Stripe buys OpenRouter to bank the token economy

Stripe's reported $7B-plus swoop on OpenRouter is the day's big money move, signaling that the plumbing of the "token economy" is now prime M&A real estate. Independent reviews of last week's Qwen 3.8 27B confirm near-frontier local performance shackled to a comically overthinking default, while cheap Chinese open-weight models keep forcing Western price cuts. And a sobering feature on what happens when the startup behind a child's AI robot goes bust.

Stripe to buy OpenRouter for $7B+, betting on the token economy

Bloomberg reports Stripe has agreed to acquire OpenRouter, the model-gateway startup that routes requests across 400-plus models for 8 million users, for more than $7 billion. That is roughly a 5x markup on the $1.3B valuation OpenRouter set in its $113M Series B in May, whose backers include Sequoia, a16z, Menlo and Alphabet's CapitalG. Stripe declined to comment; the deal puts the payments giant between apps and every model API, metering AI usage the way it meters card transactions.

Why it matters: OpenRouter is the default abstraction layer many developers use to avoid provider lock-in. Owning it hands Stripe a chokepoint on multi-model traffic and billing.

Qwen 3.8 27B reviewed: Sonnet-class, but the default reasoning setting is unhinged

Independent testing of the Apache-2 Qwen 3.8 27B lands, and the consensus is it is remarkably capable for a 17GB file: Simon Willison got his best-ever local pelican SVG, accurate vision bounding boxes, and drove a coding agent with it, while others put it near Claude Sonnet (occasionally Opus) on faithful arcade-game clones once given a good harness. The catch is a shipped default of xhigh reasoning, burning 20k-plus tokens and up to 20 minutes to draw a circle; reviewers uniformly recommend dropping to low or medium. Community work on the model's built-in Multi-Token Prediction is boosting throughput around 70%, reaching 82 tok/s single-request on an RTX 3090.

Why it matters: A genuinely useful frontier-adjacent model now fits on a laptop, but the out-of-box default is a trap. Set reasoning to low or medium unless you enjoy watching it philosophize about a circle.

Chinese models undercut US labs ~9x, and the price war keeps cutting

OpenAI cut GPT-5.6 Luna API pricing 80% (to $0.20/$1.20 per million input/output tokens) and Anthropic pitched Claude Opus 5 at roughly half its prior flagship's cost, both responding to Chinese open-weight models from DeepSeek, Moonshot's Kimi and Zhipu's GLM. One benchmark puts an equivalent job at $544 on GLM versus $4,811 on Claude, a near-ninefold gap finance teams are now spreadsheeting. Bloomberg reports the cheap models are pushing US players to rethink strategy, even as Booz Allen and others warn Chinese models generate less secure code, fueling a corporate fight over savings versus data safety and shadow AI.

Why it matters: For a large share of everyday enterprise workloads the capability gap has narrowed enough that price, not quality, is the deciding factor. The frontier labs are pricing accordingly.

When the AI-companion startup folds, the kid's robot dies

MIT Technology Review traces Moxie, the $800 AI robot marketed as a social-skills companion for neurodivergent children, through two corporate collapses that bricked the cloud-dependent device. When maker Embodied shut down in 2024, an engineer shipped OpenMoxie, open-source firmware to keep the robots running locally, but many families could not migrate before the servers went dark; a second owner then folded in 2025. The piece is a case study in the planned obsolescence of emotionally-bonded, always-online consumer AI hardware, and the thin clinical evidence behind therapeutic robots.

Why it matters: Any product that offloads its brain to a startup's servers inherits that startup's runway. 'The company folded' is now a failure mode for a child's best friend.

Amodei defends his policy agenda: open weights won't decentralize power

Anthropic CEO Dario Amodei defended his policy proposals, endorsing pre-launch model vetting and arguing that open weights will not decentralize power the way advocates claim, while saying real accomplishments (not marketing) will earn public trust. In a separate quote he conceded AI's trust problem is genuine and self-inflicted: 'the most accurate criticism is that we haven't yet delivered on our big promises to benefit the world... The thing that will work is actually curing cancer.'

Why it matters: Anthropic's regulatory line, favoring vetting and control over open release, directly shapes what open-weight developers may be allowed to ship next.

MathCode wires a coding agent to a Lean 4 proof engine

MathCode, a terminal coding assistant, takes a plain-language math problem, formalizes it into a Lean 4 theorem and attempts an agentic proof. It is backed by a persistent Lean language server (compile checks near 0.4s after warmup, versus ~30s cold), an auto-named reusable theorem and axiom library, Mathlib lemma search via leansearch and Loogle, parallel subgoal decomposition, and an Obsidian dependency graph. It runs on macOS/Linux with the codex CLI as the default backend.

Why it matters: Formal-proof scaffolding with a fast persistent REPL is exactly what turns LLM math from plausible-looking to machine-verified.

Also worth a look

Browse previous days →