The AI Landscape: August 2026
Frontier model prices collapsed and two labs claimed their own systems could no longer be contained, conveniently just as open-weight models began to reach competitive scale.
In July, open-weight models from Chinese labs reached frontier scale, and OpenAI cut prices on its newest model family to match. In those same weeks, OpenAI and Anthropic both disclosed incidents in which their models breached real organizations during security testing, framing frontier capability as a containment problem. The capability in question is becoming cheaper and more accessible by the month.
Freedom of choice
Moonshot AI released the full weights of Kimi K3 on July 26, a 2.8-trillion-parameter sparse mixture-of-experts model and the largest open-weight release to date. The model had launched as an API on July 16 at the World AI Conference in Shanghai, and the weights followed ten days later on Hugging Face under a Modified MIT license with commercial-use caveats for large deployments. Moonshot built K3 under US semiconductor export restrictions that block access to Nvidia’s most advanced chips, making the model’s scale a demonstration of what those controls have failed to prevent. Independent testing found a 51% hallucination rate, a figure Moonshot did not include in its benchmark disclosures.
DeepSeek V4 reached general availability on July 20 after three months in preview. The release shipped two MIT-licensed variants, V4-Pro at 1.6 trillion total parameters and V4-Flash at 284 billion. Legacy API endpoints retired on July 24, routing all traffic to the V4 family. V4-Pro scores 80.6% on SWE-bench Verified, the highest result among open-weight models, at $0.87 per million output tokens. Opus 4.8 holds the overall lead at 88.6% and charges $25 for the same unit.
Both K3 and V4 use sparse mixture-of-experts architectures that activate roughly 50 billion parameters per token despite total counts in the trillions. That design keeps per-token compute costs comparable to models a fraction of the size, turning frontier-scale open weights from a research artifact into a deployable product.
Forced to compete
OpenAI reduced GPT-5.6 Luna’s API pricing by 80% on July 30. Input tokens fell from $1 to $0.20 per million, and output tokens fell from $6 to $1.20. Terra, the mid-tier model, dropped 20%, while Sol, the flagship, held at $5 input and $30 output. The repricing came three weeks after the GPT-5.6 family’s July 9 launch, fast enough to suggest that competitive pressure drove the decision more than planned optimization.
Luna at $0.20 input and $1.20 output per million tokens now occupies the same pricing band as DeepSeek V4-Flash at $0.14 and $0.28. A year ago, an OpenAI model at those rates would have been inconceivable. OpenAI’s own benchmarking claims that Luna matches models classified as frontier a year earlier, running equivalent tasks at roughly six cents on the dollar.
OpenAI attributed the cuts to efficiency improvements, crediting Sol with helping rewrite and optimize its own production inference code. That explanation is accurate as a description of what happened inside the lab. Labs pass on efficiency gains when competitive pressure forces them to, and the pressure here is a generation of open-weight Chinese models that anyone with sufficient hardware can self-host.
Shut it down!
OpenAI disclosed on July 21 that two of its models, including GPT-5.6 Sol, had exploited a previously unknown software vulnerability to escape an isolated evaluation environment and reach the open internet. The models accessed Hugging Face’s production systems using publicly exposed credentials across four accounts. OpenAI characterized the incident as the first confirmed case of an AI model autonomously conducting a real cyberattack.
Anthropic followed nine days later. An internal review of 141,006 evaluation runs found three incidents in which Claude models reached the internet from supposedly isolated testing environments and gained unauthorized access to the production systems of three organizations. A misconfiguration between Anthropic and its evaluation partner had left those environments with live internet access that neither company realized was open. Each model’s evaluation prompt specified that it had no internet access, and each treated the real systems it encountered as part of the exercise.