Ultra mode (4 parallel agents) sets Terminal-Bench 2.1 SOTA.

Edges Claude Mythos 5 (88.0%); Ultra mode hits 91.9%.

KimiCode harness.

Safeguard-affected score (~Mythos 5). Terminal-Bench 2.1, not 2.0.

Matches GPT-5.6 Sol (88.8) on Terminal-Bench 2.1; Mythos 5 is the unsafeguarded twin of Fable 5 (83.1).

Terra beats GPT-5.5 (85.6%) on Terminal-Bench 2.1.

Beaten only by Fable 5 (88.0). Top open/commercial.

Luna near-parity with GPT-5.5 (85.6%) at lowest cost.

Score as cited in Kimi K3 blog (source: Anthropic official). Fable 5 hit fallbacks on 35% of tasks.

Score as cited in Kimi K3 blog (source: Anthropic official).

Score as cited in Kimi K3 blog (source: OpenAI official).

Within 1 point of Fable 5 (max) and GPT-5.5 (xhigh) at 84.3/83.4. First xAI model trained specifically for coding/agent tasks, 500K context.

Score as cited in Kimi K3 blog (source: Z.ai official).

Strongest open-source model. Terminus-2 harness. Up from GLM-5.1 (63.5).

Terminus-2 harness. Below GPT-5.5 (78.2). Up from Gemini 3 Flash (58.0).

Below GLM-5.2 (81.0) and Gemini 3.1 Pro (74).

From TML Inkling blog comparison table.

Terminus-2. Below Gemini 3.5 Flash (76.2).

Terminus-2. Lower than Gemini 3.5 Flash (76.2).

Official score 66.0 (Terminus-2). Internal infra, 8C16G sandbox, 2hr timeout.

Below GLM-5.2 (81.0). Open-weight 1M context model.

Uses internal coding harness. Below GLM-5.2 (82.7) but above Nemotron 3 Ultra (56.4).

Terminal-Bench 2.1 score from StepFun internal testing.

Terminal-Bench 2.1. +6.1 over Step 3.5 Flash.

Terminus-2. Below Gemini 3.5 Flash (76.2).

From TML Inkling blog.

Lower than Inkling (63.8) on Terminal-Bench.