Model Selection

By default, NinjaCat Agents are driven by Claude Sonnet 4.6, Anthropic's best model for complex agents and coding. However, depending on your Agent’s specific needs—whether for faster response times or enhanced reasoning—you can choose an alternative model.NinjaCat will adjust the default model when newer, better models come out after spending time testing that the output is as good or better than the prior default, and if it is more cost efficient. When the default is changed, it only impacts brand new Agents. Users will need to adjust the model for existing Agents if they would like to. NinjaCat will not adjust the model for existing agents to avoid disrupting an already perfectly crafted Agent (sometimes a model change could require tweaks to the prompt).


Available Model Options in NinjaCat

Anthropic Models

ModelRelease DateContext WindowInput $ / 1MOutput $ / 1MNotes
Claude Sonnet 5Jul 2026~200K~$3~$15Newest Sonnet-tier model. Opt-in per agent. Pricing shown at general availability; introductory pricing applies through Aug 31, 2026.
Claude Sonnet 5 - ThinkingJul 2026~200K~$3~$15Adaptive-thinking variant of Claude Sonnet 5 for deeper reasoning. Opt-in per agent.
Claude Fable 5Jun 2026~1M~$10~$50Newly restored on 7/1 following Anthropic's reinstatement announcement
Claude Opus 4.8Jun 2026~1M~$5~$25Latest Opus model from Anthropic; sharper reasoning and stronger performance across complex tasks.
Claude Opus 4.7Apr 2026~1M~$5~$25Latest Opus model; expanded ~1M context window
Claude Opus 4.7 - ThinkingApr 2026~1M~$5~$25Enhanced reasoning variant of Opus 4.7; uses more output tokens due to adaptive thinking
Claude Opus 4.6Feb 2026~200K~$5~$25Superseded by Claude Opus 4.7; although costs are the same, less tokens are used on 4.6 so should still be used for less complex tasks.
Claude Opus 4.6 - ThinkingFeb 2026~200K~$5~$25Enhanced reasoning variant of Opus 4.6
Claude Sonnet 4.6 (Default)Feb 2026~200K~$3~$15New default for all newly created Agents. Best-balanced Claude model — strong performance at moderate cost
Claude Sonnet 4.6 - ThinkingFeb 2026~200K~$3~$15Enhanced reasoning variant of Sonnet 4.6
Claude Haiku 4.5Oct 2025~200K~$1~$5Fastest & most affordable Claude option
Claude Haiku 4.5 - ThinkingOct 2025~200K~$1~$5Enhanced reasoning variant of Haiku 4.5

OpenAI Models

ModelRelease DateContext WindowInput $ / 1MOutput $ / 1MNotes
GPT-5.6 Sol (New)Aug 2026~1M~$5.00~$30Most capable GPT-5.6 tier; built for demanding coding and analysis tasks. Standard (instant) mode only — reasoning/thinking mode planned for future update.
GPT-5.6 Terra (New)Aug 2026~1M~$2.00~$12Balanced everyday performance at a lower cost than Sol. Standard (instant) mode only — reasoning/thinking mode planned for future update.
GPT-5.6 Luna (New)Aug 2026~1M~$0.20~$1.20Fast and lightweight; optimized for high-volume, simpler tasks. Most cost-efficient model option in NinjaCat — roughly 5× cheaper than Claude Haiku on both input and output. Standard (instant) mode only — reasoning/thinking mode planned for future update.
GPT-5.5 (New)Jun 2026~1M~$5.00~$30OpenAI's latest model; improvements to speed and instruction
GPT-5.4 - Instant (New)Mar 2026~1M~$2.50~$1533% fewer errors vs GPT-5.2, massive 1M context window
GPT-5.4 - Thinking (New)Mar 2026~1M~$2.50~$15Enhanced reasoning variant of GPT-5.4; most token-efficient reasoning model
GPT-5.2 - ThinkingOct 2025~400K~$1.75~$14Strong reasoning + long context. Default reassignment for agents on removed OpenAI models
GPT-5.2 - InstantOct 2025~400K~$1.75~$14Faster, lower-latency variant of GPT-5.2
GPT-5 MiniAug 2025~400K~$0.25~$2Cost-efficient; good for well-defined tasks at lower cost
GPT-5 NanoAug 2025~400K~$0.05~$0.40Cheapest & fastest GPT-5 variant; great for summarization and classification workloads

GPT-5.6 models — standard mode only: GPT-5.6 Sol, Terra, and Luna run in standard (instant) mode today. Reasoning/thinking mode is not yet available for these models and is planned for a future update.

Google Models

ModelRelease DateContext WindowInput $ / 1MOutput $ / 1MNotes
Gemini 3.5 Flash (New)Jun 2026~1M~$1.50~$9.00Google's efficiency-optimized frontier model designed for high-volume, agentic, and coding workloads
Gemini 3.1 Pro - LowFeb 2026~1M~$2.00~$12.00Google's latest Pro model; improved reasoning, multimodal, and agentic capabilities
Gemini 3.1 Pro - HighFeb 2026~1M~$2.00~$12.00Higher reasoning effort variant of Gemini 3.1 Pro
Gemini 3.1 Flash Lite - LowFeb 2026~1M~$0.25~$1.50Most cost-efficient Google model; optimized for high-volume agentic tasks
Gemini 3.1 Flash Lite - HighFeb 2026~1M~$0.25~$1.50Higher reasoning effort variant of Gemini 3.1 Flash Lite
Gemini 3 Pro - LowNov 2025~1M~$2.00~$12.00Best-in-class reasoning & multimodal from Google; massive context window
Gemini 3 Pro - HighNov 2025~1M~$2.00~$12.00Higher reasoning effort variant
Gemini 3 Flash - LowDec 2025~1M~$0.50~$3.00Fast, efficient; combines Gemini 3 Pro reasoning with Flash-level latency and cost
Gemini 3 Flash - HighDec 2025~1M~$0.50~$3.00Higher reasoning effort Flash variant

Fireworks AI Models

ModelRelease DateContext WindowInput $ / 1MOutput $ / 1MNotes
GLM-5.2 (Default)Jun 2026~128K~$1.40~$4.40Strong agentic and long-horizon task performance; well-suited to complex multi-step workflows and software-engineering tasks. Default Fireworks provider model.
DeepSeek V4 ProApr 2026~128K~$1.74~$3.48Full-capability flagship for very long document analysis or extended reasoning; cost-efficient frontier performance.
DeepSeek V4 FlashApr 2026~128K~$0.14~$0.28Budget-friendly variant of V4 — significantly cheaper than Pro with fast responses; ideal for high-volume or latency-sensitive workloads.
Kimi K2.7 CodeJun 2026~128K~$0.95~$4.00Specialized for coding and agentic coding tasks with strong token efficiency for agent workflows; supports image input (vision).
Qwen 3.7 PlusJun 2026~128K~$0.50~$3.00Cost-efficient agent model with a hybrid thinking mode (reasoning on/off); good for structured outputs and tool use; supports image input (vision).
MiniMax M3Jun 2026~128K~$0.30~$1.20Native multimodality with large-context reasoning; designed for long-context reasoning and agentic tool use; supports image input (vision).

Note: In the Agent Builder, some models offer both a standard and a "Thinking" variant. The Thinking variant supports deeper reasoning but may come with higher cost and latency. For most agents, the standard variant is recommended unless your use case requires complex multi-step reasoning.


How to Choose the Right Model

AI models are continuously improving — what is "best" today may be surpassed in weeks or months. NinjaCat will continue evaluating and adding models that demonstrate better intelligence, efficiency, or performance.

General guidance:

  • For most agents: Claude Sonnet 4.6 (default) is the best starting point — strong performance at reasonable cost.
  • For the most complex, high-effort, or long-horizon tasks: Claude Opus 4.8 or Opus 4.7 — Anthropic's highest-capability models available on the platform. (Note: Claude Fable 5 is currently unavailable due to a government directive — see the Anthropic models table above.)
  • For complex reasoning or coding tasks: Claude Opus 4.7, GPT-5.4 - Thinking, GPT-5.2 - Thinking, or GPT-5.6 Sol (standard mode).
  • For speed or cost-sensitive tasks: Claude Haiku 4.5, GPT-5.6 Luna (most cost-efficient option on the platform — ~5× cheaper than Claude Haiku), GPT-5 Nano, Gemini 3.1 Flash Lite, or Gemini 3 Flash.
  • For balanced everyday performance at a lower cost: GPT-5.6 Terra is a good middle-ground option within the GPT-5.6 family.
  • For large context windows: OpenAI GPT-5.4 series (~1M), GPT-5.6 series (~1M), or Google Gemini series (~1M).

For the latest information from each provider, see their documentation:

Anthropic Claude Models OpenAI GPT-5 Prompting Guide Google Gemini

Note: When switching between models, prompt adjustments may be required to maintain optimal Agent performance. We will provide further guidance on prompt modifications as we continue testing and learning.



Did this page help you?