# Model Economics Tighten as Cyber Agents and Post-Training Advance *By AI News Digest • July 18, 2026* Frontier-model competition is increasingly focused on price as Kimi K3 and Grok 4.5 post new comparative results. OpenAI expands defensive cyber tooling, NVIDIA emphasizes continuous post-training infrastructure, and Sakana AI tests a biologically constrained alternative to backpropagation. ## Frontier-model competition is moving from capability to economics ### Kimi K3 and Grok 4.5 sharpen the pricing contest Kimi K3 has taken first place on Frontend Code Arena and, in an internal editorial-writing benchmark, ranked first at 2840 Elo—displacing Claude Fable 5. The benchmark’s operator described it as the first open-weights model to top its board, with a reported cost of about $0.25 per script. [^1][^2] Grok 4.5 was separately reported at $0.31 per Artificial Analysis Intelligence Index task; the same comparison placed Claude Fable 5 at $2.75 and GPT-5.6 Sol at $1.04. Big Technology also reported lower-priced competitive offerings from Meta’s Muse Spark 1.1 and Grok 4.5, alongside K3’s potential to undercut U.S. labs’ pricing. [^3][^4] *Why it matters:* Big Technology’s analysis is that a larger field of frontier developers could compress model margins, shifting more of AI’s value toward products built on top of models. [^4] ## OpenAI deploys its latest model for defensive security work ### GPT-5.6 Sol reaches OpenAI’s cyber range and Codex Security OpenAI says GPT-5.6 Sol sets a new state of the art on its “The Last Ones” cybersecurity range. The company says the model is already helping teams find, validate, and fix vulnerabilities in real-world code. [^5] Access is available through OpenAI’s Codex Security plugin. [^5] *Why it matters:* This is a move from a cyber-evaluation result toward a dedicated product path for defensive software-security workflows. ## NVIDIA frames post-training as the next infrastructure race ### Vera Rubin is aimed at continuous agentic learning cycles NVIDIA says its Vera Rubin platform was codesigned for agentic post-training workloads—enabling more rollouts, more environments, and continuously repeated post-training cycles—and can train the largest models with one-fourth the GPUs of the Blackwell generation. [^6] The company also highlighted Nemotron 3 Ultra, a 550B-parameter open-weight mixture-of-experts model post-trained with NeMo RL, which scored 71.7% on SWE-bench Verified. Prime Intellect, Perplexity, and Together AI are respectively planning or using NVIDIA infrastructure for RL post-training, inference orchestration, and post-training-as-a-service. [^6] *Why it matters:* NVIDIA is positioning sustained post-training—not just initial model training—as a central compute workload for agentic AI. ## Sakana explores an alternative to backpropagation ### “Diffusing Blame” trains Dale-constrained networks with local error routing Sakana AI Labs introduced **Diffusing Blame**, a learning method designed to obey Dale’s principle: each neuron is dedicated to either excitatory or inhibitory signaling. The approach uses Error Diffusion with modulo error routing, rather than backpropagation and its requirement for transported copies of weights. [^7] The team reports competitive image-classification and reinforcement-learning results, including locomotion tasks and Craftax; the paper was accepted at ALIFE 2026. [^8][^7] *Why it matters:* The work tests whether useful representation learning and reinforcement learning can emerge under biological constraints that standard deep-learning systems typically ignore. --- ### Sources [^1]: [𝕏 post by @DavidSacks](https://x.com/DavidSacks/status/2078092271296143593) [^2]: [𝕏 post by @Whats_AI](https://x.com/Whats_AI/status/2077860441380798908) [^3]: [𝕏 post by @XFreeze](https://x.com/XFreeze/status/2078171481415290947) [^4]: [AI Model Prices Are Falling At The Worst Moment For The U.S. Frontier Labs](https://www.bigtechnology.com/p/ai-model-prices-are-falling-at-the) [^5]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2078243667081617826) [^6]: [NVIDIA Vera Rubin Maximizes Intelligence per Dollar for Post-Training Workloads — a Key Metric for Agentic AI](https://blogs.nvidia.com/blog/nvidia-vera-rubin-post-training-intelligence-per-dollar) [^7]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2078136419521048905) [^8]: [𝕏 post by @hardmaru](https://x.com/hardmaru/status/2078156625479921847)