Compare model APIs for agents that need tools, JSON mode, long context, and a budget you can live with.
Agent runs spit out a lot of tokens and burn money when tool loops go wrong. Before you wire one up, check tool calling, JSON reliability, context length, latency, and output price. Then set a budget so a bad loop cannot empty the account overnight.
Source basis: NextModel capability mapping and supported-parameter metadata when available. · Updated 2026-07-01
How to use this shortlist
Fit score
Start with the shortlist, then test real prompts and compare monthly cost before production routing in Canada.
Gemini 4 Argon via NextModel: 1M-token context; $2 / 1M tokens in / $10 / 1M tokens out per 1M tokens. Capabilities: streaming, tool, json.
Claude Haiku 5.5 via NextModel: 1M-token context; Starting at $0.1 / 1M tokens in / $0.5 / 1M tokens out per 1M tokens. Capabilities: streaming, tool, json.
Claude Sonnet 5.5 via NextModel: 1M-token context; $2 / 1M tokens in / $10 / 1M tokens out per 1M tokens. Capabilities: streaming, tool, json.
GPT‑6 Luna via NextModel: 1M-token context; Starting at $0.1 / 1M tokens in / $0.5 / 1M tokens out per 1M tokens. Capabilities: streaming, tool, json.
Comparison table
Use this view when narrowing a production shortlist, building a fallback policy, or comparing model economics for Canada-based teams.
| Model | Provider | Input | Output | Context | Capabilities | Best for | Latency | Status | Source |
|---|---|---|---|---|---|---|---|---|---|
| Gemini 4 Argongoogle/gemini-4-argon | $2 / 1M tokens | $10 / 1M tokens | 1M | StreamingTool callingJSON modeVision | image understanding, multimodal chat | 2974-10589ms | Catalog | Platform curated | |
| Claude Haiku 5.5anthropic/claude-haiku-5.5 | Anthropic | Starting at $0.1 / 1M tokens | $0.5 / 1M tokens | 1M | StreamingTool callingJSON modeVision | image understanding, multimodal chat | 1321-2017ms | Catalog | Platform curated |
| Claude Sonnet 5.5anthropic/claude-sonnet-5.5 | Anthropic | $2 / 1M tokens | $10 / 1M tokens | 1M | StreamingTool callingJSON modeVision | image understanding, multimodal chat | 1230-2392ms | Catalog | Platform curated |
| GPT‑6 Lunaopenai/gpt-6-luna | OpenAI | Starting at $0.1 / 1M tokens | $0.5 / 1M tokens | 1M | StreamingTool callingJSON modeVision | image understanding, multimodal chat | 1722-2681ms | Catalog | Platform curated |
| GPT‑6 Solopenai/gpt-6-sol | OpenAI | Starting at $2 / 1M tokens | $10 / 1M tokens | 1M | StreamingTool callingJSON modeVision | image understanding, multimodal chat | 1752-3715ms | Catalog | Platform curated |
| Claude Opus 5.5anthropic/claude-opus-5.5 | Anthropic | $4 / 1M tokens | $20 / 1M tokens | 1M | StreamingTool callingJSON modeVision | image understanding, multimodal chat | 1845-5119ms | Catalog | Platform curated |
| GPT-6 Astraopenai/gpt-6-astra | OpenAI | Starting at $10 / 1M tokens | $50 / 1M tokens | 1.1M | StreamingTool callingJSON modeVision | image understanding, multimodal chat | 2408-7945ms | Catalog | Platform curated |
| Claude Fable 5.1anthropic/claude-fable-5.1 | Anthropic | $10 / 1M tokens | $50 / 1M tokens | 1M | StreamingTool callingJSON modeVision | image understanding, multimodal chat | 3187-5152ms | Catalog | Platform curated |
FAQ
Tool calling, structured JSON, enough context for the task, and instructions that stick. Everything else is secondary.
They generate long traces: planning text, tool results stuffed back into context, and retries. Cap steps and log token use per run.
No. Route planning or simple tool picks to cheaper models when quality allows, and reserve stronger models for hard steps.
Related rankings
Related guides