Snapshot · 2026-06-11 · read the note ↗
The Frontier Models Room
A weekly read on the AI model landscape — frontier closed/open, small & on-device, agentic frameworks, and classifiers.
Comparison & trade-offs
Side-by-side at a glance. Click any card above for full detail. Use the tabs to filter by category.
| Model / tool | Category | Latest | License | Headline | Best for | Watch-out |
|---|---|---|---|---|---|---|
| Claude (Opus 4.8) Anthropic | Frontier · Closed | Opus 4.8 · Fable 5 (new top tier) 2026-05-28 | Proprietary (API) | #1 Artificial Analysis Intelligence Index — 61.4 | Complex multi-file coding and refactors | Most expensive flagship pricing among peers |
| GPT-5.5 (OpenAI) OpenAI | Frontier · Closed | GPT-5.5 / 5.5 Pro 2026-04-23 | Proprietary (API) | 82.7% Terminal-Bench 2.0 (SOTA) · #2 Intelligence Index ~60.2 | Agentic coding and autonomous computer-use tasks | Trails Anthropic on SWE-bench Pro |
| Gemini 3.5 (Google) Google DeepMind | Frontier · Closed | 3.5 Flash (GA) · 3.5 Pro (rolling out) 2026-05-19 | Proprietary (API) | Frontier coding at Flash price + ~4x speed | High-volume agentic workloads | Flagship 3.5 Pro not yet fully GA |
| Grok 4.3 (xAI) xAI | Frontier · Closed | Grok 4.3 2026-04-30 | Proprietary (API) | Intelligence Index ~53 at $1.25 / $2.50 per M | Cost-sensitive high-context tasks | Lower absolute intelligence than the top three |
| Amazon Nova 2 Amazon | Frontier · Closed | Nova 2 Pro / Premier 2026 | Proprietary (AWS Bedrock) | 1M context, low price; AA Index ~19–23 | Cost-sensitive AWS workloads | Well below frontier intelligence |
| DeepSeek V4 DeepSeek-AI | Frontier · Open | V4-Pro (1.6T-A49B) · V4-Flash (284B-A13B) 2026-04-24 | MIT | 80.6% SWE-bench Verified · 93.5% LiveCodeBench | Open coding and math at frontier level | 1.6T params — impractical to self-host for most |
| Kimi K2.6 (Moonshot) Moonshot AI | Frontier · Open | K2.6 (1T-A32B) 2026-04-20 | Modified MIT | 58.6 SWE-bench Pro (open-weight leader) · AA Index 54 | Open agentic coding and long-horizon tasks | 1T params — heavy to self-host |
| GLM-5.1 (Z.ai) Z.ai (Zhipu) | Frontier · Open | GLM-5.1 (744B-A40B) 2026-03-27 | MIT | 77.8% SWE-bench Verified (GLM-5) · 58.4 SWE-bench Pro (5.1) | Open coding and agentic workflows | Text-only (no native multimodal) |
| Qwen3.6 (Alibaba, open) Qwen Team, Alibaba | Frontier · Open | Qwen3.6-27B Dense · 3.6-35B-A3B MoE 2026-04-22 | Apache-2.0 (open models) | 27B dense reportedly matches Claude 4.5 Opus on Terminal-Bench 2.0 | Self-hosted coding at moderate cost | Version sprawl is confusing |
| Llama 4 (Meta) Meta | Frontier · Open | Scout (109B-A17B) · Maverick (400B-A17B) 2025-04-05 | Llama 4 Community License (custom) | 10M-token context (Scout); now ~14 months old | Existing Llama-based stacks | ~14 months old — off the open frontier |
| Mistral Large 3 Mistral AI | Frontier · Open | Large 3 (675B-A41B) 2025-12-02 | Apache-2.0 | #2 OSS non-reasoning on LMArena at launch | Sovereign / EU-hosted deployments | Trails the Chinese open frontier on coding / agentic |
| gpt-oss (OpenAI) OpenAI | Frontier · Open | gpt-oss-120b · gpt-oss-20b 2025-08-05 | Apache-2.0 | 120B ≈ o4-mini reasoning; AA Index 33 | Local / single-GPU reasoning | ~10 months old and lagging newer open models |
| NVIDIA Nemotron 3 NVIDIA | Frontier · Open | Nemotron 3 Ultra (550B-A55B) 2026-06-04 | OpenMDW-1.1 | AA Intelligence Index 48 — best US open model | US / Western open-weight deployments | Still trails the Chinese open frontier (~6 index pts) |
| MiniMax M3 MiniMax | Frontier · Open | M3 2026-06-01 | Likely commercially restricted (verify) | AA Index 55 (claimed, top open) — unverified | Cost-sensitive long-context tasks (once weights land) | Benchmarks not independently verified |
| Microsoft Phi-4 Microsoft Research | Small · Reasoning | Phi-4-reasoning-vision-15B · Phi-4-mini-reasoning (3.8B) 2026-03-04 | MIT | Phi-4-mini-reasoning rivals o1-mini on math | On-device / cheap high-volume reasoning | Narrow strengths (math/science-leaning) |
| Google Gemma 4 (small/edge) Google DeepMind | Small · On-device | Gemma 4 E2B / E4B / 12B (+ QAT) 2026-04-02 | Apache-2.0 | Sub-1GB on mobile (E2B QAT, 0.84GB text) | On-device / mobile multimodal | Tops out at 31B — not a trillion-param competitor |
| Qwen3.5 Small Series Qwen Team, Alibaba | Small · Multimodal | 0.8B / 2B / 4B / 9B 2026-03-02 | Apache-2.0 | 9B matches/surpasses gpt-oss-120B on several evals | Edge / IoT / mobile (0.8B–2B) | Headline numbers still secondary-sourced |
| HuggingFace SmolLM3 Hugging Face | Small · Fully open | SmolLM3-3B 2025-07-08 | Apache-2.0 | Punches above its weight at the 3B scale | Reproducible research base | No newer version since Jul 2025 |
| IBM Granite 4.0 Nano IBM | Small · Enterprise edge | Nano 350M / ~1B 2025-10-30 | Apache-2.0 | 70–80% memory cut vs transformer-only | On-device enterprise automation | Lower raw benchmarks than Qwen/Gemma at size |
| Liquid AI LFM2 Liquid AI | Small · Fastest on-device | LFM2.5 · LFM2-VL · LFM2.5-8B-A1B 2026-01-06 | Open weights (LFM Open License) | ~2x faster CPU decode/prefill vs Qwen3 | Smartphones, laptops, wearables | Smaller ecosystem than Qwen / Gemma |
| Apple Foundation Models 3 Apple | Small · On-device | AFM 3 Core / Core Advanced 2026-06 | Proprietary (on-device API) | Native multimodal on-device, no cloud round-trip | Privacy-sensitive iOS / macOS app features | Not open-weight (Apple API only) |
| OpenClaw Peter Steinberger / OpenClaw Foundation | Agent · Autonomous | rolling (foundation-stewarded) 2026-01-30 | MIT | ~310K–350K+ GitHub stars (most in history) | Power-user personal automation | Serious security risk (Cisco found a data-exfiltration skill) |
| OpenManus (+ Manus) FoundationAgents (MetaGPT team) | Agent · Autonomous | rolling 2025-03 | Apache-2.0 | ~52K–56K GitHub stars | Open autonomous research / task agents | Less polished than the original Manus product |
| OpenHands All-Hands-AI | Agent · Coding | v1.6.0 2026-03-30 | MIT | ~72% SWE-bench Verified (w/ Claude Sonnet 4.5) | Autonomous bug-fixing / issue resolution | Best results need a strong (paid) backing model |
| LangGraph / LangChain LangChain Inc. | Framework · Orchestration | 1.0 GA (Oct 2025), 1.x 2025-10-22 | MIT | Industry-default agent stack | Production multi-agent systems | More concepts to learn than lightweight libs |
| CrewAI CrewAI Inc. | Framework · Multi-agent | ~v1.14 2026 | MIT | ~47.8K★ · ~2B executions in 12 months | Multi-agent business workflows | Less low-level control than LangGraph |
| Microsoft Agent Framework Microsoft | Framework · Enterprise | v1.0 GA 2026-04-07 | MIT | Consolidates AutoGen + Semantic Kernel | Enterprise / Azure agent apps | Migration needed from AutoGen / SK |
| OpenAI Agents SDK OpenAI | Framework · Orchestration | v0.17.1 2026-05 | MIT | ~26K+★ · Swarm deprecated in its favor | OpenAI-stack multi-agent apps | OpenAI-centric |
| Google ADK 2.0 | Framework · Enterprise | ADK 2.0 2026-05-19 | Apache-2.0 | Multi-language; on-device Android (Gemini Nano) | Enterprise agents on Google Cloud | Google Cloud / Gemini-leaning |
| smolagents (HuggingFace) Hugging Face | Framework · Code-first | rolling 2025 | Apache-2.0 | ~26K★ · CodeAgent ~30% fewer steps | Learning agent fundamentals | Fewer enterprise features |
| Claude Agent SDK Anthropic | Framework · Vendor | rolling (Managed Agents beta) 2026 | Proprietary (Anthropic) | Subagents, MCP, managed cron agents | Claude-powered coding / ops agents | Tied to Claude models (not model-agnostic) |
| Dify LangGenius | Platform · No-code | ~1.1.2 2026 | Dify Open Source License (≈Apache-2.0+) | ~138K★ · 1M+ apps deployed | Fast RAG / agent app building | License has a few restrictions vs pure OSS |
| n8n (AI agents) n8n GmbH | Platform · Automation | rolling 2026 | Fair-code (Sustainable Use License) | ~150K–192K★ · $2.5B valuation | AI-in-the-loop business automation | Fair-code, not OSI-open |
| ModernBERT Answer.AI + LightOn | Encoder · Backbone | base 149M / large 395M 2024-12 | Apache-2.0 | 8K context; SOTA base-encoder GLUE at release | Classification / NER / retrieval backbones | Raw GLUE still trails DeBERTa-v3-large |
| DeBERTa-v3 Microsoft | Encoder · NLU king | v3 (base 86M / large 304M) 2021 | MIT | ~91.4 GLUE average — still the top encoder | Max-accuracy short-text classification | Only 512-token context |
| Qwen3-Embedding Qwen Team, Alibaba | Embedding · Open leader | 0.6B / 4B / 8B 2025-06-05 | Apache-2.0 | #1 open on MMTEB (70.58) · 80.68 MTEB-Code | Production RAG / semantic search | 8B is heavier than tiny embedders |
| EmbeddingGemma | Embedding · On-device | 300M 2025-09-04 | Gemma license | Best open multilingual model under 500M on MTEB | On-device / offline semantic search | Below big embedders on absolute MTEB |
| Gemini Embedding 001 | Embedding · API leader | gemini-embedding-001 2025-05-20 | Proprietary (API) | 68.32 MTEB (#1) · #1 MMTEB multilingual | Best-quality managed RAG | Proprietary + per-token cost |
| Qwen3-Reranker Qwen Team, Alibaba | Reranker · Open leader | 0.6B / 4B / 8B 2025-06 | Apache-2.0 | Leading open reranker, 100+ languages | Second-stage RAG reranking | Cross-encoders are slower than bi-encoders |
| Cohere Rerank 4 Cohere | Reranker · API leader | v4.0 (Pro / Fast) 2026 | Proprietary (API) | ~1629 ELO (Agentset reranker leaderboard) | Managed RAG reranking | Proprietary + per-call cost |
| gpt-oss-safeguard OpenAI | Classifier · Safety | 120B / 20B 2025-10-29 | Apache-2.0 | Policy-at-inference classification with visible reasoning | Custom content moderation | Heavier than a small classifier |
| Llama Guard 4 Meta | Classifier · Safety | 12B 2025 | Llama Community License | Native text+image moderation (MLCommons taxonomy) | Multimodal content moderation | 12B is heavier than small classifiers |
| GLiNER2 Urchade / community | NLP · NER / extraction | GLiNER2 2025 | Apache-2.0 | Beats DeBERTa-v3 and LLMs on zero-shot NER | Zero-shot NER / entity extraction | Narrower than a general LLM |
Which model should I use?
Quick decision guide by scenario.
Best model for complex agentic coding
Claude (Opus 4.8 / Fable 5) leads SWE-bench and long-horizon coding; GPT-5.5 tops terminal and computer-use agents. For an open, self-hosted option, DeepSeek V4 or Kimi K2.6 are the strongest.
Cheapest near-frontier intelligence at scale
Gemini 3.5 Flash ($1.50/$9, ~4x speed) and Grok 4.3 ($1.25/$2.50) give the best price-to-intelligence. Among open models, a self-hosted Qwen3.6-27B is extremely cost-effective.
Best open-weight model to self-host
At the trillion-param tier: DeepSeek V4 (MIT), Kimi K2.6 (multimodal), or GLM-5.1 (MIT). For a single-GPU-friendly dense model, Qwen3.6-27B (Apache-2.0). Best US-open model: NVIDIA Nemotron 3 Ultra.
Best small model for on-device / mobile
Gemma 4 E2B/E4B (Apache-2.0, native audio+vision, sub-1GB) or Qwen3.5-4B (native multimodal). For raw speed, Liquid LFM2; for regulated edge, IBM Granite 4 Nano; on Apple devices, AFM 3.
Best tiny model for cheap reasoning
Phi-4-mini-reasoning (3.8B, MIT) rivals o1-mini on math and beats larger R1-distilled models — the best reasoning-per-parameter pick for high-volume or on-device math/science.
Best embeddings for RAG
Managed: Gemini Embedding 001 (#1 MTEB). Open / self-hosted: Qwen3-Embedding-8B (#1 open, best code retrieval). On-device: EmbeddingGemma (runs in <200MB RAM).
Best reranker for retrieval
Managed: Cohere Rerank 4 (Pro for accuracy, Fast for latency) or Zerank-2. Open / self-hosted: Qwen3-Reranker or bge-reranker-v2-m3.
Best classifier for short-text NLU / zero-shot
DeBERTa-v3 still has the highest raw accuracy for short-text classification and zero-shot NLI; ModernBERT when you need 8K context or speed; GLiNER2 for zero-shot 'extract any entity' NER.
Best content-safety / guardrail classifier
gpt-oss-safeguard (open; interprets your policy at inference with visible reasoning) for flexible moderation; Llama Guard 4 for native text+image filtering; ShieldGemma for lightweight category checks.
Best framework to build production agents
LangGraph for durable graph orchestration; Microsoft Agent Framework or Google ADK for enterprise; OpenAI Agents SDK or smolagents for lightweight; Claude Agent SDK for Claude-native coding/ops agents.
Best autonomous 'personal agent' to run
OpenClaw is the most popular open personal agent (the most-starred repo ever) — but treat security seriously (it has had data-exfiltration findings; sandbox it). OpenManus is the open successor to the now-collapsed Manus for general task automation.
Best autonomous software-engineering agent
OpenHands is the top open-source coding-agent platform (~72% SWE-bench Verified with a strong backing model like Claude Sonnet 4.5).
In the pipeline
What the labs have announced, are rolling out, or are rumored to ship next. Treat rumored items with caution.
OpenAI GPT-5.6 Rumored
Widely expected late June 2026 per prediction markets and a brief Codex routing-log sighting — but unconfirmed. GPT-5.5 remains the current confirmed flagship.
Google Gemini 3.5 Pro Rolling out
Announced at I/O; GA expected June 2026 with a 2M-token context and a 'Deep Think' reasoning mode. The 3.5 Flash tier is already GA.
xAI Grok 5 In training
Training on the Colossus 2 supercluster, reportedly up to 10 trillion parameters, with a Q2/Q3 2026 target. Parameter count and timeline are rumored, not official.
DeepSeek R2 Uncertain
A dedicated R2 reasoning model is rumored (a 32B variant was reported), but it doesn't appear in DeepSeek's official API list — treat as unconfirmed. V4's Think modes already cover much of the reasoning role.
MiniMax M3 weights Pending
Announced June 1 with the top claimed open-weight intelligence score; full weight release and independent benchmark verification are still pending.
Meta Llama 5 ('Avocado') Rumored ~2027
Meta shipped its first closed model (Muse Spark) in Apr 2026; the open Llama roadmap now appears to lag the closed line. Behemoth (2T) was shelved.
Anthropic Fable / Mythos line New tier
A new top 'Fable 5' tier launched Jun 9 2026 above Opus 4.8, with a gated 'Mythos 5' sibling via a limited research program. Headline figures are still firming up.
On the watchlist
Emerging models & tools being tracked but not yet promoted to a full card — pending proven significance.
NeoBERT ↗
250M modern encoder (Mila), ~89 GLUE — matches RoBERTa-large with ~100M fewer params. A strong compact ModernBERT alternative.
Why watch: Could become a go-to compact encoder; watch for adoption and a clear license.
EuroBERT / mmBERT / MrBERT ↗
Multilingual modern encoders (210M–2.1B) covering 35–110+ languages, all built on the ModernBERT architecture; MrBERT (BSC, 2026) adds legal/bio/Spanish spinoffs.
Why watch: The fast-moving multilingual-encoder frontier; promote whichever wins production adoption.
Ant Ling / Ring (1T) ↗
Ant Group's trillion-scale open MoEs; Ring-2.6-1T is a reasoning model with customizable inference intensity and extreme token efficiency (MIT, per secondary sources).
Why watch: Another credible Chinese trillion-param open entrant; verify license + benchmarks before promoting.
Tencent Hunyuan 3 / StepFun / Baichuan ↗
More open / open-weight Chinese frontier models shipping fast (e.g. Hunyuan Hy3-preview 295B-A21B).
Why watch: The open frontier is crowded and Chinese-led; track which sustain momentum.
Agno ↗
High-performance Python agent runtime (~40K★) claiming ~2µs agent instantiation; markets itself as far faster than LangGraph for high-concurrency swarms.
Why watch: Performance-first alternative to CrewAI/LangGraph; watch enterprise uptake.
Pydantic AI ↗
Type-safe agent framework (v1.x stable; v2 in beta) emphasizing structured outputs and validation.
Why watch: Popular with the Pydantic crowd; promote to a full card when v2 lands.
Zerank-2 (ZeroEntropy) ↗
Tops the Agentset reranker leaderboard (~1638 ELO), edging Cohere Rerank 4 on instruction-following.
Why watch: Newer managed reranker; watch availability and independent evals.