Inkling vs Qwen 3.6 Plus

Choose Inkling if you want…

• ✓Higher intelligence (Artificial Analysis index 46)

Choose Qwen 3.6 Plus if you want…

• ✓Lower price ($1.13 / 1M blended)
• ✓Larger context window (1M)

Inkling

Inkling is Thinking Machines Lab's first open-weights model, and on just4o.chat it runs through the Vercel AI Gateway under thinkingmachines/inkling. It is a multimodal MoE with 975B total parameters and 41B active, trained to reason over text, images, and audio in a shared space rather than bolting on separate encoders. Controllable thinking effort lets you trade latency for depth, function calling supports agentic workflows, and a 256k context window on the gateway is roomy enough for long multimodal sessions. It is a base-tier model on just4o.chat, billed at 4 base requests per send before length multipliers, at $1.00 per million input and $4.05 per million output tokens (cached input $0.17). Practical caveats: no native web search on this route, and chat attachments currently emphasize image input even though the model itself understands audio. For teams that want an open, customizable multimodal foundation with ZDR-compatible gateway routing, Inkling is a strong addition.

Qwen 3.6 Plus

At $0.50 per million input tokens, Qwen 3.6 Plus punches well above its price band — scoring 78.8 on SWE-bench Verified and 61.6 on Terminal-Bench 2.0, where it outpaces Claude 4.5 Opus on agentic coding tasks. The 1 million token context window lets you drop in entire codebases for security audits, multi-file refactors, or long-horizon agent sessions without chunking or worrying about cost. Always-on chain-of-thought reasoning is baked into the architecture rather than toggled per request, and native tool-calling makes it well-suited for multi-step workflows. Developers building high-volume API applications have reported generating hundreds of millions of tokens during its preview period — its first-day usage crossed one trillion tokens across platforms. That said, the long context is not a silver bullet: retrieval accuracy degrades in the middle of very long inputs, and real-world testing has surfaced instruction-following inconsistencies and occasional tool-calling failures that more mature providers handle more reliably. For cost-sensitive production deployments where coding and document analysis are the core workload, few models compete at this price.