Good morning. Anthropic’s Opus 4.8, which we covered yesterday, has had a day to settle and the reviews are coming in — mixed, as expected. Elsewhere: OpenRouter just raised nine figures to keep proxying everyone else’s models, EY Canada got caught publishing a report full of fake citations, and GitHub Copilot users are watching their monthly bills jump by orders of magnitude.
Opus 4.8 lands, community verdict is “fine.” A day after the release, HN’s verdict is roughly aligned with Anthropic’s own framing — a “modest but tangible improvement.” Simon Willison’s pelican-on-bicycle test shows a meaningful gap between low and high thinking levels, and one commenter reported Opus 4.8 in ultracode mode one-shotting a working RTS game in a single HTML file. Early ArtificialAnalysis numbers suggest GPT-5.5 still wins on cost per task, solving problems with about half the output tokens. Several users are also hitting “can’t modify thinking blocks” errors in Claude Code, which is dampening the launch.
OpenRouter raises $113M Series B. CapitalG led, with NVIDIA, Databricks, and Snowflake along for the ride, on the back of growth from 5 to 25 trillion weekly tokens in six months — roughly 41 million tokens per second through a team of fifty. The announcement leans on the routing, failover, and key-management story, which HN commenters who actually use the product seem to back: billing caps, low-friction model switching, and per-key limits came up repeatedly as reasons they pay the proxy tax. The skeptical camp had two points worth airing: the “Open” branding is misleading since nothing is open source, and there’s a real question of whether a proxy business holds up once the frontier consolidates around two or three labs.
EY Canada publishes a cybersecurity report full of hallucinated citations. GPTZero’s investigation found that most URLs in EY’s 2025 loyalty-systems report are broken or fabricated, over half the cited titles don’t match real sources, and several statistics appear invented. The report is credited to two partners and a senior manager, and has already been picked up by downstream news outlets and AI search tools. The most depressing HN comment on the thread: 12 to 18 months ago this would have ended careers, and now it’ll barely register.
GitHub Copilot’s new token billing is not landing well. Copilot moves to token-based pricing on June 1, and TechCrunch collected some eye-watering before/after examples from users — $29/month becoming $750, $50 becoming $3,000. The split in the comments is roughly between developers canceling outright and developers arguing the spikes are concentrated among people using agent loops carelessly. Either way, the era of flat-rate AI coding subsidies looks like it’s ending.
SoftBank commits up to €75B to French data centers. The plan, reported by TechCrunch, targets 5 GW of AI capacity with an initial 3.1 GW phase in Hauts-de-France by 2031 — SoftBank’s largest European AI infrastructure bet to date, and another sign that the Stargate model is being exported. France gets cheap nuclear baseload and a willing host government, which goes a long way to explaining the location.
Meta is building an AI pendant. TechCrunch reports testing within the year, built on the Limitless acquisition from late 2025, alongside an expanded glasses lineup and a “Wearables for Work” subscription to help close Reality Labs’ $4B quarterly hole. HN was unkind. One commenter asked which products Meta has actually built (rather than bought) since Facebook itself, and another pointed to the recent reporting on Ray-Ban Meta footage being reviewed by overseas contractors as a reason no one should want a Meta-branded microphone on their chest.
Gemini Spark, Google’s always-on agent, gets a hands-on. TechCrunch’s reviewer found Spark genuinely useful for inbox triage and finding savings across Gmail, Calendar, and Docs, with the caveat that Google still can’t articulate why personal users specifically need this. The novelty is that it runs in the cloud continuously rather than requiring an open laptop, which makes the productivity-suite tie-in actually work.
AI job grief, fake Black TikTok influencers, and a clean inference repo. A few smaller items worth a look: a post on AI job grief makes the case that displaced knowledge workers experience identity loss in a way the discourse hasn’t quite caught up with, and the HN thread is full of people in exactly that position. The Verge documented AI-generated Black influencers used to dropship Shein products at marked-up prices via manufactured emotional appeals. And on the more cheerful end, Tiny-vLLM is a C++/CUDA inference engine with a lesson-style README that commenters are comparing favorably to early llama.cpp as a learning resource.
That’s the brief. Worth watching today: whether Copilot’s pricing backlash forces a walk-back before the June 1 cutover, and whether the Opus 4.8 thinking-block bug gets patched before the week is out.