Model Landscape — June 2026

Purpose: Comprehensive reference of frontier LLM models, their specs, pricing, modalities, and best-fit use cases for building a model-agnostic agent harness.

Date researched: 2026-06-23

Source: OpenRouter model pages (live data)

1. What Was Researched

A curated set of 25 models across 11 providers was profiled using OpenRouter's standardized cost/benchmark data. Models were selected to represent the full spectrum needed by a multi-model agent harness:

• Frontier reasoning models (primary agent loop)
• Coding-specialized models (subagent coding tasks)
• Fast/flash models (high-throughput, low-latency tier)
• Nano/mini models (cost-efficient inner loops, classification)
• Audio/voice models (voice integration)
• Embedding models (RAG, semantic search, memory)
• Reranking models (retrieval pipeline optimization)

2. Sources

All data sourced from OpenRouter model pages on 2026-06-23:

3. Model Profiles — Frontier Reasoning Models

3.1 xAI Grok 4.3

Description: Reasoning model suited for agentic workflows, instruction-following tasks, and applications requiring high factual accuracy. Supports configurable reasoning effort levels. 1M context with no output limit makes it excellent for long-document analysis, deep research, and multi-step agentic tasks.

Best Use Cases for Agent Harness:

• Primary agent loop model (long-horizon tasks)
• Deep research and analysis subagent
• Multi-step reasoning and planning
• Image understanding in agentic workflows

3.2 Z.ai GLM 5.2

Description: Large-scale reasoning model suited for long-horizon agent workflows, project-level software engineering, and complex multi-step automation. Particularly strong at coding and tool use across long-running tasks — maintains engineering context and follows standards consistently through full development workflows.

Best Use Cases for Agent Harness:

• Project-level coding agent (requirements → deployment in single task)
• Long-running automation workflows
• Agent-orchestrated engineering pipelines
• Cost-effective alternative to top-tier reasoning models ($0.98 in vs $1.25 Grok)

3.3 Anthropic Claude Fable 5

Description: Mythos-class model built for autonomous knowledge work and coding. Designed for long-running, complex, asynchronous tasks — hours/days/weeks of work. Executes well-scoped tasks with few mistakes, automatically self-correcting through verification loops. Ships with robust safeguards.

Best Use Cases for Agent Harness:

• Maximum-autonomy agent: tasks that take hours/days
• Asynchronous background agent with self-correction
• Complex multi-step knowledge work
• ⚠️ Cost warning: 8–20x more expensive than competitors. Reserve for highest-value tasks.

3.4 Anthropic Claude Opus 4.8

Description: Anthropic's established frontier coding and reasoning model. Similar pricing to Fable 5 but positioned as the reliable workhorse for complex coding and knowledge work.

Best Use Cases for Agent Harness:

• Production-grade coding agent
• Complex reasoning tasks requiring high reliability
• Same cost tier as Fable 5 — use when proven reliability is preferred over newest capabilities

3.5 OpenAI GPT-5.5

Description: OpenAI's latest frontier model. Successor to GPT-5, representing the highest capability tier from OpenAI's lineup.

Best Use Cases for Agent Harness:

• Top-tier reasoning and generation
• When OpenAI ecosystem compatibility is required
• Pair with GPT-5.4 Mini / Nano for cost tiering

3.6 NVIDIA Nemotron 3 Ultra (550B / 55B active)

Description: Open frontier-reasoning and orchestration model. Hybrid Transformer-Mamba MoE architecture with 55B active parameters. Designed for agent orchestration, coding agents, deep research, and complex enterprise tasks. High-throughput inference designed for high-volume agent pipelines.

Best Use Cases for Agent Harness:

• Best value frontier reasoning ($0.50 input — cheapest 1M-context model)
• Agent orchestration layer
• High-volume pipeline processing
• Self-hosted option (open weights)
• ⚠️ Output limited to 16K tokens — not ideal for long-form generation

3.7 DeepSeek V4 Pro

Description: DeepSeek's flagship reasoning model. Known for strong coding and reasoning capabilities at competitive pricing.

Best Use Cases for Agent Harness:

• Cost-effective reasoning and coding
• Alternative to Grok/GLM for primary agent loop
• Strong on benchmarks relative to price

3.8 Qwen 3.7 Max

Description: Qwen's maximum-capability model. Part of the Qwen 3.7 family representing Alibaba's frontier AI capabilities.

Best Use Cases for Agent Harness:

• Primary reasoning when Qwen ecosystem is preferred
• Multi-language tasks (strong CJK support)
• Alternative frontier model for diversity in model routing

4. Model Profiles — Coding-Specialized Models

4.1 MoonshotAI Kimi K2.7 Code

Description: Coding-focused model built to complete end-to-end programming tasks reliably over long contexts. Native multimodal MoE architecture. Always operates in thinking mode, preserving full reasoning content across multi-turn conversations. Targets long-horizon coding, agentic task decomposition, and multi-turn dialogue.

Best Use Cases for Agent Harness:

• Dedicated coding subagent — best price/performance for code
• Long-horizon code generation (256K in + 256K out)
• Agentic task decomposition
• Self-hosted coding agent (open weights, 32B active)
• ⚠️ Always-on thinking — not configurable, plan for reasoning token costs

5. Model Profiles — Fast / Flash Tier

5.1 Google Gemini 3.5 Flash

Description: Google's flash-tier model optimized for speed and cost efficiency while maintaining strong reasoning capabilities. Gemini 3.5 Flash extends the proven Flash lineage with improved capabilities.

Best Use Cases for Agent Harness:

• High-throughput inner loops
• Tool call classification and routing
• Fast summarization and extraction
• Cost-efficient bulk processing

5.2 StepFun Step 3.7 Flash

Description: StepFun's flash-tier model designed for rapid, low-latency responses.

Best Use Cases for Agent Harness:

• Ultra-fast classification and routing
• Low-latency inner loops
• Budget-conscious high-volume tasks

5.3 DeepSeek V4 Flash

Description: DeepSeek's flash variant optimized for speed and cost. Maintains strong reasoning at dramatically reduced latency.

Best Use Cases for Agent Harness:

• Fast coding assist (inner loop)
• High-throughput data processing
• Cost-optimized bulk operations

6. Model Profiles — Mini / Nano Tier (Cost-Efficient)

6.1 OpenAI GPT-5.4 Mini

Description: OpenAI's cost-optimized model balancing capability and efficiency. Successor to GPT-4o mini.

Best Use Cases for Agent Harness:

• Classification and routing decisions
• Simple tool-call parsing
• Cost-efficient multi-turn dialogue
• Inner-loop validation

6.2 OpenAI GPT-5.4 Nano

Description: OpenAI's smallest and cheapest model. Designed for high-volume, low-complexity tasks.

Best Use Cases for Agent Harness:

• Intent classification (near-zero cost)
• Token counting and metadata extraction
• Simple formatting and transformation
• Heartbeat / health-check pings
• Guard-rail input validation

6.3 Qwen 3.7 Plus

Description: Qwen's mid-range model offering strong capabilities at moderate cost. Good balance between Qwen Max and budget models.

Best Use Cases for Agent Harness:

• Cost-effective general reasoning
• Multi-language support (strong CJK)
• Mid-tier subagent tasks

6.4 MiniMax M3

Description: MiniMax's latest model offering competitive performance.

Best Use Cases for Agent Harness:

• General-purpose subagent tasks
• Provider diversity for resilience

7. Model Profiles — Audio / Voice Models

7.1 xAI Grok Voice TTS 1.0

Description: xAI's voice synthesis model. Generates natural speech from text input.

Best Use Cases for Agent Harness:

• Voice output for conversational agent interfaces
• Spoken notifications and alerts
• Accessibility features
• Voice-enabled agent interactions

7.2 OpenAI GPT Audio

Description: OpenAI's full audio model supporting both audio input and output. Enables native voice conversations.

Best Use Cases for Agent Harness:

• Full voice conversation pipeline
• Audio transcription + response generation
• Voice-first agent interfaces

7.3 OpenAI GPT Audio Mini

Description: Cost-optimized version of GPT Audio. Lower latency and cost for voice interactions.

Best Use Cases for Agent Harness:

• Cost-efficient voice interactions
• Real-time voice agent (latency-sensitive)
• Voice command parsing

8. Model Profiles — Embedding Models

8.1 Google Gemini Embedding 2

Description: Google's latest text embedding model for semantic search and retrieval.

Best Use Cases for Agent Harness:

• Semantic memory — embed and retrieve agent context
• RAG pipeline (document embeddings)
• Similarity search for skill/tool matching
• Codebase semantic indexing

8.2 OpenAI Text Embedding 3 Large

Description: OpenAI's highest-quality embedding model. Supports configurable dimensions (256–3072) via dimensions parameter.

Best Use Cases for Agent Harness:

• Highest-quality semantic search for agent memory
• Knowledge base embeddings
• Code search and retrieval
• When embedding quality matters more than cost

8.3 OpenAI Text Embedding 3 Small

Description: OpenAI's cost-optimized embedding model. 5x cheaper than text-embedding-3-large.

Best Use Cases for Agent Harness:

• Cost-efficient embeddings for high-volume indexing
• Real-time embedding during agent conversations
• When speed/cost preferred over maximum quality
• Lightweight memory storage

9. Model Profiles — Reranking Models

9.1 Cohere Rerank 4 Pro

Description: Cohere's highest-quality reranking model. Cross-encoder architecture for maximum relevance scoring accuracy.

Best Use Cases for Agent Harness:

• RAG pipeline reranking — improve retrieval quality
• Tool/skill selection (rank candidate tools by relevance)
• Document relevance scoring for context selection
• Multi-document synthesis prioritization

9.2 Cohere Rerank 4 Fast

Description: Speed-optimized reranker for latency-sensitive applications.

Best Use Cases for Agent Harness:

• Real-time reranking during agent conversations
• High-throughput document filtering
• When latency matters more than maximum accuracy

9.3 Cohere Rerank v3.5

Description: Previous-generation reranker. Still widely deployed and well-tested.

Best Use Cases for Agent Harness:

• Stable, battle-tested reranking
• Fallback option if Rerank 4 has availability issues

10. Key Findings — Pricing Comparison

Frontier Reasoning Models (per 1M tokens)

Key insight: A 20x pricing spread exists between the cheapest (Nemotron $0.50 in) and most expensive (Claude $10 in) frontier models. A model-agnostic harness can exploit this by routing tasks based on complexity and value.

Cost-per-1K-turn Estimate (1K input + 2K output average)

11. Architecture Implications for Agent Harness

11.1 Model Tiering Strategy

The harness should implement a 5-tier model routing system:

┌─────────────────────────────────────────────────────────┐
│ TIER 0: NANO                                            │
│ GPT-5.4 Nano                                            │
│ Use: Classification, routing, validation, guard rails   │
│ Cost: ~$0.0001/turn                                     │
├─────────────────────────────────────────────────────────┤
│ TIER 1: MINI / FLASH                                    │
│ GPT-5.4 Mini, Gemini 3.5 Flash, DeepSeek V4 Flash      │
│ Use: Tool-call parsing, summarization, fast responses   │
│ Cost: ~$0.001/turn                                      │
├─────────────────────────────────────────────────────────┤
│ TIER 2: MID-RANGE                                       │
│ Qwen 3.7 Plus, MiniMax M3, Step 3.7 Flash              │
│ Use: General subagent tasks, multi-turn dialogue        │
│ Cost: ~$0.003/turn                                      │
├─────────────────────────────────────────────────────────┤
│ TIER 3: FRONTIER                                        │
│ Grok 4.3, GLM 5.2, Nemotron 3 Ultra, DeepSeek V4 Pro   │
│ Kimi K2.7 Code (coding), Qwen 3.7 Max                  │
│ Use: Primary agent loop, complex reasoning, coding      │
│ Cost: ~$0.005–0.01/turn                                 │
├─────────────────────────────────────────────────────────┤
│ TIER 4: ULTRA-PREMIUM                                   │
│ Claude Fable 5, Claude Opus 4.8, GPT-5.5               │
│ Use: Maximum-autonomy, days-long tasks, highest quality │
│ Cost: ~$0.10+/turn                                      │
└─────────────────────────────────────────────────────────┘

11.2 Specialized Model Slots

Beyond the tiering system, the harness should have dedicated slots for:

11.3 Context Window Implications

Design principle: The harness should track context consumption per model and automatically route to larger-context models when conversation history grows.

11.4 Open Weights vs Closed

Design principle: The harness should support both API-based and local (Ollama, vLLM, TGI) model backends through a unified OpenAI-compatible client. Study LiteLLM's BaseConfig pattern only if you build a self-hosted translation proxy; for most setups, point at Ollama or OpenRouter directly.

12. What Is Confirmed

1. 20x pricing spread exists between cheapest and most expensive frontier models — routing by task value is critical
2. 1M context windows are now standard for top-tier models (5/6 frontier models support it)
3. Open-weight frontier models (GLM 5.2, Nemotron 3, Kimi K2.7 Code) are viable alternatives to closed models
4. Specialized models (embedding, reranking, voice) require dedicated harness integration slots — not just text-in/text-out
5. Configurable reasoning effort (Grok: none/low/medium/high; GLM: high/xhigh) is a new dimension the harness must support
6. MoE architectures are dominant in cost-effective frontier models (Nemotron 55B/550B, Kimi 32B/1T)

13. What Is Uncertain

1. Benchmark data: OpenRouter pages render benchmarks client-side; exact benchmark scores were not captured in this pass. Live browser scraping or API access would be needed for Artificial Analysis benchmark comparisons.
2. Exact pricing for some models (GPT-5.5, GPT-5.4 Mini/Nano, Qwen models, DeepSeek models, Step 3.7 Flash, MiniMax M3): These were fetched but specific dollar amounts were in client-rendered JavaScript. The meta descriptions for these models did not include explicit pricing.
3. Rate limits and quotas: Not captured from OpenRouter pages. These vary significantly by provider and plan.
4. Latency benchmarks: Time-to-first-token and tokens-per-second data requires Artificial Analysis integration.

14. How This Applies to Building a Model-Agnostic Agent Harness

This research directly informs:

1. Model registry design — the harness needs a configuration system that tracks all 25+ model specs (pricing, context, modalities, reasoning config)
2. Cost-optimized routing — the 20x pricing spread means intelligent routing between tiers saves 90%+ on API costs
3. Multi-modal support — the harness can't just handle text; it needs audio I/O, file I/O, embedding, and reranking endpoints
4. Reasoning effort control — new API parameter (reasoning_effort) must be surfaced in the harness configuration
5. Provider fallback — models with 20 providers (GLM 5.2) vs 4 providers (Nemotron 3) have different reliability profiles
6. Self-hosting path — open-weight models (GLM 5.2, Nemotron 3, Kimi K2.7) enable cost-free local deployment for high-volume use

Document created: 2026-06-23 | All data sourced from OpenRouter model pages