Grok 4.3 Release, DeepSeek V4 Pro, and Codex Product Expansion
Patch internal systems to evaluate Grok 4.3's 40% lower input pricing and DeepSeek V4 Pro's 4× lower inference FLOPs for cost‑effective agent workloads.
Patch internal inference stack to use Grok 4.3 for cost‑effective workloads and test DeepSeek V4 Pro's 1M context for agentic coding.
Summary
xAI released Grok 4.3, scoring 53 on the Intelligence Index, up 4 points from Grok 4.20, with 40% lower input and 60% lower output pricing. The model gained 321 Elo on GDPval‑AA, reaching 1500, and achieved 98% on τ²‑Bench Telecom and 81% on IFBench, though its non‑hallucination dropped by 8 points. DeepSeek V4 Pro offers 1M context, hybrid CSA/HCA attention, 10% KV cache, and 4× lower inference FLOPs, making it comparable to Codex and Claude Code for multi‑turn agentic coding without custom setup. The model supports multi‑step research/coding loops on Fireworks inference with stable traces and is an open‑weight 1.6T/49B active system, closing the gap with top closed‑source models. Codex continues to expand its product, adding a device toolbar, CI status in chat, migration/import tooling, and a viral pets system, emphasizing a cohesive environment over a single endpoint.
These releases highlight a trend toward cost‑effective, high‑performance agents that can be integrated into existing workflows.
The focus on lower pricing, higher context, and production‑ready harnesses reflects the industry’s shift toward scalable, enterprise‑grade AI solutions.
Key changes
- Grok 4.3 scores 53 on Intelligence Index, up 4 points from Grok 4.20, with 40% lower input and 60% lower output pricing.
- Grok 4.3 gains 321 Elo on GDPval‑AA, reaching 1500, and hits 98% on τ²‑Bench Telecom.
- DeepSeek V4 Pro offers 1M context, hybrid CSA/HCA attention, 10% KV cache, and 4× lower inference FLOPs.
- DeepSeek V4 Pro is comparable to Codex and Claude Code for multi‑turn agentic coding, requiring no custom setup.
- DeepSeek V4 Pro supports multi‑step research/coding loops on Fireworks inference with stable traces.
- DeepSeek V4 Pro is an open‑weight 1.6T/49B active model, closing the gap with top closed‑source models.