Octopus Prime
AIs Read Here Start Free Trial
Model guide

Model guide

The model that's best at one job often isn't the best at another, and for the same coding task the most expensive model here costs 100 times as much as the cheapest. Pick a use case to see how each model scores against what one task costs. Higher is better, further left is cheaper.

Prices and benchmarks checked

Coding

One task: one bug fix in a real codebase: about 100,000 tokens read (files, history, tool output) and 10,000 written. 21 models with API prices, 3 local.

  • Best-value line: nothing else is both cheaper and higher scoring
  • Other models with API prices
  • Local: free to run, needs your own hardware
Coding: Octopus Score against cost per task 21 models with API prices and 3 local models. The table below the chart lists every value. Coding: Octopus Score against cost per task 21 models with API prices and 3 local models. The table below the chart lists every value.

Hover over or tap a dot to see the model, its score and its cost.

Show as a table
Model Octopus Score Cost per task Data
Claude Opus 5.5 Anthropic 99 $0.60 All benchmarks
Claude Sonnet 5.5 Anthropic 93 $0.30 All benchmarks
GPT-6 Astra OpenAI 93 $1.50 Partial (4 of 5 benchmarks)
GPT-6.1 Sol OpenAI 89 $0.30 Partial (3 of 5 benchmarks)
Claude Fable 5.1 Anthropic 88 $1.50 All benchmarks
Gemini 4 Argon Google 85 $0.30 Partial (4 of 5 benchmarks)
Muse Wonder Meta 81 $0.60 Partial (3 of 5 benchmarks)
DeepSeek V4.1 Flash DeepSeek 76 $0.042 Partial (2 of 5 benchmarks)
Kimi K3 Moonshot AI 67 $0.45 Partial (3 of 5 benchmarks)
GLM-5.3 Z.ai 66 $0.184 All benchmarks
Claude Haiku 5.5 Anthropic 65 $0.015 Partial (3 of 5 benchmarks)
Qwen3.8 27B Qwen (Alibaba) 64 Local Partial (2 of 5 benchmarks)
DeepSeek V4 Pro DeepSeek 64 $0.172 Partial (2 of 5 benchmarks)
GPT-5.6 Terra OpenAI 62 $0.32 Partial (3 of 5 benchmarks)
Grok 4.7 xAI 60 $0.26 Partial (4 of 5 benchmarks)
GLM-5.3 Flash Z.ai 58 $0.02 All benchmarks
Muse Spark 1.3 Meta 57 $0.168 All benchmarks
Qwen3.8 Max Qwen (Alibaba) 56 $0.26 Partial (4 of 5 benchmarks)
Gemini 3.8 Flash Google 53 $0.113 All benchmarks
Gemma 4 31B Google 49 Local Partial (2 of 5 benchmarks)
Gemma 4 26B A4B Google 46 Local Partial (2 of 5 benchmarks)
Mistral Medium 3.5 Mistral 43 $0.225 Partial (2 of 5 benchmarks)
Mistral Large 3 Mistral 41 $0.065 Partial (2 of 5 benchmarks)
GPT-6 Luna OpenAI 33 $0.015 Partial (2 of 5 benchmarks)

Agents & tool use

One task: one multi-step agent run with tools: about 50,000 tokens read and 3,000 written. 15 models with API prices.

  • Best-value line: nothing else is both cheaper and higher scoring
  • Other models with API prices
  • Local: free to run, needs your own hardware
Agents & tool use: Octopus Score against cost per task 15 models with API prices. The table below the chart lists every value. Agents & tool use: Octopus Score against cost per task 15 models with API prices. The table below the chart lists every value.

Hover over or tap a dot to see the model, its score and its cost.

Show as a table
Model Octopus Score Cost per task Data
Gemini 4 Argon Google 97 $0.13 Partial (2 of 3 benchmarks)
Claude Opus 5.5 Anthropic 95 $0.26 All benchmarks
Claude Sonnet 5.5 Anthropic 94 $0.13 All benchmarks
Claude Fable 5.1 Anthropic 87 $0.65 All benchmarks
GPT-6 Astra OpenAI 83 $0.65 Partial (2 of 3 benchmarks)
GPT-6.1 Sol OpenAI 80 $0.13 Partial (2 of 3 benchmarks)
Claude Haiku 5.5 Anthropic 74 $0.0065 Partial (2 of 3 benchmarks)
GLM-5.3 Z.ai 69 $0.083 All benchmarks
Grok 4.7 xAI 67 $0.118 All benchmarks
Qwen3.8 Max Qwen (Alibaba) 63 $0.118 Partial (2 of 3 benchmarks)
GLM-5.3 Flash Z.ai 61 $0.009 All benchmarks
Gemini 3.8 Flash Google 61 $0.049 All benchmarks
GPT-5.6 Terra OpenAI 60 $0.136 All benchmarks
Muse Spark 1.3 Meta 57 $0.075 All benchmarks
GPT-6 Luna OpenAI 42 $0.0065 Partial (2 of 3 benchmarks)

Writing

One task: one 1,500-word draft from a brief: about 3,000 tokens read and 2,500 written. 20 models with API prices, 6 local.

  • Best-value line: nothing else is both cheaper and higher scoring
  • Other models with API prices
  • Local: free to run, needs your own hardware
Writing: Octopus Score against cost per task 20 models with API prices and 6 local models. The table below the chart lists every value. Writing: Octopus Score against cost per task 20 models with API prices and 6 local models. The table below the chart lists every value.

Hover over or tap a dot to see the model, its score and its cost.

Show as a table
Model Octopus Score Cost per task Data
Gemini 4 Argon Google 99 $0.031 Partial (2 of 3 benchmarks)
Claude Fable 5.1 Anthropic 93 $0.155 All benchmarks
Muse Wonder Meta 92 $0.062 Partial (2 of 3 benchmarks)
GPT-6 Astra OpenAI 88 $0.155 All benchmarks
Claude Opus 5.5 Anthropic 88 $0.062 All benchmarks
Claude Sonnet 5.5 Anthropic 86 $0.031 Partial (2 of 3 benchmarks)
Qwen3.8 Max Qwen (Alibaba) 85 $0.021 Partial (2 of 3 benchmarks)
GPT-6.1 Sol OpenAI 84 $0.031 Partial (2 of 3 benchmarks)
Kimi K3 Moonshot AI 82 $0.046 All benchmarks
GLM-5.3 Z.ai 80 $0.015 All benchmarks
GLM-5.3 Flash Z.ai 79 $0.0017 Partial (2 of 3 benchmarks)
Qwen3.7 Plus Qwen (Alibaba) 76 $0.0052 Partial (2 of 3 benchmarks)
Mistral Medium 3.5 Mistral 68 $0.023 Partial (2 of 3 benchmarks)
Muse Spark 1.3 Meta 68 $0.014 All benchmarks
Gemini 3.8 Flash Google 64 $0.012 All benchmarks
Grok 4.7 xAI 64 $0.021 Partial (2 of 3 benchmarks)
DeepSeek V4.1 Flash DeepSeek 54 $0.0039 All benchmarks
DeepSeek V4 Pro DeepSeek 54 $0.014 All benchmarks
GPT-5.6 Terra OpenAI 51 $0.036 Partial (2 of 3 benchmarks)
Gemma 4 31B Google 49 Local All benchmarks
Gemma 4 26B A4B Google 46 Local All benchmarks
Mistral Large 3 Mistral 42 $0.0053 All benchmarks
Qwen3.8 27B Qwen (Alibaba) 36 Local Partial (2 of 3 benchmarks)
Mistral Small 3.2 24B Mistral 33 Local All benchmarks
gpt-oss-120b OpenAI 28 Local All benchmarks
gpt-oss-20b OpenAI 23 Local All benchmarks

Research & long documents

One task: one question over a long report: about 150,000 tokens read and 3,000 written. 21 models with API prices, 7 local.

  • Best-value line: nothing else is both cheaper and higher scoring
  • Other models with API prices
  • Local: free to run, needs your own hardware
Research & long documents: Octopus Score against cost per task 21 models with API prices and 7 local models. The table below the chart lists every value. Research & long documents: Octopus Score against cost per task 21 models with API prices and 7 local models. The table below the chart lists every value.

Hover over or tap a dot to see the model, its score and its cost.

Show as a table
Model Octopus Score Cost per task Data
Gemini 4 Argon Google 100 $0.33 Partial (3 of 4 benchmarks)
Claude Opus 5.5 Anthropic 94 $0.66 Partial (3 of 4 benchmarks)
Gemini 3.8 Flash Google 94 $0.124 Partial (3 of 4 benchmarks)
GPT-6 Astra OpenAI 93 $1.65 Partial (3 of 4 benchmarks)
GPT-6.1 Sol OpenAI 93 $0.33 Partial (3 of 4 benchmarks)
Claude Fable 5.1 Anthropic 92 $1.65 Partial (2 of 4 benchmarks)
Qwen3.8 Max Qwen (Alibaba) 87 $0.318 All benchmarks
GLM-5.3 Flash Z.ai 85 $0.024 Partial (3 of 4 benchmarks)
Claude Sonnet 5.5 Anthropic 84 $0.33 Partial (3 of 4 benchmarks)
GPT-5.6 Terra OpenAI 84 $0.336 Partial (3 of 4 benchmarks)
Qwen3.7 Plus Qwen (Alibaba) 84 $0.065 Partial (2 of 4 benchmarks)
Qwen3.8 27B Qwen (Alibaba) 83 Local Partial (2 of 4 benchmarks)
DeepSeek V4 Pro DeepSeek 83 $0.21 All benchmarks
GLM-5.3 Z.ai 82 $0.223 All benchmarks
Grok 4.7 xAI 82 $0.318 Partial (3 of 4 benchmarks)
Kimi K3 Moonshot AI 81 $0.495 All benchmarks
DeepSeek V4.1 Flash DeepSeek 79 $0.049 Partial (3 of 4 benchmarks)
Gemini 3.5 Flash-Lite Google 77 $0.052 Partial (2 of 4 benchmarks)
GPT-6 Luna OpenAI 76 $0.017 Partial (2 of 4 benchmarks)
Gemma 4 26B A4B Google 74 Local Partial (2 of 4 benchmarks)
Grok 4.3 xAI 71 $0.195 Partial (2 of 4 benchmarks)
Claude Haiku 5.5 Anthropic 66 $0.017 Partial (2 of 4 benchmarks)
Gemma 4 31B Google 62 Local Partial (3 of 4 benchmarks)
gpt-oss-120b OpenAI 60 Local Partial (2 of 4 benchmarks)
Mistral Small 3.2 24B Mistral 51 Local Partial (2 of 4 benchmarks)
gpt-oss-20b OpenAI 51 Local Partial (2 of 4 benchmarks)
Llama 4 Scout Meta 49 Local Partial (2 of 4 benchmarks)
Mistral Medium 3.5 Mistral 48 $0.247 Partial (2 of 4 benchmarks)

Customer chat

One task: one customer conversation: about 8,000 tokens read (instructions, policy, history) and 1,000 written. 20 models with API prices, 6 local.

  • Best-value line: nothing else is both cheaper and higher scoring
  • Other models with API prices
  • Local: free to run, needs your own hardware
Customer chat: Octopus Score against cost per task 20 models with API prices and 6 local models. The table below the chart lists every value. Customer chat: Octopus Score against cost per task 20 models with API prices and 6 local models. The table below the chart lists every value.

Hover over or tap a dot to see the model, its score and its cost.

Show as a table
Model Octopus Score Cost per task Data
Gemini 4 Argon Google 99 $0.026 All benchmarks
Claude Opus 5.5 Anthropic 92 $0.052 Partial (3 of 4 benchmarks)
Muse Wonder Meta 91 $0.052 Partial (2 of 4 benchmarks)
Gemini 3.8 Flash Google 89 $0.0097 Partial (3 of 4 benchmarks)
GPT-6 Astra OpenAI 88 $0.13 All benchmarks
Claude Fable 5.1 Anthropic 88 $0.13 Partial (3 of 4 benchmarks)
GPT-6.1 Sol OpenAI 86 $0.026 Partial (3 of 4 benchmarks)
Muse Spark 1.3 Meta 85 $0.014 Partial (2 of 4 benchmarks)
Kimi K3 Moonshot AI 82 $0.039 Partial (3 of 4 benchmarks)
DeepSeek V4 Pro DeepSeek 81 $0.015 All benchmarks
GLM-5.3 Flash Z.ai 81 $0.0017 Partial (2 of 4 benchmarks)
Gemma 4 26B A4B Google 80 Local Partial (3 of 4 benchmarks)
Qwen3.8 Max Qwen (Alibaba) 79 $0.022 Partial (3 of 4 benchmarks)
Claude Sonnet 5.5 Anthropic 77 $0.026 Partial (3 of 4 benchmarks)
Qwen3.7 Plus Qwen (Alibaba) 76 $0.0048 Partial (2 of 4 benchmarks)
GLM-5.3 Z.ai 75 $0.016 Partial (3 of 4 benchmarks)
Grok 4.7 xAI 75 $0.022 Partial (2 of 4 benchmarks)
DeepSeek V4.1 Flash DeepSeek 74 $0.0036 Partial (3 of 4 benchmarks)
GPT-5.6 Terra OpenAI 74 $0.028 Partial (3 of 4 benchmarks)
Qwen3.8 27B Qwen (Alibaba) 70 Local Partial (2 of 4 benchmarks)
Gemma 4 31B Google 69 Local All benchmarks
Mistral Medium 3.5 Mistral 69 $0.019 Partial (2 of 4 benchmarks)
Mistral Large 3 Mistral 65 $0.0055 Partial (2 of 4 benchmarks)
gpt-oss-120b OpenAI 56 Local Partial (3 of 4 benchmarks)
Mistral Small 3.2 24B Mistral 51 Local Partial (2 of 4 benchmarks)
gpt-oss-20b OpenAI 38 Local Partial (2 of 4 benchmarks)

Images (vision)

One task: one image plus a question: about 2,000 tokens read and 500 written. 17 models with API prices, 5 local.

  • Best-value line: nothing else is both cheaper and higher scoring
  • Other models with API prices
  • Local: free to run, needs your own hardware
Images (vision): Octopus Score against cost per task 17 models with API prices and 5 local models. The table below the chart lists every value. Images (vision): Octopus Score against cost per task 17 models with API prices and 5 local models. The table below the chart lists every value.

Hover over or tap a dot to see the model, its score and its cost.

Show as a table
Model Octopus Score Cost per task Data
Qwen3.8 Max Qwen (Alibaba) 100 $0.007 All benchmarks
GPT-6.1 Sol OpenAI 98 $0.009 All benchmarks
Claude Opus 5.5 Anthropic 98 $0.018 All benchmarks
Gemini 3.8 Flash Google 97 $0.0034 All benchmarks
Muse Spark 1.3 Meta 97 $0.0046 All benchmarks
Claude Fable 5.1 Anthropic 96 $0.045 All benchmarks
GPT-6 Astra OpenAI 96 $0.045 All benchmarks
GLM-5.3 Flash Z.ai 92 $0.00055 All benchmarks
GPT-5.6 Terra OpenAI 91 $0.01 All benchmarks
Gemini 3.5 Flash-Lite Google 91 $0.0019 All benchmarks
Claude Sonnet 5.5 Anthropic 90 $0.009 All benchmarks
Qwen3.7 Plus Qwen (Alibaba) 90 $0.0016 All benchmarks
Gemma 4 31B Google 89 Local All benchmarks
DeepSeek V4.1 Flash DeepSeek 89 $0.0012 All benchmarks
Grok 4.7 xAI 86 $0.007 All benchmarks
Qwen3.8 27B Qwen (Alibaba) 84 Local All benchmarks
Grok 4.3 xAI 84 $0.0037 All benchmarks
Gemma 4 26B A4B Google 83 Local All benchmarks
Mistral Medium 3.5 Mistral 72 $0.0067 All benchmarks
Mistral Large 3 Mistral 72 $0.0018 All benchmarks
Mistral Small 3.2 24B Mistral 57 Local All benchmarks
Llama 4 Scout Meta 54 Local All benchmarks

Cheap high-volume

One task: one small job such as tagging or extracting from a short message: about 1,000 tokens read and 100 written. 20 models with API prices, 6 local.

  • Best-value line: nothing else is both cheaper and higher scoring
  • Other models with API prices
  • Local: free to run, needs your own hardware
Cheap high-volume: Octopus Score against cost per task 20 models with API prices and 6 local models. The table below the chart lists every value. Cheap high-volume: Octopus Score against cost per task 20 models with API prices and 6 local models. The table below the chart lists every value.

Hover over or tap a dot to see the model, its score and its cost.

Show as a table
Model Octopus Score Cost per task Data
Gemini 4 Argon Google 100 $0.003 All benchmarks
Claude Opus 5.5 Anthropic 96 $0.006 All benchmarks
Muse Wonder Meta 94 $0.006 Partial (2 of 3 benchmarks)
Gemini 3.8 Flash Google 93 $0.0011 All benchmarks
Claude Fable 5.1 Anthropic 93 $0.015 Partial (2 of 3 benchmarks)
Kimi K3 Moonshot AI 92 $0.0045 All benchmarks
Claude Sonnet 5.5 Anthropic 91 $0.003 All benchmarks
GPT-6.1 Sol OpenAI 91 $0.003 All benchmarks
Muse Spark 1.3 Meta 90 $0.0017 Partial (2 of 3 benchmarks)
GPT-6 Astra OpenAI 90 $0.015 All benchmarks
Qwen3.8 Max Qwen (Alibaba) 90 $0.0026 All benchmarks
GLM-5.3 Z.ai 89 $0.0018 All benchmarks
DeepSeek V4.1 Flash DeepSeek 88 $0.00042 All benchmarks
GLM-5.3 Flash Z.ai 88 $0.0002 All benchmarks
GPT-5.6 Terra OpenAI 87 $0.0032 All benchmarks
DeepSeek V4 Pro DeepSeek 86 $0.0017 All benchmarks
Grok 4.7 xAI 85 $0.0026 Partial (2 of 3 benchmarks)
Qwen3.7 Plus Qwen (Alibaba) 83 $0.00056 All benchmarks
Gemma 4 31B Google 79 Local All benchmarks
Gemma 4 26B A4B Google 76 Local All benchmarks
Qwen3.8 27B Qwen (Alibaba) 74 Local Partial (2 of 3 benchmarks)
Mistral Medium 3.5 Mistral 72 $0.0022 Partial (2 of 3 benchmarks)
Mistral Large 3 Mistral 68 $0.00065 Partial (2 of 3 benchmarks)
gpt-oss-120b OpenAI 59 Local All benchmarks
Mistral Small 3.2 24B Mistral 52 Local All benchmarks
gpt-oss-20b OpenAI 49 Local All benchmarks
From the data

Fun facts

  • 1/13 the price

    For images, GLM-5.3 Flash scores 92, within 10 points of the top score (Qwen3.8 Max, 100), at $0.00055 a task instead of $0.007.

  • 100×

    The price gap between the cheapest and the most expensive model for the same coding task: $0.015 against $1.50.

  • $0.20

    A thousand small jobs (tagging, sorting, extracting) with GLM-5.3 Flash, which scores 88 for high-volume work. The same thousand with the top scorer, Gemini 4 Argon (100): $3.00.

  • 3 leaders

    No single model wins everything. Gemini 4 Argon tops 5 of the 7 use cases (agents & tool use, writing, research & long documents, customer chat and cheap high-volume); Claude Opus 5.5 tops coding; Qwen3.8 Max tops images.

  • Free

    Qwen3.8 27B runs on a computer with roughly 24 GB of memory and scores 83 for research & long documents, ahead of Kimi K3 (81) at $0.495 a task.

  • 5×

    What a token the model writes costs, compared with a token it reads, for the typical (median) model in this guide. So asking for short answers when you don't need long ones saves more per token than trimming what you send.

OpenClaw

How to spend less with OpenClaw

Checked against the OpenClaw docs (2026.10.5). Commands go in any chat with your agent; config lives in your Gateway's config file.

  1. Give each agent the model its job needs

    An agent that files emails or posts a daily summary doesn't need the top model. Set a cheaper model per agent with agents.entries.<agent>.model, and keep the expensive model for the agent that does the hard thinking. Per-agent settings

  2. Switch one chat's model with /model

    /model <provider/model> -s changes the model for the current conversation only; /model default -s switches back. Use the top model for the hard question, then drop back down. Slash commands

  3. Start a new session for a new topic

    Every message resends the conversation so far. /new starts a fresh session so you stop paying for old history; /compact shrinks a long session you want to keep. Token use and costs

  4. Plan with a top model, do the work with a cheaper one

    Ask the top model for the plan, then switch the chat with /model (or hand the steps to a sub-agent) and let a cheaper model carry it out. Plans are short; doing the work is where the tokens go.

  5. Use cheaper models for sub-agents, heartbeats and automations

    Sub-agents follow agents.defaults.subagents.model, heartbeats follow heartbeat.model (and heartbeat.lightContext skips your instruction files on those runs), and each automation job takes its own --model. Sub-agents · Heartbeat · Automations

  6. Keep instruction files lean

    AGENTS.md, SOUL.md, USER.md and MEMORY.md go into every run, so every extra paragraph is paid for again on every message. /context list shows what each file costs; /usage cost shows what you've spent. Context

Methodology

How the Octopus Score works

The Octopus Score is our own composite of public benchmark results, 0 to 100 per use case. For each benchmark, the best model in this guide gets 100. Benchmarks scored in percent count as a share of the best result (60% against a best of 80% gives 75). Arena scores, which come from people voting between two answers, become the chance of beating the best model head to head, doubled so a tie is 100: 100 points behind gives 72, 400 points behind gives 18.

Each use case blends its benchmarks with the weights below. When a model is missing a benchmark, the other weights are scaled up to fill the gap and the model is marked "partial data". A model appears in a use case only with results from at least 2 of its benchmarks (images has only one) that together carry 40% of its weight. Where a source lists several thinking settings of one model, we use the best-scoring one.

Cost per task is the provider's list price for input and output tokens (standard rate, no batch or caching discounts) times a fixed token mix per use case. Real costs vary: models that think out loud write more tokens than the mix assumes, and caching makes repeated prompts cheaper. Local models have no API price; they're free to run on your own hardware and are plotted on their own.

Benchmarks measure what they measure. Use the score to narrow the list, then try two or three models on your own work.

Benchmarks and weights per use case, and the token mix for one task
Use case Benchmarks (weight) One task (tokens in / out)
Coding Terminal-Bench 4.0 (25%), CursorBench (20%), FrontierSWE (15%), LMArena WebDev (20%), LMArena Coding (20%) 100,000 / 10,000
Agents & tool use APEX-Agents (45%), Terminal-Bench 4.0 (30%), CursorBench (25%) 50,000 / 3,000
Writing LMArena Creative Writing (40%), EQ-Bench Creative Writing (35%), LMArena Instruction Following (25%) 3,000 / 2,500
Research & long documents LMArena Longer Query (30%), GPQA Diamond (25%), Context Arena MRCR v2 (128k) (25%), SimpleQA Verified (20%) 150,000 / 3,000
Customer chat LMArena Multi-Turn (35%), LMArena Instruction Following (25%), SimpleQA Verified (20%), Vectara factual consistency (20%) 8,000 / 1,000
Images (vision) LMArena Vision (100%) 2,000 / 500
Cheap high-volume LMArena Text (35%), LMArena Instruction Following (35%), GPQA Diamond (30%) 1,000 / 100

Sources

  • LMArena leaderboard dataset: LMArena Text, LMArena Coding, LMArena WebDev, LMArena Vision, LMArena Creative Writing, LMArena Longer Query, LMArena Multi-Turn, LMArena Instruction Following. Arena scores from the LMArena leaderboard dataset (style control on), CC BY 4.0.
  • Epoch AI Benchmarking Hub: GPQA Diamond, SimpleQA Verified, APEX-Agents, CursorBench, FrontierSWE. Epoch AI, ‘Capabilities & benchmarking’, epoch.ai/benchmarks, CC BY 4.0: GPQA Diamond and SimpleQA Verified run by Epoch AI; APEX-Agents (Mercor), CursorBench (Cursor) and FrontierSWE results as compiled by Epoch AI.
  • Terminal-Bench 4.0: Terminal-Bench 4.0. Terminal-Bench 4.0 leaderboard, tbench.ai (best published agent run per model).
  • Context Arena: Context Arena MRCR v2 (128k). Long-context retrieval: Context Arena, MRCR v2 with 8 needles, score up to 128k tokens.
  • Vectara Hallucination Leaderboard: Vectara factual consistency. Factual consistency (100% minus hallucination rate) from the Vectara Hallucination Leaderboard, Apache 2.0.
  • EQ-Bench Creative Writing v3: EQ-Bench Creative Writing. Creative writing: EQ-Bench Creative Writing v3 (eqbench.com), Elo score.
  • Prices: each provider's own pricing page, list prices in US dollars per million tokens. 24 models with API prices, 7 local models.

Not used: Artificial Analysis (its terms forbid republishing its data or using it for model rankings); SWE-bench (its leaderboard data is licensed for non-commercial use only); Aider polyglot leaderboard (no results for 2026 models); Berkeley Function Calling Leaderboard (no results for 2026 models); τ²-bench and τ-knowledge (Sierra) (current results for only 2 of the models in this guide); MMMU leaderboard (no results for the current models in this guide).

Prices used

List prices per million tokens, checked October 21, 2026
Model Input Output Note
Claude Fable 5.1 Anthropic $10 $50
Claude Opus 5.5 Anthropic $4 $20
Claude Sonnet 5.5 Anthropic $2 $10
Claude Haiku 5.5 Anthropic $0.1 $0.5 Prompts up to 100,000 tokens ($0.50 / $2.50 above)
GPT-6 Astra OpenAI $10 $50 Prompts up to 272,000 tokens
GPT-6.1 Sol OpenAI $2 $10 Prompts up to 272,000 tokens
GPT-6 Luna OpenAI $0.1 $0.5 Prompts up to 272,000 tokens
GPT-5.6 Terra OpenAI $2 $12 Prompts up to 272,000 tokens
Gemini 4 Argon Google $2 $10 Launch price until November 30, 2026, then $4 / $20
Gemini 3.8 Flash Google $0.75 $3.75 Until December 31, 2026, then $1.50 / $7.50
Gemini 3.5 Flash-Lite Google $0.3 $2.5
Grok 4.7 xAI $2 $6 Prompts under 200,000 tokens
Grok 4.3 xAI $1.25 $2.5 Prompts under 200,000 tokens
DeepSeek V4 Pro DeepSeek $1.32 $3.96 Peak-hour price; half price off-peak
DeepSeek V4.1 Flash DeepSeek $0.3 $1.2 Peak-hour price; half price off-peak
Mistral Medium 3.5 Mistral $1.5 $7.5
Mistral Large 3 Mistral $0.5 $1.5
Muse Wonder Meta $4 $20 Standard mode
Muse Spark 1.3 Meta $1.25 $4.25 Standard mode
Qwen3.8 Max Qwen (Alibaba) $2 $6 International (Singapore) price
Qwen3.7 Plus Qwen (Alibaba) $0.4 $1.6 International (Singapore) price, prompts up to 256,000 tokens
Kimi K3 Moonshot AI $3 $15
GLM-5.3 Z.ai $1.4 $4.4
GLM-5.3 Flash Z.ai $0.15 $0.5
gpt-oss-120b OpenAI Free to run about 65 GB download (4-bit), roughly 88 GB of memory
gpt-oss-20b OpenAI Free to run about 14 GB download (4-bit), roughly 24 GB of memory
Gemma 4 31B Google Free to run about 19 GB download (4-bit), roughly 24 GB of memory
Gemma 4 26B A4B Google Free to run about 16 GB download (4-bit), roughly 24 GB of memory
Qwen3.8 27B Qwen (Alibaba) Free to run about 18 GB download (4-bit), roughly 24 GB of memory
Llama 4 Scout Meta Free to run about 67 GB download (4-bit), roughly 88 GB of memory
Mistral Small 3.2 24B Mistral Free to run about 15 GB download (4-bit), roughly 24 GB of memory

Octopus Score is our own composite of public benchmarks, updated . Methodology. Benchmark results belong to their publishers, credited above. Model names are trademarks of their owners; Octopus Prime is not affiliated with any model provider.

Put the right model on every agent. Octopus Prime keeps your OpenClaw agents working, whichever models they run.

Start free trial See how it works

30 days free · no card · from $5/month