3 min read · 14 sources

Meta shipped a new model, Google shipped another one, and 15 robotaxis showed up on roads that flunk most human drivers. Two frontier labs, one medieval city, and a quiet antitrust win round out the day.

Gemini 3.8 Flash’s $0.75/M input tokens intro rate undercuts most competitors while matching larger models on coding benchmarks.

Meta puts Muse Spark 1.3 behind the API counter

Source: bloomberg.com ↗

Meta has released Muse Spark 1.3, and this time the bill arrives before the weights. Developers can pay for access now, with the model rolling into Meta’s social surfaces shortly after. Meta hasn’t committed to releasing 1.3 weights; it still plans to drop weights for 1.2.

For engineers, the practical question is whether the API is the product or the weights are. If 1.3 stays closed and 1.2 ships openly, the open release becomes the regression test bench while the hosted tier carries the actual capability gap.

Google's third Flash in six weeks, with a Cyber variant bolted on

Source: arstechnica.com ↗

Gemini 3.8 Flash is out in two flavours: a workhorse Flash for agentic and coding work, and Flash Cyber, tuned for vulnerability detection. Cyber is gated to trusted testers and governments for now. Through the end of the year, input is $0.75/M tokens and output is $3.75/M; that doubles to $1.50 and $7.50 in January.

Google’s own benchmarks show modest gains over 3.7 Flash on most evals and bigger jumps on coding. The interesting part is the cadence: three Flash releases in roughly six weeks suggests Google is treating pricing as a tunable knob too, not just capability. If you budget on the intro rate, plan for the January step before you commit.

Chemogenetics hits its first human trials in China

Source: neuroai.science ↗

Bryan Roth disclosed at the BRAIN Initiative conference that seven chemogenetics trials are running in China, the first human use of the technique. Chemogenetics modifies neurons to respond to inert designer drugs, so a patient can be switched on or off remotely and at genetic resolution. That’s a different shape from TMS or small molecules: it’s irreversible, gene-therapy-adjacent, and targets specific cell types.

The implications aren’t abstract. Anyone building BCIs or neuromodulation stacks now has to design alongside a therapy layer that can hard-wire receptor behaviour into tissue. Targets mentioned include epilepsy, Parkinson’s, and pain.

Robotaxis meet London's medieval road geometry

Source: nytimes.com ↗

Fifteen Uber robotaxis built on Wayve’s stack are now available through the Uber app while safety testing runs. The framing in the piece is the licence test: more than half of London’s human applicants fail, so the bar for an autonomous driver in the same streets is the unspoken punchline.

European rollouts have been slow because the streets aren’t a grid and the regulators aren’t permissive. If Wayve’s learned-driving approach holds up in this environment, it weakens the “AVs only work in Phoenix” argument a bit further.

GitHub Copilot's bill: cut tokens, but not at the cost of the task

GitHub’s [post on Copilot efficiency](https://github.blog/ai-and-ml/github-copilot/how-we-make-ai-coding-more-cost-efficient