Summary
The morning slot is two stories and a lot of landfill. Claude Opus 5 shipped in a five-post thread from @ClaudeDevs plus same-hour confirmations from Cursor and AWS, and the substance is in the platform plumbing rather than the benchmark card: effort now defaults to high with a dial in both directions, fast mode runs about 2.5x faster at 2x base price, tools can be added or removed mid-conversation without invalidating the prompt cache, and classifier-blocked requests get routed to a recommended model instead of failing. Cursor published the number that matters, 66.7 versus Fable 5's 66.5 on CursorBench 3.2 at $3.91 against $8.77 per task, and buried in the same table is Grok 4.5 High also at 66.7 for $1.51. The louder story is Jensen Huang joining X to post NVIDIA's open-weights letter, which pulled amplification from a Google DeepMind researcher, tinygrad, and xAI staff within hours, and turned into a pile-on over OpenAI declining to sign while Sam Altman posted supportively. Underneath both sits the real hardware signal, a cluster of four NVIDIA posts on the Korea buildout, President Lee Jae Myung meeting Huang ahead of the San Francisco AI Summit, and a KAIST joint AI research lab in Seoul, which is the social surface of the $500B SK Group partnership. One repost is worth pulling directly: Nick Dorsey citing SemiAnalysis to argue most of AMD's Advancing AI roadmap was NVIDIA's first, which lands the same day SemiAnalysis published its full AMD analysis. Everything else is volume. Two accounts alone contributed 34 posts of French cultural politics and Los Angeles homelessness fights, and the Starship Flight 13 celebration cluster carries no AI content beyond the space-datacenter hobbyhorse.
Posts
Claude Opus 5 ships in Claude Code and the Claude Platform (cluster of 5: @ClaudeDevs · docs). The thread's four substantive claims, in order of how much they matter to anyone running long agent sessions. Prompt caching survives tool changes: you can add or remove tools mid-conversation without invalidating the cache, which is the specific prefix-invalidation event that has forced full prefill recomputes in agent harnesses. Fallback routing: requests blocked by the safety classifier are routed to a recommended model based on the request rather than returning an error. Effort defaults to high in both Claude Code and the Platform, dialable down for faster cheaper turns or up for hard problems, with a fast mode at roughly 2.5x speed for 2x base price. Migration is a skill, not a doc:
/claude-api migratein Claude Code updates model strings and suggests prompt changes tuned for Opus 5. The first two are optimizations the research literature normally treats as things you build around a model, now inside the API. See Claude Opus 5, KV cache, LLM routing.Cursor publishes the cost-per-task table, and it says more than Anthropic's pitch does (cluster of 2: @cursor_ai · CursorBench 3.2). Cursor reports Opus 5 matching Fable 5 at default effort, 66.7 against 66.5, and notes Opus 5 works with Zero Data Retention where Fable does not. The full leaderboard is the interesting artifact: Fable 5 Max leads at 70.5% for $17.32 per task, Opus 5 Max is 0.5 points behind at $8.23, Opus 5 High hits 66.7% for $3.91, and Grok 4.5 High ties that exact 66.7% for $1.51. Three models within 0.2 points across a 5.8x price spread. CursorBench 3.2 scores agents on ambiguous multi-file tasks pulled from real Cursor sessions, so it is closer to production shape than a static coding benchmark, and the crowding it shows is precisely the pool diversity that makes routing worth doing. See LLM routing.
Anthropic's own researcher concedes the qualitative gap (@_sholtodouglas). "Half the cost, pareto mogging intelligence. I think fable still has more sparks of genius, but they work very well together." Worth noting because it is the vendor-side admission that the win is on the cost-per-task frontier, not on raw capability, and because "they work very well together" is a routing argument coming from inside Anthropic.
Opus 5 lands on Amazon Bedrock same day (@mattsgarman · AWS blog). AWS's CEO frames it as built for coding and long-running agents, with zero data retention and zero operator access, and notes the Claude Platform is also available on AWS for customers who want Anthropic's native experience rather than Bedrock's.
Jensen Huang joins X with NVIDIA's open-weights letter (cluster of 6: @JensenHuang · @nvidia · @hexiang · @__tinygrad__ · @stepango · @brivael · letter PDF). His first post ever is a signed letter arguing that AI will be built by every country, that open models strengthen safety and cybersecurity while enabling sovereignty, and that the world needs both frontier closed and frontier open models. The lines being quoted back are sharper than the framing: "closed models are single points of failure," "relying solely on closed models is not inherently safe," and a defense of distillation as reflecting "a long tradition of learning from, building upon, and improving existing technologies," with a call to avoid premature restrictions. The amplification pattern is the actual signal. A Google DeepMind researcher posted "Be open," tinygrad framed it as the company that made modern AI possible choosing to empower others and predicted the unaligned companies would find themselves isolated, an xAI engineer welcomed him, and Elon Musk endorsed it outright. That distillation sentence lands the same day the White House accused a Chinese lab of exactly that practice. See the open-weights letter.
OpenAI's non-signature becomes the story (cluster of 4: @ns123abc). Four posts hammer the same point, that the company with "Open" in its name did not sign and has lobbied against open weights since DeepSeek R1, against Sam Altman's post that "i want the US to win in AI both in open source and proprietary models, and i am glad to see this." The framing is partisan and the account is an aggregator, but the underlying fact is checkable and the coalition split it describes is real: every company selling compute or a cloud platform signed, and the two labs with the strongest frontier positions did not.
NVIDIA's Korea buildout gets a four-post rollout (cluster of 4: @nvidia · @nvidia on KAIST). "This is the Golden Age of Korea," partnering across government, industry, and technology on the K-AI vision spanning chips, frontier AI, physical AI and robotics, and AI factories. President Lee Jae Myung met Huang ahead of the San Francisco AI Summit, and KAIST launched what NVIDIA calls the first NVIDIA joint AI research lab in Asia, focused on agentic AI for Korean industry. The corporate-relations gloss hides the substance, which surfaced in the afternoon slot and in The Information the same day: this is the social face of a $500B partnership with SK Group, owner of SK hynix, including joint next-generation HBM development and a 2-gigawatt Vera Rubin DSX factory. Memory supply is the binding constraint on inference economics, so the HBM co-development line is the part to track, not the ribbon-cuttings. See NVIDIA and SK, memory hierarchy.
AMD accused of shipping NVIDIA's roadmap late, on the day SemiAnalysis published its AMD deep dive (@ns123abc). A repost of investor Nick Dorsey, who calls the AMD Advancing AI event exciting and then argues most of the showcased roadmap "was put forward by someone else a little while ago," citing SemiAnalysis, with a copied-homework list long enough to need a second tweet to clear the character limit. Investor-flavored and one-sided, but the underlying source published in full today, and its actual verdict is more interesting than the dunk: AMD's silicon is genuinely ahead on integration (first 2nm datacenter part, 12 HBM4 stacks for 432GB against Rubin's 288GB), while its microarchitecture converges on Hopper and its distributed inference stack has no wide expert parallelism anywhere on the new silicon. See Can AMD break the CUDA moat.
xAI signals a two-week model cadence (cluster of 2: @ns123abc · @brivael). Elon Musk says Grok 4.6 ships in two weeks and Grok 4.7 in four, two weeks after Grok 4.5 went live. Both reposts frame it as acceleration. The practical consequence is unglamorous: any static model-selection configuration is stale within a month, which is an argument for routing by measured per-task performance rather than by pinned model name.
Grok Build reportedly has no agent-spawning limit (@ns123abc). "I kept spawning 10s of agents to test the limit, Grok 4.5 has no limits." An anecdote rather than a measurement, but it is the second consecutive day of Grok Build parallel-agent claims after Workflows launched, and unbounded fan-out is a real cost-and-safety surface rather than a feature.
Grok Imagine on the craft of prompting a candid performance (cluster of 4: @imagine · @heavypulp). Bloopers from an Odyssey dialogue scene, with a genuinely useful observation: you cannot prompt "she laughs" and get anything natural. Their working prompt describes the mechanics instead, "the first laugh already escapes through her nose, one loud snort she claps a hand over, the hand loses: the laugh bursts around it, a half-second of pure stunned silence, and then she is gone completely." They also note that writing a reaction in heavy detail makes the generated performance more distinct. Small but concrete evidence that video-model controllability is currently a specification problem rather than a capability ceiling.
Starship Flight 13 succeeded (cluster of 4: @ns123abc · @stepango). All 33 Raptor engines fired clean, hot-staging worked, the first Starlink V3 batch deployed with contact on all 20 satellites, the booster splashed down in the Gulf, and the ship survived reentry to soft-land in the Indian Ocean despite losing heat shield tiles on ascent. Two adjacent one-liners from the same account, "SpaceX 🤝 NVIDIA" and "SpaceX 🤝 Thinking Machines," are context-free teasers with nothing behind them in this slot. Relevant here only through the orbital-datacenter thread these accounts keep pushing.
tinygrad on Radeon and self-driving (@__tinygrad__). "These Radeon are practicing driving. Trying to beat FSD." One line, no data, but it is the same account running consumer-AMD training experiments that the wiki has tracked before, and it sits alongside their open-weights endorsement in the same slot.
Opaque and off-topic volume. @brivael contributed 18 posts, almost all French domestic politics, plus one substantive AI take (that building frontier models in Europe is structurally impossible on culture, capital, and market grounds, in response to criticism of Mistral abandoning large language models for B2B specialization) and a promo for his own video product Argil v2. @spencerpratt contributed 16 posts of Los Angeles homelessness and Hollywood merger politics. @HouseGOP, @AustinJustice, @WHFraudTF, @DoWCTO, @lynnmartin, @magicsilicon, and @JonasBadalic round out the noise with congressional messaging, crime statistics, a NYSE anniversary party, and a one-word Starship joke. Skip.
No curated reposts this slot. @bayesiansapien posted no retweets in the 24-hour lookback, so every item above comes from the AI handle feed rather than the curated layer.