Stripe buys the AI gateway OpenRouter for more than $7 billion
Stripe has agreed to acquire OpenRouter for over $7 billion, according to Bloomberg and TechCrunch. OpenRouter is the single API that around 8 million developers use to reach 400-plus models from OpenAI, Anthropic, Google, Meta and DeepSeek, picking one per request on price and capability; it reportedly handles around 25 trillion tokens a week — roughly 1.3 quadrillion annualized. The price is about 5.4× the $1.3 billion valuation OpenRouter raised at in May, and roughly 30% below the ~$10 billion the Wall Street Journal floated in July.
It puts a payments company in charge of the layer that both routes and bills AI traffic — the exact point where a request is matched to a model and metered. For a European buyer that layer is also a control point: whoever owns the gateway sees the prompt, decides which provider and jurisdiction handle it, and now settles the invoice too. The deal doesn't change which model answers a given call, but it concentrates who sits between a company and every provider it uses.
GLM-5.3 tops the open coding charts — but Z.ai is holding the weights back
Z.ai shipped GLM-5.3 on 14 August, and the notable part is what it didn't ship: the open weights. The model reuses the same 743B-parameter mixture-of-experts base as GLM-5.2, so every reported gain comes from scaled-up post-training rather than a new architecture. Z.ai's own numbers show Terminal-Bench 3.0 climbing from 4.6 to 28.3, and CyberGym vulnerability-discovery at 84.5%, which it places just ahead of Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%) — reported figures, still awaiting independent confirmation.
Rather than repeat GLM-5.2's MIT-licensed, weights-in-days release, Z.ai says it will stage the open weights and wider API access only after a cyber-safety review, roughly two weeks out — around the end of the month. A Chinese open-weight leader voluntarily gating its own release on cyber-capability concerns is a first, and the trigger is precisely the CyberGym score it is advertising. For any team that treats “open weights” as a sovereignty guarantee, it is the recurring lesson: the artifact isn't yours until it's on the repo, whatever the launch post promises.
Google's Gemini 3.7 Flash posts big coding gains at a price that expires
Google released Gemini 3.7 Flash on 13 August, three weeks after 3.6 Flash, and the coding scores moved a long way in that window: FrontierCode 1.1 Main from 34.4% to 43.6%, DeepSWE v1.1 from 49% to 65.3%, and a Code Arena Elo of 1588 — ahead of Claude Sonnet 5 (1541) and GPT-5.6 Terra (1523). It is aimed squarely at coding, agents and enterprise document processing.
The pricing is the part to put in a calendar. Through 31 December the model runs at $0.75 per million input tokens and $3.75 output; on 1 January 2027 those list prices double to $1.50 and $7.50. That is the same shape as Anthropic's Sonnet 5 cliff (promo pricing ends 31 August) and DeepSeek's weekend hike: an introductory rate with an expiry date attached. Budget on the January number, not the one on the pricing page today.
Quick Hits
- DeepSeek's price hike is live — At 16:00 UTC on 16 August, DeepSeek moved V4-Flash and V4-Pro to peak/off-peak billing; V4-Pro output went from a $0.87 flat rate to $3.96 per million at peak, half that off-peak, per Quartz. The cheap-China-API discount just narrowed sharply.
- Claude went dark for about 42 minutes — On the evening of 16 August (21:58–22:40 UTC), authentication failures cascaded into degraded performance across claude.ai, Claude Code and Cowork; the Claude API and Console stayed up, and Anthropic hasn't disclosed a cause (BleepingComputer).
- Anthropic nudged its own risk rating up — Its 14 August risk report raised the odds of catastrophic harm from misalignment from “very low” to “low,” citing added uncertainty from this summer's cyber-eval incidents rather than any failed test, and disclosed a shelved internal “Model 2” it says is more capable than Mythos 5.