%)
- Healthcare: 57.8% HealthBench Professional (Sonnet 4.6 scored 44.2%)
Sonnet 5 introduces a 1M-token context window and configurable effort levels — from lightweight (fast, cheap) to extra-high (approaching Opus quality at proportionally higher cost). It uses a new tokenizer that may produce up to 1.35× more tokens for the same text.
Introductory pricing of $2/$10 per million tokens runs through August 31, 2026, after which it moves to standard Sonnet pricing ($3/$15).
Google: Gemini 3.5 Flash
Gemini 3.5 Flash is the strongest agentic and coding Flash model Google has shipped — and it beats Gemini 3.1 Pro on every coding, agentic, and multimodal benchmark Google evaluated, while running roughly 4× faster.
Key numbers: 83.6% MCP Atlas (beats 3.1 Pro), 76.2% Terminal-Bench 2.1 (beats 3.1 Pro), 57.9% Finance Agent v2 (+14.9 over 3.1 Pro), 1656 Elo GDPval-AA (top quadrant), 55.1% SWE-Bench Pro, 284 t/s output speed.
The speed is the headline. At roughly 284 tokens per second, Flash 3.5 is about 4× faster than other frontier models. This makes it ideal for iterative coding cycles, multi-step tool-use loops, and any workflow where the model makes many sequential calls and speed directly impacts productivity.
It supports configurable thinking levels (minimal/low/medium/high), a 1M-token context window, and a 90% cache discount on repeated input.
Trade-offs versus 3.1 Pro: Flash 3.5 regresses slightly on pure abstract reasoning (HLE: 40.2% vs 44.4%; ARC-AGI-2: 72.1% vs 77.1%) and 128K long-context retrieval. For hard abstract puzzles or needle-in-haystack at extreme depth, 3.1 Pro remains the better choice.
xAI: Grok 4.5
Grok 4.5 is xAI's flagship for software engineering, agentic execution, and technical reasoning. Released July 8, 2026, it scores #4 on the Artificial Analysis Intelligence Index (54) — behind only Fable 5, Opus 4.8, and GPT-5.5 — at a fraction of their cost per task.
The standout story is token efficiency. On SWE-Bench Pro, Grok 4.5 used approximately 15,954 output tokens per task, compared to 67,020 for Claude Opus 4.8 — a 4.2× difference. That means equivalent-quality code completion at dramatically lower cost in production agent pipelines.
Where Grok 4.5 leads:
- Agentic tool use: τ³-Banking 33% (#1 of 28 models evaluated)
- Terminal work: 82–83.3% Terminal-Bench 2.1
- Engineering: 64.7% SWE-Bench Pro, 62% DeepSWE
- Scientific reasoning: 93% GPQA Diamond
- Coding agent performance: AA Coding Agent Index 76 (on par with GPT-5.5)
Grok 4.5 features a 500K-token context window, configurable reasoning effort (high is the default and cannot be fully disabled), built-in server-side tools (web search, X search, code execution), and text + image input.
Important caveats: The hallucination rate increased to 54% (from 25% on Grok 4.3) per Artificial Analysis's Omniscience Index. The 500K context is smaller than the 1M offered by most peers. And the published evidence is heavily weighted toward engineering — broader legal, financial, and writing performance is less documented.
NVIDIA: Nemotron 3 Ultra
Nemotron 3 Ultra is NVIDIA's most capable model — 550 billion total parameters with 55 billion active, using the same hybrid Mamba-Transformer LatentMoE architecture as the existing Nemotron 3 Super, scaled up. Released June 4, 2026.
The headline is throughput. Nemotron 3 Ultra achieves 5.9× higher inference throughput than comparable open models while maintaining competitive accuracy. In agent benchmarks, it completed tasks using 30% fewer total tokens than comparable models — lowering cost-to-task-completion materially.
Where it excels:
- Long context: RULER @1M: 94.7% — best-in-class for open models
- Competitive math: IOI 2025: 570, IMOAnswerBench: 88.6–92.3%
- Code: LiveCodeBench v6: 89.0%, SWE-Bench Verified: 70.7–71.9%
- Agent productivity: PinchBench: 90.0%, IFBench: 81.7%
- Scientific reasoning: GPQA (no tools): 87.0%
It supports 11 natural languages and 43 programming languages, was pre-trained on 20 trillion tokens, and ships under an open-weights license (OpenMDW 1.1).
All models. Same privacy. Always.
Every model added today operates under the same zero-data-retention guarantee as the rest of the ARMES platform:
- Providers process your request and immediately discard it
- No model is trained on your data
- No provider retains your prompts or outputs
- ARMES does not transmit your identity as part of the request
- Your conversations remain for your eyes only
This is not conditional on the model you choose, the plan you are on, or the feature you use. It is the architecture.
Where to find these models
All six models are available now for manual selection on Pro and Ultra tiers. Open your chat, click the model selector, and choose any of them directly.
As we evaluate each model on real ARMES workloads — The current model line up are always visible at armes.ai/ai-access.
The model pool keeps growing
ARMES now provides access to 60+ AI models from 17 labs — all private, all zero-data-retention, all available from one platform. We add models when they meet our quality and privacy requirements, and we never remove access to models you have come to rely on.
The full, always-current model reference is at armes.ai/ai-access.
Joseph Founder, ARMES
Written by
ARMES Team
From the team building ARMES — private AI that puts every frontier model in one place.