Usage became visible in four different units. GitHub converts Copilot agent work into token-priced credits; Meta's reported pendant would collect ambient context; Gemini Embedding 2 puts four modalities into one retrieval space; and Cassandra spends pruned weights and cache entries to draft reasoning tokens on edge hardware.
1. GitHub Copilot moves agent usage to token-based credits
Starting June 1, Copilot plans will include GitHub AI Credits consumed by input, output, and cached tokens at each model's listed API rate. Base seat prices remain unchanged, code completions and Next Edit stay included, and paid users or organizations can buy additional usage after their monthly allotment.
Long autonomous sessions will no longer look equivalent to a quick chat in billing. Preview bills, pooled budgets, model-and-task attribution, and hard ceilings can expose runaway agents before overage. Extreme screenshots are anecdotal, but they correctly reveal how request counts hid differences in cached, input, and generated tokens.
Sources: GitHub's usage-billing announcement · TechCrunch on developer reaction
2. Meta reportedly plans to test an AI pendant
Meta is reportedly developing an AI pendant for testing within the next year, building on its 2025 acquisition of Limitless. A memo viewed by The Information also described more AI glasses and a Wearables for Work subscription, but Meta has not publicly announced the pendant or a launch schedule.
An ambient device that records conversations makes bystander consent and retention product-defining. A credible prototype would expose unmistakable recording state, local deletion, narrow defaults, enterprise administration, and a workflow whose benefit exceeds a phone or existing glasses. Otherwise the pendant adds a microphone and another data path without distinct utility.
Sources: TechCrunch on Meta's reported pendant · The Information's cited memo report
3. Gemini Embedding 2 places four modalities in one retrieval space
Google researchers describe Gemini Embedding 2 as a native model for interleaved text, images, audio, and video in one representation space. The paper reports 62.9 Recall@1 on MSCOCO, 68.8 NDCG@10 on VATEX, 69.9 on multilingual MTEB, and 84.0 on MTEB Code.
The authors report task averages that can hide corpus-specific failures. One index could simplify cross-modal search and recommendation, but corpus tests can isolate language coverage, segment length, cross-modal relevance, latency, drift, and permission filtering so a shared vector space does not flatten source access controls.
Sources: Gemini Embedding 2 research paper, version 1
4. Cassandra drafts reasoning tokens with a pruned copy of the same model
The Cassandra preprint proposes self-speculative decoding for AI reasoning models that prunes and truncates selected weights and key-value cache entries to draft tokens, then verifies candidates at full precision. The authors report up to 2.41-times speedup over BF16 and 1.81 times more tokens under the same memory budget than Eagle-3 on Llama 3 8B.
Cassandra also assumes a proposed encoder-decoder hardware module, so the peak result is an algorithm-hardware co-design rather than software ready for current devices. End-to-end power, latency, draft acceptance, memory, and output equivalence on implemented hardware would connect the simulated advantage to an edge deployment.