Choose the right model for every agent.
Compare execution, research, and image-generation models by capability, context, provider, and CT cost before routing work through agents, SDKs, or project tasks.
Start now
Compare frontier and efficient models for coding, operations, browser work, and product tasks.
Choose deeper research paths when agents need stronger coverage, synthesis, and source review.
Create and edit visuals from the same workspace agents use for files, plans, and deliverables.
Review CT multipliers, defaults, context windows, and BYOM options before routing work.
Compute Tokens are ACP’s unified billing unit for both model usage and runtime. Model costs are passed through directly, and the CT Cost and CT Baseline columns below show weighted average consumption and relative cost using the same baseline across the docs and platform.
Relative CT multipliers are calculated against a stable billing baseline. This is a pricing reference, not the default execution model.
New and existing managed system agents use DeepSeek V4 Pro by default, except the Developer agent, which uses Kimi K2.6 for coding-oriented execution.
Kimi K2.6
DeepSeek V4 ProDeepSeek V4 ProDeepSeek V4 ProCT Cost uses this representative token mix so model costs stay comparable at a glance.
Model comparison
| Model | Intelligence | Speed | Context | CT Cost | CT Baseline |
|---|---|---|---|---|---|
 Claude Opus 4.7 Anthropic’s latest flagship for the most complex reasoning and agentic coding work. | Medium | 1M | 855 CT / 1M | 5.0x | |
Claude Opus 4.6 Previous-generation flagship Claude model for complex reasoning and coding. | Medium | 1M | 855 CT / 1M | 5.0x | |
Claude Sonnet 4.5 Best balance of intelligence and speed. Excellent for coding and daily tasks. | Fast | 200k | 513 CT / 1M | 3.0x | |
Claude Haiku 4.5 Fast and cost-effective for quick tasks and high-volume operations. | Very Fast | 200k | 171 CT / 1M | 1x (baseline) | |
 GPT-5.5 Pro OpenAI highest-accuracy model for the hardest professional and agentic work. | Medium | 1M | 6000 CT / 1M | 35.1x | |
GPT-5.5 OpenAI frontier model for coding, professional work, and long-context agents. | Fast | 1M | 955 CT / 1M | 5.6x | |
GPT-5.4 OpenAI flagship model for advanced coding, planning, and computer use. | Fast | 1M | 478 CT / 1M | 2.8x | |
GPT-5.4 mini OpenAI mini model tuned for fast coding, subagents, and computer-use tasks. | Fast | 400k | 143 CT / 1M | 0.84x | |
GPT-5.4 nano Lightweight OpenAI model for simple instructions, routing, and high-volume tasks. | Very Fast | 400k | 39 CT / 1M | 0.23x | |
 Gemini 3.1 Pro Higher-reasoning Gemini model for complex planning and analysis. | Fast | 1M | 513 CT / 1M | 3.0x | |
Gemini 3 Flash Fast default model for everyday agent execution. | Very Fast | 1M | 171 CT / 1M | 1x (baseline) | |
Gemini 3.1 Flash Fast Gemini tier that currently routes through the same runtime as Gemini 3 Flash. | Very Fast | 1M | 171 CT / 1M | 1x (baseline) | |
 DeepSeek V4 ProDefault DeepSeek flagship model running through Clawcode for agentic coding and long-context reasoning. | Fast | 1M | 48 CT / 1M | 0.28x | |
DeepSeek V4 Flash Fast DeepSeek V4 model running through Clawcode for efficient agent execution. | Very Fast | 1M | 16 CT / 1M | 0.09x | |
 Kimi K2.6 Moonshot flagship served via Cloudflare Workers AI for long-horizon coding and agentic execution. | Fast | 262k | 148 CT / 1M | 0.87x |
Research models
| Model | Intelligence | Speed | Context | Pricing | Best for |
|---|---|---|---|---|---|
Gemini 3.1 FlashDefault Faster deep research path tuned for speed and lower cost. | Very Fast | 1M | 171 CT / 1M | Quick researchSummaries | |
Gemini 3.1 Pro Higher-depth deep research path tuned for stronger coverage. | Fast | 1M | 513 CT / 1M | Deep researchComplex reasoning |
Image generation models
| Model | Quality | Speed | Mode | Pricing | Best for |
|---|---|---|---|---|---|
GPT Image 2Default Highest-fidelity OpenAI image model with quality-based CT billing. | Highest | Medium | Generation and editing | 1-12 CT/img | Final assetsText rendering |
Gemini 3.1 Flash Image Preview Fast multimodal image generation and editing preview with a flat CT estimate. | High | Very Fast | Generation and editing | 7 CT/img | Fast iterationLower-cost drafts |
Bring your own models
Team and Enterprise workspaces can connect external model providers and route agent execution through their own model inventory. That includes OpenAI-compatible endpoints, custom inference backends, and workspace-level model catalogs exposed through ACP’s external model connection flow.