GPT-6 Astra leads the BrowseComp benchmark for agentic web research, though Kimi K3 offers near-identical performance at a fraction of the cost.

$GPT-6 Astra leads the BrowseComp benchmark for agentic web research, though Kimi K3 offers near-identical performance at a fraction of the cost.$

OPEN LINK ↗

// 1d agoBENCHMARK RESULT

GPT-6 Astra leads the BrowseComp benchmark for agentic web research, though Kimi K3 offers near-identical performance at a fraction of the cost.

ANNOUNCEMENT PRODUCT X

The latest BrowseComp leaderboard reveals GPT-6 Astra at the top with a 91.5% score, closely trailed by Kimi K3 at 91.2% and Claude Opus 5 at 90.8%. However, Kimi K3 costs significantly less at $14.25 per million output tokens compared to Astra's $50 per million. The post highlights the viability of using routing tools like Merge Gateway to dynamically allocate tasks between models based on their performance and cost efficiency.

// ANALYSIS

The tightening gap in flagship model capabilities makes cost the primary differentiator for scaled applications.

–GPT-6 Astra holds the performance edge but demands a premium price.
–Kimi K3 provides an exceptional price-to-performance ratio for web research tasks.
–Claude Opus 5 sits competitively in the middle ground of both cost and capability.
–Gateway routing solutions are increasingly necessary to optimize unit economics in LLM applications.

// TAGS

aillmbenchmarkbrowsecompgpt-6-astrakimi-k3claude-opus-5merge-gateway

DISCOVERED

1d ago

2026-09-22

PUBLISHED

1d ago

2026-09-22

RELEVANCE

8/ 10

AUTHOR

merge_api

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE28m ago

Keel drops local-first macOS coding workspace

Keel is an open-source, local-first macOS coding workspace written in Rust and GPUI that introduces a host-verified decision architecture for AI agents. Instead of giving models full execution control, Keel prepares and verifies candidate routes before execution, supporting both local Core ML inference and hosted selectors via the Agent Client Protocol.

MODEL1h ago

Google drops Gemini 3.8 Flash TTS

Google DeepMind has launched Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, offering expressive audio generation with granular control over style, pacing, and emotional inflection. Available via Google AI Studio and the Gemini API, the release includes a high-fidelity Flash tier for creative media alongside an ultra-fast Flash-Lite tier for real-time voice applications.

NEWS1h ago

Grok Build generates 3D jet from single image

In a demonstration shared on X, a creator showed xAI's Grok 4.7 transforming a single 2D image into a fully realized 3D model of a jet directly within Grok Build. The workflow illustrates expanding multimodal and spatial reasoning capabilities within xAI's development environment, bridging the gap between 2D concept images and interactive 3D assets.