Qwen 3.6 and Grok 4.3 in Critique: cheaper routing, stronger coding signal, and one absurd xAI price cut

This update is less about adding names to a dropdown and more about cleaning up our routing ladder. The old Qwen3.5-27B slot had become hard to justify once Qwen shipped a 35B/3B-active successor that is better on repository reasoning, better on frontend tasks, and still cheap enough to use as a routine specialist. On the xAI side, Grok 4.3 gives us a simpler story: reasoning is always on, multimodal input stays available, the context window remains large, and the credit floor falls sharply enough that it stops being a novelty lane and starts being a real option.

PART ONE - GROK 4.3 REPLACES GROK 4.2

xAI’s current developer docs surface grok-4.3 as the active flagship in the overview flow, with a 1M-token context class and text-plus-image input support. Their model docs also keep the important behavior constraint from the Grok 4 family: reasoning is built in, and there is no separate reasoning-effort dial for the standard Grok 4 line. That makes Grok 4.3 a clean fit for Critique. We do not need to explain hidden mode switches to users, and we do not need one price for “thinking” and another for “non-thinking.”

PART TWO - QWEN3.6-35B-A3B REPLACES QWEN3.5-27B

This is the cleaner benchmark story in the release. Qwen’s official Hugging Face card for Qwen3.6-35B-A3B lists a 35B total / 3B active MoE architecture, 262,144 native context, and extension to roughly 1.01M tokens. The published coding-agent table is mixed on the SWE rows, but it shows clear gains on the repo-scale and front-end-shaped tasks we care about most: Terminal-Bench 2.0, Claw-Eval average, SkillsBench, QwenClawBench, NL2Repo, and QwenWebBench.

One important correction to the simplistic launch line: the official Qwen table does not show a clean sweep on every single coding number. Qwen3.5-27B still posts higher values on some SWE rows. What it does show is that Qwen3.6-35B-A3B is stronger on the repo-scale, terminal, browser, and agentic workflow benchmarks that matter more to Critique’s review and Remedy loops, while also costing less in our catalog. That is enough to retire the older slot.

PART THREE - LING-2.6-FLASH JOINS AT 1 CREDIT

InclusionAI positions Ling-2.6-Flash as a fast instruct model for real-world agents: 104B total parameters, 7.4B active parameters, and a focus on token-efficient execution rather than theatrical benchmark chasing. The official Hugging Face card emphasizes lower token usage across coding, document processing, and lightweight workflows. That makes it a natural cheap specialist lane in Critique, especially for teams that want something faster and broader than a tiny extraction model but still do not want to pay mid-tier prices.

PART FOUR - QWEN3.6-MAX-PREVIEW JOINS AT 8 CREDITS

Alibaba positions Qwen3.6-Max-Preview as the higher-end proprietary Qwen lane. The official Model Studio docs list a 256K context window with thinking mode, function calling, and structured output support. Alibaba’s launch writeup says the preview model improves on Qwen3.7-Plus by +9.9 on SkillsBench, +6.3 on SciCode, +5.0 on NL2Repo, and +3.8 on Terminal-Bench 2.0, then summarizes the release by saying it leads six major coding benchmarks in their internal comparison set.

PART FIVE - GLM DOWN, KAT UP, MIMO UP

Three more credit moves round out the refresh. z-ai/glm-5.1 drops from 4 credits to 3, making it a more attractive mid-tier generalist. kwaipilot/kat-coder-pro-v2 rises from 1 credit back to its normal 2-credit shelf price; the lower number was a month-long partnership discount with the Kwaipilot team, not the permanent list price. And xiaomi/mimo-v2.5 rises from 0.5 to 1.5 credits, ending the unusually cheap launch positioning it held earlier.