What Changed in AI in the Last 90 Days (Quick Round-up)
The shifts that actually matter for builders - late May to mid-August 2026 The last three months did not produce a single "GPT-5 moment." There was no single release that reset the conversation the way earlier step-changes once did. Instead, the ground moved in several places at once: a wave of frontier and open-weight model launches in July, growing candor about how badly long-context windows…
In the past 90 days, the artificial intelligence landscape has seen a series of notable developments for those who develop AI systems. While no single model launch like GPT-5 altered the conversation, several key changes impacted the field.
Firstly, in terms of model capabilities, OpenAI released GPT-5.6 in three tiers - Sol, Terra, and Luna - following a government review. The fastest tier reportedly achieved a processing speed of 750 tokens per second on Cerebras hardware. Anthropic increased its lineup with Opus 5, priced similarly to competitors and outperforming them in benchmark tests. xAI released Grok 4.5 and Grok 4.6, the latter introducing a 500K-token-context model optimized for coding and long-running agents.
Google's Gemini Flash line saw three releases - 3.5, 3.6, and 3.7 Flash - with pricing reductions on each iteration.
Secondly, the emergence of agentic systems proved significant. Computer-use agents crossed a crucial threshold on the OSWorld-Verified leaderboard, with the top model (Qwen3.8 Max) reaching 86% across 22 evaluated models. This range is where real productivity begins. However, this progress came with a security concern highlighted by xAI's Grok Bot, which shares a single cloud computer, cookie store, and credential pool among all its bots.
This setup lacks security boundaries, making prompt-injection attacks feasible. The takeaway is that building multi-step agents requires treating the account or persistent VM as the trust boundary, not the individual agent.
Thirdly, infrastructure and serving improvements continued to make frontier capabilities more affordable. Techniques such as speculative decoding and aggressive quantization helped reduce costs without compromising quality. Streaming architectures and KV-cache management also improved for handling long-context usage in production. The gap between benchmarked performance and affordability is narrowing, making power and grid capacity more critical bottlenecks than GPU supply.
Fourthly, evaluation and safety underwent scrutiny. Long-context claims were found to be overstated, with most models only using 50-65% of their advertised context window. Additionally, positional bias came to light, affecting content retrieval accuracy based on its position within the document. OpenAI paused its deployment-focused RL training after concerns about a potential cybersecurity breach and the exposure of Hugging Face systems during testing. This pause led to a rewriting of their Preparedness Framework with stronger monitoring.
Lastly, smaller but important shifts included the maturation of the RAG vs. fine-tuning debate towards a more balanced approach focusing on efficiency and cost. Overall, the AI landscape remains dynamic, with a growing emphasis on efficiency, cost, and security.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
