New SOTA. 21.7 points above GPT-5.5 (58.6%). Agentic coding at scale.

Unsafeguarded Mythos-class model via Project Glasswing. Same underlying model as Fable 5 with cyber safeguards lifted.

24.4pp above Opus 4.6 (53.4%). Massive jump on harder coding eval.

Second only to Fable 5 (80.3). Beats GLM-5.2 (62.1).

Third on SWE-bench Pro behind Fable 5 (80.4) and Opus 4.8 (69.2). Notable for token efficiency: 4.2x fewer output tokens than Opus 4.8.

Beats GPT-5.5 (59.4%); trails Claude Mythos 5 (80.3%).

Up 10.9pp from Opus 4.6. Largest single-version jump on SWE-bench Pro.

Balanced-tier Terra beats GPT-5.5 (59.4%) at lower cost.

Cheapest tier Luna beats GPT-5.5 (59.4%) on SWE-bench Pro.

Strongest open-source score. Closes gap to Opus 4.8 (69.2). Up from GLM-5.1 (58.4).

Above GPT-5.5 (58.6), below GLM-5.2 (62.1).

Above GPT-5.5 (58.6). Competitive with GLM-5.2 (62.1).

Open-weight SOTA on SWE-bench Pro. 1T/32B MoE matching GPT-5.4 at 57.7%.

Matches Kimi K2.6 SOTA. Uses fewer tokens than GPT-5.4.

Same as GPT-5.5, listed separately in Pro column.

#1 open source, #3 globally. Beats GPT-5.4 (57.7) and Claude Opus 4.6 (57.3). First Chinese model to top SWE-bench Pro. Zero Nvidia hardware.

SWE-bench Pro. Competitive with DeepSeek V4 Pro (55.4).

Harder coding benchmark variant. GPT-5.4 splits coding, reasoning, and computer use into one unified model.

Cited alongside GPT-5.4 (57.7) and GLM-5.1 (58.4) in multiple independent comparisons. Opus 4.6 was previous SOTA before GLM-5.1.

Barely edges past Opus 4.5 (57.1). Top coding model from China. Proprietary, not open-weight.

Leads SWE-bench Pro. 25% faster than GPT-5.2 Codex. Covers 1,865 multi-language tasks.

SWE-bench Pro. 2nd place among open models, behind DeepSeek V4 Pro (55.4) is actually ahead.

Matches GPT-5.3 Codex level. Nearly approaches Opus best on SWE-Pro.

Max reasoning. Below GPT-5.5 (58.6) and Kimi K2.6 (58.6).

Public set. Below Opus 4.7 (64.3). Up from Gemini 3 Flash (49.6).

Smallest GPT-5.4 variant. 2x faster than GPT-5 mini at similar coding quality.

Open-weights MoE; trails GLM-5.2 (62.1) and Fable 5 (80.0) on harder SWE variant.

Competitive with GPT-5.4 (57.7%). Behind Claude Mythos Preview (77.8%).

Competitive with Qwen 3.5 397B MoE on agent coding.

Scored by Anthropic using Scale AI variant. Lower than GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%) on this harder variant.

SWE-bench Pro Public.

Impressive for 13B activated — competitive with much larger models.

Fastest, cheapest GPT-5.4 variant. Designed for classification, data extraction, and coding subagents.

Trails Claude Mythos (77.8%) significantly but competitive with other frontier models.

Internal evaluation framework with minimal tool set. 1T param MoE, 32B active.

3B active params MoE beating Gemma 4 31B (35.7%) on SWE-bench Pro.

Flagship coding model from ByteDance Seed team.

From TML Inkling blog. SWE-bench Pro Public.

Nearly matches Pro variant (46.9) on harder SWE subset.

SWE-Agent score. Competitive with much larger models on harder SWE subset.

3B active params matching models with 10x more active compute on SWE-bench Pro.

Struggles on harder SWE tasks vs MoE competitors.

MoE 4B active/26B total. Struggles with complex agentic coding tasks.