Merge Gateway finds GLM 5.3 beats Claude

// 1h agoBENCHMARK RESULT

Merge Gateway finds GLM 5.3 beats Claude

Merge Gateway evaluated five open-weight models—GLM 5.3, GLM 5.3 Flash, DeepSeek V4 Pro, DeepSeek V4 Flash, and Kimi K3—against Anthropic’s Claude Sonnet 5 across 20 real coding challenges, with GLM 5.3 solving 14 tasks versus Claude’s 12 at $0.317 per successful solve versus $3.415 and a faster median completion time (170s versus 197s). DeepSeek V4 Flash matched Claude’s 12/20 solve rate at $0.032 per successful task—106x cheaper but slower at 261 seconds—and all tested models are accessible through Merge Gateway’s unified API.

// ANALYSIS

Open-weight models have caught up with frontier proprietary models on coding agent tasks, radically undermining closed-source pricing power.

• The economic moat of proprietary frontier models is breaking: DeepSeek V4 Flash delivering the exact same task completion rate as Claude Sonnet 5 at 106x lower cost makes premium API pricing hard to sustain for high-throughput coding agents.

• GLM 5.3 establishes a new benchmark standard: By outperforming Claude Sonnet 5 on task resolution (14/20 vs. 12/20), cost ($0.317 vs. $3.415), and speed (170s vs. 197s), GLM 5.3 proves that open-weight options can lead across the entire Pareto frontier.

• Speed remains the compromise for ultra-cheap inference: DeepSeek V4 Flash was the slowest model tested at 261s per solve, whereas GLM 5.3 Flash led raw execution speed at 111s despite a lower task completion rate (9/20).

• Dynamic routing becomes indispensable: With cost and latency variances spanning orders of magnitude across capable models, unified LLM gateways are essential for optimizing agent workflows in production.

// TAGS

merge-gatewayllm-benchmarkscoding-benchmarksclaude-sonnet-5glm-5.3deepseek-v4kimi-k3open-weights

DISCOVERED

1h ago

2026-09-15

PUBLISHED

1h ago

2026-09-15

RELEVANCE

8/ 10

AUTHOR

merge_api

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

LAUNCH30m ago

MYWAI ports VILMA to Arduino AI boards

Developed under the European ARISE I-GENIUS research initiative alongside Centro Ricerche FIAT and the University of Genoa, MYWAI's VILMA (Visual Imitation Learning for Manipulation Activities) toolkit enables robotic manipulators to acquire kitting and pick-and-place tasks by watching a single human demonstration. The end-to-end pipeline extracts 3D hand landmarks via MediaPipe, tracks objects with YOLOv8, and models motion trajectories using Dynamic Movement Primitives (DMPs) that adapt dynamically to new object coordinates. The toolkit has now been ported to edge hardware, running vision inference and motion adaptation directly on Qualcomm-powered Arduino UNO Q and VENTUNO Q boards interfaced with ROS 2 and robotic arms like the Fairino FR10.

RESEARCH41m ago

CoG adds cognitive cycles to GraphRAG

Cognition on Graph (CoG) is a training-free, open-source framework that alternates planning, graph-text exploration, synthesis, and reflection for multi-hop question answering. Across seven benchmarks, it reports stronger accuracy while accessing 12.4× fewer documents than ToG-2.

LAUNCH2h ago

Bolt Forge launches open models, 50x usage

Bolt.new launched Forge, an experimental agent running GLM 5.3, Kimi K3, and DeepSeek V4 Pro. Individual Pro users receive up to 50x more Forge usage through October 14 in exchange for optional anonymized prompts, code, and fix traces that help Arcee AI train an open-weight model. [Bolt announcement](https://bolt.new/blog/what-is-bolt-forge)