This brief covers AI news from 2026-09-13 UTC.
Payments giant Stripe buys the model router, DeepSeek ships a frontier-adjacent open model, and Washington moves closer to binding AI safety rules.
Stripe acquires OpenRouter for $7.5 billion
What happened: Stripe announced it completed its acquisition of OpenRouter, the AI model routing and gateway platform, in a deal valued at roughly $7.5 billion, up from OpenRouter’s $1.3 billion valuation in its $113 million Series B in May. OpenRouter will keep operating independently under its existing name and roadmap.
Why it matters: Developers routing requests across hundreds of models through one API now sit under a payments company that wants to meter token consumption alongside revenue.
DeepSeek releases V4.1 Flash under MIT license
What happened: DeepSeek released V4.1-Flash, built on a new causal encoder-decoder architecture with 8B active parameters for prefill and 16B for decode, native vision, a 1M-token context window, and KV cache cut to roughly 1/8 of its predecessor. It retires the V4 Pro line and ships under an MIT license.
Why it matters: MIT terms mean commercial use, redistribution, and fine-tuning with no revenue thresholds, so teams with sensitive data can self-host instead of calling a third-party API.
Google closes Mechanize talent deal for over $1.5B
What happened: Google completed a talent acquisition of AI startup Mechanize for more than $1.5 billion, moving over a dozen employees including co-founder Tamay Besiroglu into DeepMind to work on midtraining. Former chief of staff Guive Assadi remains at Mechanize as CEO. Google and Besiroglu declined to comment.
Why it matters: Midtraining, the phase between pretraining and fine-tuning where models learn to reason and use tools, is where Google is placing its bet on closing coding gaps.
Senate negotiators draft binding AI safety duty of care
What happened: Senate Majority Leader John Thune, Commerce Committee Chairman Ted Cruz, and Sen. Amy Klobuchar are drafting legislation that would impose a legal duty of care on developers of the most powerful AI models and grant the government authority to block unsafe releases, Reuters reported. No public draft exists yet.
Why it matters: A federal preemption clause under discussion could override state AI laws, so builders selling frontier models face one federal standard instead of a patchwork.
Salesforce in talks to buy Listen Labs for about $2B
What happened: Salesforce is negotiating to acquire Listen Labs, a three-year-old voice AI startup that runs automated customer interviews, for roughly $2 billion, near 67 times its $30 million annualized revenue. The talks are preliminary and could still fall apart.
Why it matters: The reported target feeds motivational context into Salesforce’s Agentforce agents, a signal that agent differentiation is being bought through data rather than model weights.
Agnes AI ships 33B multimodal model under Apache 2.0
What happened: Agnes AI released Agnes-3.0-Flash, a 33B open-weights multimodal model under Apache License 2.0 with a 262, 144-token context window and text, image, and video support. Of 72 decoder layers, 54 use a gated delta-rule recurrent design and 18 use global attention, so only 18 layers hold KV cache.
Why it matters: The bf16 checkpoint runs on a single H200 or H100, giving smaller teams a self-hostable multimodal option without a cluster.
Fujitsu to export Fugaku-derived AI chips next year
What happened: Fujitsu will begin exporting AI semiconductors developed with technology used in its Fugaku supercomputer to the U.S. and Asia as early as next year, targeting growing demand for CPUs for inference.
Why it matters: Another inference CPU supplier entering the export market gives buyers an alternative as server lead times stretch and prices climb.