[INFORMATIONAL] Model ranking/recommendations - May 2026

### Description

**_(I took the time to research/generate this for my own usage, sharing in case it's useful to someone else)_**


# Model Recommendations for opencode Agent Roles - May 2026

## Purpose

Recommend the top models for each opencode/oh-my-opencode-slim agent role, ranked by quality, cost, and suitability as of May 2026. Each tier provides top-5 closed and top-5 open-weight options.

## Methodology

Sources:
- Artificial Analysis: throughput, pricing, quality index
- LMSYS Chatbot Arena: Elo ratings
- SWE-bench Verified: software engineering task completion
- Terminal-Bench: agentic terminal performance
- GPQA Diamond: hard science reasoning
- BenchLM, TokenMIX, OpenRouter rankings: cross-platform aggregation
- Aider Polyglot, LM Studio, OpenRouter community benchmarks: coding signals



### Extra context

Tier-to-variant mapping: `low` = simple, `medium` = standard, `high` = complex, `max` = reasoning.

## Reasoning Tier (max variant)

Used by: Oracle, Council summarizer, Council delta.

Criteria: deep reasoning, multi-file code generation, agentic planning, scientific reasoning. Latency is secondary to quality.

### Closed Source

| Rank | Model | Key Strengths | Approx Pricing (per M tokens) | Notes |
|------|-------|---------------|-------------------------------|-------|
| 1 | Claude Opus 4.7 | Best multi-file coding, Terminal-Bench #1, strongest agentic tool use | $15/$75 | Anthropic flagship for complex software engineering. 200K context. |
| 2 | GPT-5.5 | Highest AA Index (60), Terminal-Bench #2, fastest reasoning-class inference | $10/$40 | Best throughput of any frontier model. Excels at agentic chaining. |
| 3 | Gemini 3.1 Pro | GPQA 94.3%, best reasoning per dollar, 2M context | $2/$10 | Best value in reasoning tier. Unmatched context window. |
| 4 | Claude Mythos | Experimental Anthropic variant, reportedly SWE-bench 83.1% | ~$15/$75 | Pre-release. For early adopters running Anthropic's latest reasoning architecture. |
| 5 | GPT-5.5 Pro | Extended reasoning variant of GPT-5.5, highest SWE-bench (84.2%) | $15/$60 | Best raw SWE-bench score. Use Oracle when correctness must not fail. |

### Open Weight

| Rank | Model | Key Strengths | Approx Pricing (per M tokens) | Notes |
|------|-------|---------------|-------------------------------|-------|
| 1 | DeepSeek V4 Pro | MIT license, 80.6% SWE-bench | $0.55/$0.87 | Open-weight champion. Self-hostable. 128K context. |
| 2 | Kimi K2.6 | 79.4% SWE-bench, Chinese + English, 128K context | $0.50/$0.80 | Moonshot AI's latest. Strong multilingual code reasoning. |
| 3 | Qwen 3.7 Max | GPQA 92.1%, multilingual | $0.60/$1.00 | Best for multilingual reasoning. 131K context. |
| 4 | GLM-5.1 | Zhipu AI flagship, competitive with Kimi on code | $0.45/$0.75 | Solid all-rounder on Chinese cloud providers. |
| 5 | Llama 4 Maverick | 405B-class MoE, uncensored variant | $0.50/$0.80 | Apache 2.0. Best for self-hosting at scale. |

## Complex Tier (high variant)

Used by: Council gamma.

Criteria: strong coding and tool use, lower latency requirement than reasoning tier. Balance quality and speed.

### Closed Source

| Rank | Model | Key Strengths | Approx Pricing (per M tokens) | Notes |
|------|-------|---------------|-------------------------------|-------|
| 1 | Claude Sonnet 4.6 | Best quality-to-latency ratio in Claude lineup | $3/$15 | Nearly Opus-quality at 1/5 the cost. |
| 2 | GPT-5.4 | Fast, 128K context, strong Terminal-Bench scores | $2.50/$10 | Less capable than GPT-5.5 but significantly cheaper. |
| 3 | Gemini 3 Flash | 97% of Gemini 3.1 Pro quality at 1/4 the cost | $0.50/$3 | Near reasoning-tier quality at standard-tier prices. |
| 4 | Grok 4.1 | xAI's latest, 128K context, strong math/code reasoning | $3/$12 | Less proven on SWE-bench. Competitive on math and science. |
| 5 | GPT-5.4 Pro | Higher-correctness variant of GPT-5.4 | $5/$15 | Middle ground between Sonnet 4.6 and Opus 4.7 pricing. |

### Open Weight

| Rank | Model | Key Strengths | Approx Pricing (per M tokens) | Notes |
|------|-------|---------------|-------------------------------|-------|
| 1 | DeepSeek V4 Pro | Cross-listed from reasoning tier | $0.55/$0.87 | Versatile enough for both tiers. Self-host alongside reasoning workload. |
| 2 | Kimi K2.6 | Cross-listed. Excellent code at open-weight pricing | $0.50/$0.80 | Good second source for council diversity. |
| 3 | MiMo-V2.5-Pro | 01.AI's latest, 128K context | $0.40/$0.70 | Budget option for complex code generation. |
| 4 | Mistral Large 3 | European languages, good tool use | $0.40/$0.70 | Apache 2.0. Best for EU compliance. |
| 5 | Qwen 3.5-397B | Alibaba MoE model | $0.35/$0.60 | Slightly behind Kimi and DS V4 but ahead of older open models. |

## Standard Tier (medium/low variant)

Used by: Librarian, Explorer, Observer, Designer, Scout.

Criteria: low latency, cost-sensitive, high throughput. These roles are I/O-bound, not reasoning-bound.

### Closed Source

| Rank | Model | Key Strengths | Approx Pricing (per M tokens) | Notes |
|------|-------|---------------|-------------------------------|-------|
| 1 | Gemini 3 Flash | 0.5s TTFT, 97% quality retention from Pro, 1M context | $0.50/$3 | Best standard-tier option. Excellent for librarians and explorers. |
| 2 | Claude Haiku 4.5 | Fastest Claude variant, 200K context | $0.25/$1.25 | Best tool-use in standard tier. Slightly slower than Gemini Flash. |
| 3 | GPT-5.4 Mini | Fast, 128K context, good for structured data extraction | $0.15/$0.60 | Best for high-volume scraping. |
| 4 | GPT-5.4 Nano | Smallest GPT, extremely fast | $0.05/$0.20 | Lowest latency of any closed model. Ideal for rapid-fire explorer queries. |
| 5 | Gemini 3.1 Flash-Lite | Budget Flash variant | $0.15/$0.60 | Best value for Observer and basic designer tasks. |

### Open Weight

| Rank | Model | Key Strengths | Approx Pricing (per M tokens) | Notes |
|------|-------|---------------|-------------------------------|-------|
| 1 | DeepSeek V4 Flash | 79% SWE-bench at 1/10 the cost, 128K context | $0.14/$0.28 | Budget champion. Self-hostable. Good enough for most standard tasks. |
| 2 | MiniMax M2.7 | Hugging Face leader in several categories | $0.10/$0.20 | Excellent for RAG-heavy librarian tasks. |
| 3 | DeepSeek V3.2 | Previous-gen, proven reliability | $0.10/$0.20 | Fallback. Mature ecosystem and hosting options. |
| 4 | Mistral Small 4 | Apache 2.0, fast code summarization | $0.08/$0.16 | Best for self-hosted setups on modest hardware (8B class). |
| 5 | Qwen 3.5-72B | Balanced quality and cost | $0.12/$0.25 | Good default for designer tasks. |

## Simple Tier (low variant)

Used by: Fixer, Council alpha.

Criteria: maximum speed, minimum cost. Single-file edits that should be near-instant.

### Closed Source

| Rank | Model | Key Strengths | Approx Pricing (per M tokens) | Notes |
|------|-------|---------------|-------------------------------|-------|
| 1 | Gemini 3 Flash | Cross-listed from standard tier | $0.50/$3 | Fast, cheap, and good quality. |
| 2 | GPT-5.4 Nano | 0.2s TTFT, $0.05/M in | $0.05/$0.20 | Best latency. Perfect for Fixer's iteration loop. |
| 3 | Gemini 3.1 Flash-Lite | Extremely cheap | $0.15/$0.60 | Runs on every council invocation. |
| 4 | Mistral Small 3.2 | Fastest Mistral variant | $0.10/$0.20 | European option. Fast inference. |
| 5 | GPT-5 Nano | Minimal GPT-5 model | $0.03/$0.12 | Cheapest closed model. Council alpha only (diversity of thought). |

### Open Weight

| Rank | Model | Key Strengths | Approx Pricing (per M tokens) | Notes |
|------|-------|---------------|-------------------------------|-------|
| 1 | DeepSeek V4 Flash | Cross-listed from standard tier | $0.14/$0.28 | Fast, good, cheap. |
| 2 | Phi-4 Mini | Microsoft's 3.8B | $0.03/$0.06 | Best model under 5B parameters. Self-host on CPU. |
| 3 | Gemma 4 E4B | Google's 4B MoE | $0.03/$0.06 | Strongest in its weight class. |
| 4 | GPT-OSS-20B | Community fine-tune, 20B, MIT license | $0.05/$0.10 | Distilled from GPT-4 lineage. |
| 5 | Qwen 2.5 Coder 7B | Proven code specialist | $0.04/$0.08 | Mature ecosystem. Well-tested for simple code changes. |

## Specialist Picks

### Observer (Vision/Multimodal)

- **Closed 1**: Gemini 3.1 Pro - 2M context, best multimodal understanding
- **Closed 2**: GPT-5.5 Vision - good vision, better code from screenshots
- **Closed 3**: Claude Opus 4.7 - best detailed visual reasoning
- **Open 1**: Qwen 3.5-Omni - strongest open-weight multimodal
- **Open 2**: Kimi-VL2.6 - good UI understanding
- **Open 3**: Llama 4 Maverick (multimodal variant) - Meta's best open vision model

### Designer (UI-to-Code / Visual Design)

- **Closed 1**: Claude Opus 4.7 - best UI-to-code conversion
- **Closed 2**: GPT-5.5 - fastest UI iteration
- **Closed 3**: Gemini 3.1 Pro - design system consistency
- **Open 1**: UI2Code^N - specialized for visual design
- **Open 2**: Qwen 3.5-Omni - UI component creation
- **Open 3**: DeepSeek V4 Pro - generalist strong at UI code

### Librarian (Search / RAG)

- **Closed 1**: Gemini 3.1 Pro - 2M context for massive codebase search
- **Closed 2**: Gemini 3 Flash - 97% quality at 1/4 cost
- **Closed 3**: Claude Haiku 4.5 - best tool use for search chains
- **Open 1**: DeepSeek V4 Flash - cheap search-heavy workloads
- **Open 2**: MiniMax M2.7 - best retrieval quality among open-weight models
- **Open 3**: Qwen 3.5-72B - balance of search comprehension and cost

## Quick-Reference Table

| Tier | Closed #1 | Closed #2 | Closed #3 | Open #1 | Open #2 | Open #3 |
|------|-----------|-----------|-----------|---------|---------|---------|
| Reasoning | Claude Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro | DeepSeek V4 Pro | Kimi K2.6 | Qwen 3.7 Max |
| Complex | Claude Sonnet 4.6 | GPT-5.4 | Gemini 3 Flash | DeepSeek V4 Pro | Kimi K2.6 | MiMo-V2.5-Pro |
| Standard | Gemini 3 Flash | Claude Haiku 4.5 | GPT-5.4 Mini | DeepSeek V4 Flash | MiniMax M2.7 | DeepSeek V3.2 |
| Simple | Gemini 3 Flash | GPT-5.4 Nano | Gemini 3.1 Flash-Lite | DeepSeek V4 Flash | Phi-4 Mini | Gemma 4 E4B |

## Caveats

- Pricing is approximate and provider-dependent. Self-hosted open-weight costs exclude hardware.
- Rankings reflect May 2026. The model landscape shifts monthly.
- Claude Mythos is speculative. Include only if on the cutting edge.
- Open-weight rankings assume self-hosting. Cloud-hosted pricing varies.
- For council diversity, avoid using the same model for multiple councillors.
- Specialized models (Phi-4 Mini, Qwen 2.5 Coder 7B) may outperform generalists on their specific task.

Navigation Menu

[INFORMATIONAL] Model ranking/recommendations - May 2026 #498

Description

Model Recommendations for opencode Agent Roles - May 2026

Purpose

Methodology

Extra context

Reasoning Tier (max variant)

Closed Source

Open Weight

Complex Tier (high variant)

Closed Source

Open Weight

Standard Tier (medium/low variant)

Closed Source

Open Weight

Simple Tier (low variant)

Closed Source

Open Weight

Specialist Picks

Observer (Vision/Multimodal)

Designer (UI-to-Code / Visual Design)

Librarian (Search / RAG)

Quick-Reference Table

Caveats

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Rank	Model	Key Strengths	Approx Pricing (per M tokens)	Notes
1	Claude Opus 4.7	Best multi-file coding, Terminal-Bench #1, strongest agentic tool use	$15/$75	Anthropic flagship for complex software engineering. 200K context.
2	GPT-5.5	Highest AA Index (60), Terminal-Bench #2, fastest reasoning-class inference	$10/$40	Best throughput of any frontier model. Excels at agentic chaining.
3	Gemini 3.1 Pro	GPQA 94.3%, best reasoning per dollar, 2M context	$2/$10	Best value in reasoning tier. Unmatched context window.
4	Claude Mythos	Experimental Anthropic variant, reportedly SWE-bench 83.1%	~$15/$75	Pre-release. For early adopters running Anthropic's latest reasoning architecture.
5	GPT-5.5 Pro	Extended reasoning variant of GPT-5.5, highest SWE-bench (84.2%)	$15/$60	Best raw SWE-bench score. Use Oracle when correctness must not fail.

Rank	Model	Key Strengths	Approx Pricing (per M tokens)	Notes
1	DeepSeek V4 Pro	MIT license, 80.6% SWE-bench	$0.55/$0.87	Open-weight champion. Self-hostable. 128K context.
2	Kimi K2.6	79.4% SWE-bench, Chinese + English, 128K context	$0.50/$0.80	Moonshot AI's latest. Strong multilingual code reasoning.
3	Qwen 3.7 Max	GPQA 92.1%, multilingual	$0.60/$1.00	Best for multilingual reasoning. 131K context.
4	GLM-5.1	Zhipu AI flagship, competitive with Kimi on code	$0.45/$0.75	Solid all-rounder on Chinese cloud providers.
5	Llama 4 Maverick	405B-class MoE, uncensored variant	$0.50/$0.80	Apache 2.0. Best for self-hosting at scale.

Rank	Model	Key Strengths	Approx Pricing (per M tokens)	Notes
1	Claude Sonnet 4.6	Best quality-to-latency ratio in Claude lineup	$3/$15	Nearly Opus-quality at 1/5 the cost.
2	GPT-5.4	Fast, 128K context, strong Terminal-Bench scores	$2.50/$10	Less capable than GPT-5.5 but significantly cheaper.
3	Gemini 3 Flash	97% of Gemini 3.1 Pro quality at 1/4 the cost	$0.50/$3	Near reasoning-tier quality at standard-tier prices.
4	Grok 4.1	xAI's latest, 128K context, strong math/code reasoning	$3/$12	Less proven on SWE-bench. Competitive on math and science.
5	GPT-5.4 Pro	Higher-correctness variant of GPT-5.4	$5/$15	Middle ground between Sonnet 4.6 and Opus 4.7 pricing.

Rank	Model	Key Strengths	Approx Pricing (per M tokens)	Notes
1	DeepSeek V4 Pro	Cross-listed from reasoning tier	$0.55/$0.87	Versatile enough for both tiers. Self-host alongside reasoning workload.
2	Kimi K2.6	Cross-listed. Excellent code at open-weight pricing	$0.50/$0.80	Good second source for council diversity.
3	MiMo-V2.5-Pro	01.AI's latest, 128K context	$0.40/$0.70	Budget option for complex code generation.
4	Mistral Large 3	European languages, good tool use	$0.40/$0.70	Apache 2.0. Best for EU compliance.
5	Qwen 3.5-397B	Alibaba MoE model	$0.35/$0.60	Slightly behind Kimi and DS V4 but ahead of older open models.

Rank	Model	Key Strengths	Approx Pricing (per M tokens)	Notes
1	Gemini 3 Flash	0.5s TTFT, 97% quality retention from Pro, 1M context	$0.50/$3	Best standard-tier option. Excellent for librarians and explorers.
2	Claude Haiku 4.5	Fastest Claude variant, 200K context	$0.25/$1.25	Best tool-use in standard tier. Slightly slower than Gemini Flash.
3	GPT-5.4 Mini	Fast, 128K context, good for structured data extraction	$0.15/$0.60	Best for high-volume scraping.
4	GPT-5.4 Nano	Smallest GPT, extremely fast	$0.05/$0.20	Lowest latency of any closed model. Ideal for rapid-fire explorer queries.
5	Gemini 3.1 Flash-Lite	Budget Flash variant	$0.15/$0.60	Best value for Observer and basic designer tasks.

Rank	Model	Key Strengths	Approx Pricing (per M tokens)	Notes
1	DeepSeek V4 Flash	79% SWE-bench at 1/10 the cost, 128K context	$0.14/$0.28	Budget champion. Self-hostable. Good enough for most standard tasks.
2	MiniMax M2.7	Hugging Face leader in several categories	$0.10/$0.20	Excellent for RAG-heavy librarian tasks.
3	DeepSeek V3.2	Previous-gen, proven reliability	$0.10/$0.20	Fallback. Mature ecosystem and hosting options.
4	Mistral Small 4	Apache 2.0, fast code summarization	$0.08/$0.16	Best for self-hosted setups on modest hardware (8B class).
5	Qwen 3.5-72B	Balanced quality and cost	$0.12/$0.25	Good default for designer tasks.

Rank	Model	Key Strengths	Approx Pricing (per M tokens)	Notes
1	Gemini 3 Flash	Cross-listed from standard tier	$0.50/$3	Fast, cheap, and good quality.
2	GPT-5.4 Nano	0.2s TTFT, $0.05/M in	$0.05/$0.20	Best latency. Perfect for Fixer's iteration loop.
3	Gemini 3.1 Flash-Lite	Extremely cheap	$0.15/$0.60	Runs on every council invocation.
4	Mistral Small 3.2	Fastest Mistral variant	$0.10/$0.20	European option. Fast inference.
5	GPT-5 Nano	Minimal GPT-5 model	$0.03/$0.12	Cheapest closed model. Council alpha only (diversity of thought).

Tier	Closed #1	Closed #2	Closed #3	Open #1	Open #2	Open #3
Reasoning	Claude Opus 4.7	GPT-5.5	Gemini 3.1 Pro	DeepSeek V4 Pro	Kimi K2.6	Qwen 3.7 Max
Complex	Claude Sonnet 4.6	GPT-5.4	Gemini 3 Flash	DeepSeek V4 Pro	Kimi K2.6	MiMo-V2.5-Pro
Standard	Gemini 3 Flash	Claude Haiku 4.5	GPT-5.4 Mini	DeepSeek V4 Flash	MiniMax M2.7	DeepSeek V3.2
Simple	Gemini 3 Flash	GPT-5.4 Nano	Gemini 3.1 Flash-Lite	DeepSeek V4 Flash	Phi-4 Mini	Gemma 4 E4B

[INFORMATIONAL] Model ranking/recommendations - May 2026 #498

Description

Description

Model Recommendations for opencode Agent Roles - May 2026

Purpose

Methodology

Extra context

Reasoning Tier (max variant)

Closed Source

Open Weight

Complex Tier (high variant)

Closed Source

Open Weight

Standard Tier (medium/low variant)

Closed Source

Open Weight

Simple Tier (low variant)

Closed Source

Open Weight

Specialist Picks

Observer (Vision/Multimodal)

Designer (UI-to-Code / Visual Design)

Librarian (Search / RAG)

Quick-Reference Table

Caveats

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions