OpenAI publishes the o3 system card, with cost-per-answer figures.

OpenAI has released Sora 2, with synchronised audio output, longer maximum clip durations, and noticeable improvements in physical plausibility over the original v1 preview. Audio-synced video at this quality tier expands the addressable use cases beyond marketing B-roll. Worth a scoped evaluation if your team works on creative tooling or short-form content.

Mistral has cut Mistral Large 2 inference and fine-tuning pricing and added a managed fine-tuning path through La Plateforme aimed at teams that want a closed-source alternative to OpenAI and Anthropic. Pricing pressure at the closed-source frontier is real now. If you have been on a single-vendor contract, the renewal conversation just got more interesting.

Microsoft has continued the Phi-3 line of small language models, with sizes from 3.8B aimed at on-device inference and a quality bar Microsoft is benchmarking against larger frontier models. If small-model claims hold on your eval set, the cost-per-good-output for a class of tasks moves a step. Useful if you have been priced out of frontier inference for high-volume workloads.

Researchers from Google DeepMind and Berkeley have argued that, for a fixed quality target, scaling inference-time compute can sometimes substitute for additional pre-training compute, with worked-example trade-offs across maths and reasoning. If the trade-off generalises, smaller models with more thinking time are a viable architecture for cost-sensitive deployments. Worth a careful read before the next budget cycle.

OpenRouter has expanded its public dashboards to include p50 and p99 latency per model and current pricing across providers, alongside the existing fallback-routing primitives. Like-for-like latency data has been the missing input for production routing decisions. Worth bookmarking even if you do not route through OpenRouter directly.

NVIDIA's NIM offering bundles open-weight models into containerised microservices with TensorRT-LLM optimisations, an OpenAI-compatible API, and Helm charts for Kubernetes deployment. If you self-host inference on NVIDIA GPUs, NIM closes the gap between 'open-weight model' and 'production-ready endpoint' without a long DevOps tail.

Hugging Face has released smolagents, a deliberately small Python library for agent loops that writes actions as Python code rather than JSON tool calls, with built-in sandboxing options. Code-as-action is a noticeable ergonomic upgrade over JSON tool calling for many tasks. Worth a half-hour even if you stay on your current framework.

Hugging Face has continued the SmolLM family of small open-weight models (135M / 360M / 1.7B) trained on a curated dataset, with the training recipes published for reproducibility. Small-and-reproducible is the missing tier between 'too big to host' and 'hosted-API'. If you want to fine-tune your own model without renting an H100 cluster, this is the level the field is converging on.