AI News: August 14, 2026

1. DeepSeek Ships an Improved V4-Pro and Open-Sources Its Harness Agent Stack

DeepSeek. DeepSeek graduated its V4-Pro model out of testing and released its Harness v0.1 agent software under an MIT license, while sharply raising API prices, with cache-hit costs climbing roughly sixfold. The open agent stack gives developers a reference implementation for building on DeepSeek models, but the pricing change hits repeated-file and long-running agent workflows hardest. Teams running cost-sensitive agents on DeepSeek will need to weigh the stronger model against the new economics. Source

2. Ant Group’s Ling 3.0 Flash Claims the Open-Weight Lead in Its Size Class

Ant Group. Ant Group released Ling 3.0 Flash, positioning it as the strongest open-weight model in its size class. The compact release gives developers a capable option for cost- and latency-sensitive deployments where larger frontier models are impractical. It adds to a steady stream of competitive open-weight models coming out of Chinese labs. Source

3. Writer Builds a New Model on GLM-5.2 With a Harness Aimed at Token Costs

Writer. Enterprise AI vendor Writer introduced a new model built on GLM-5.2, paired with an upgraded harness designed to contain token costs in production. The company is targeting enterprises worried about runaway inference spend as agentic workloads scale. It reflects a broader shift toward optimizing the full serving stack, not just the base model, to keep deployments economical. Source

4. Databricks Raises $5 Billion at a $190 Billion Valuation

Databricks. Databricks closed a $5 billion round at a $190 billion valuation after investor demand far outran its original plan to raise about $1 billion. CEO Ali Ghodsi pointed to the high cost of AI infrastructure as the rationale for the larger raise. The round underscores how much capital data and AI platform companies are pulling in to fund compute and expansion. Source

5. Anthropic Reportedly Weighs a $6 Billion Acquisition of Decart

Anthropic. Anthropic is reportedly weighing its largest-ever acquisition, a roughly $6 billion deal for Decart, a startup working on generative video and chip optimization and known for its Oasis 3 world model. The reported price is about 50% above Decart’s $4 billion valuation from May and would expand Anthropic’s multimodal and video capabilities. The talks are unconfirmed and not finalized, so terms could change or the deal could fall through. Source

6. IBM Partners With OpenAI to Certify Tens of Thousands of Consultants

IBM. IBM and OpenAI formed an enterprise partnership under which IBM will train and certify tens of thousands of its consultants on OpenAI technology. The arrangement extends OpenAI’s distribution deeper into large enterprises through IBM’s consulting arm. For buyers, it signals more channel support and integration services around OpenAI models in complex deployments. Source

7. Suno Studio 2.0 Adds a Chat Interface to Drive Music Production

Suno. Suno released Studio 2.0, an update that turns the platform into a fuller production suite for Premier subscribers, adding a chat interface to direct AI instrument generation and unlimited 32-bit audio exports. The chat-driven controls let users steer a digital audio workstation conversationally rather than through manual editing. The unlimited-export feature sits awkwardly against Suno’s recent download limits meant to curb AI-music spam. Source

8. Apple Reportedly in Talks to License News for Siri

Apple. Apple is reportedly in talks to pay news publishers nine-figure sums to license current news content for Siri, according to a new report. The approach would ground Siri’s answers in fresh, licensed information rather than open-web scraping, a contrast to how some rivals source news. The negotiations are early and unconfirmed, but they hint at how Apple plans to make Siri more useful for timely queries. Source

9. Researchers Say Several Predicted “Automated AI Research” Milestones Have Already Fallen

Research. A researcher at the Institute for AI Policy and Strategy interviewed 25 frontier-lab scientists about recursive self-improvement and catalogued the milestones they predicted on the path to automated AI research, several of which have already been reached. The findings suggest progress toward systems that accelerate their own development is tracking ahead of some insiders’ expectations. For practitioners, it is a signal to watch how much of the research loop labs are beginning to automate. Source

10. Fable 5’s Slow Uptake Suggests a Ceiling on Frontier AI Spend

Industry. New figures reported by The Decoder indicate that Anthropic’s Fable 5, widely regarded as one of the most capable models available, accounts for only about 6% of the tokens Anthropic sells to US companies. The analysis argues that corporate willingness to pay a premium for frontier capability may be hitting a ceiling, with many workloads served well enough by cheaper models. It is a useful counterpoint to headline benchmark races when planning model spend. Source