Summary
A loaded slot, all from the AI account feed. No curated retweets from @bayesiansapien in the past 24h. The dominant signal is the Anthropic launch cluster: Opus 4.8 hits the SWE-bench Pro leaderboard at 69.2% (up from 64.3%), Dynamic Workflows ships in Claude Code as a research preview where Claude writes a JavaScript orchestration script that fans out to up to 1,000 subagents with adversarial verification, and @ns123abc reports the Series H closed at $65B on a $965B valuation with run-rate revenue at $47B. Two adjacent signals: AWS CEO Matt Garman announces Resilient Network Graphs (33% better throughput, 40% less network power) and confirms Opus 4.8 on Bedrock plus a new $6B Snowflake partnership. Kilo Code separately reports running an end-to-end webhook service for $1.65 total using xAI's new Grok Build 0.1 ($1/$2 per 1M input/output, 256K context), the cheapest credible frontier-coding-model build to date. Cursor confirms Opus 4.8 is live in their product and notes input tokens have crossed output tokens as the dominant cost. Smaller items: kipply (Anthropic) crowdsourcing examples of Claude thinking too much or too little, NVIDIA's CFO recognition, a Tesla FSD anecdote, and a Frontier Tower demo day amplification from Scoble. Non-AI noise from @brivael, @spencerpratt, @NICKIMINAJ, and @WHFraudTF in the slot.
Posts
Claude Opus 4.8 + Dynamic Workflows launch cluster (@ClaudeDevs thread, Opus 4.8 blog via AWS, Dynamic Workflows blog, Claude Code docs). Anthropic shipped two product surfaces on 2026-05-28. Opus 4.8 is a same-price replacement for 4.7 with three explicit improvements: 69.2% on SWE-bench Pro (up from 64.3%), four-times-lower rate of letting flaws in its own code pass unremarked, and stronger long-horizon task persistence. Opus 4.8 defaults to "high" effort, which spends about the same tokens as 4.7's default on coding while performing better; for difficult tasks and async work, "xhigh" mode is available, and Anthropic raised Claude Code rate limits to cover the increased token usage. Pair it with
/goalin Claude Code or with Outcomes on Claude Managed Agents for long-horizon work. Dynamic Workflows is a Claude Code feature where Claude writes a JavaScript orchestration script on the fly that fans work out across tens to hundreds of parallel subagents, runs adversarial verification, and returns only the verified result. Trigger by including "workflow" in a prompt or setting/effort ultracode. Available on Max, Team, Enterprise, and via API (including Bedrock, Vertex AI, Foundry). Background execution, resumable from cached results. The first bundled workflow,/deep-research, is explicitly designed to disprove its own findings. The same/claude-api migrateskill updates project model strings to 4.8. See today's digest deep dive for the synthesis and the ai-industry summary page for the connections to the May routing cluster.Anthropic Series H at $65B, $965B valuation, $47B run-rate (@ns123abc, Les Echos via @brivael repost). Two reposts independently surface the same scoop: Anthropic closed a $65B Series H at a $965B post-money valuation, with run-rate revenue crossing $47B earlier in May. The Les Echos piece (in French, paywalled) frames this as Anthropic "creating an unprecedented gap" with OpenAI and approaching the $1T valuation mark. Simon Willison's RSS post (also in today's digest) tracks the run-rate ramp: $14B on 2026-02-12, $30B on 2026-04-06, $47B on 2026-05-07. The growth curve is faster than any company has scaled organic revenue at this level, per Axios. Cross-check Gary Marcus's "tokenmaxxing" critique in today's Industry Pulse; the $47B figure rests on enterprise stickiness that has not been independently audited.
AWS Opus 4.8 on Bedrock + Resilient Network Graphs + $6B Snowflake deal (@mattsgarman Opus 4.8, AWS blog Opus 4.8, @mattsgarman Snowflake). Three same-day posts from the AWS CEO. First, Opus 4.8 is live on Amazon Bedrock and on the Claude Platform on AWS, framed as the model that "sustains work across long, complex tasks without losing context or going off track." The AWS blog adds practical guidance: token efficiency claims on agentic workflows, and operational guarantees around regional data residency and enterprise security. Second, the broader Resilient Network Graphs announcement is the substrate move covered in today's digest as a Deep Dive: flat random-graph data-center fabric, 33% better throughput, 40% less network power. Third, a new $6B multi-year Snowflake deal built on Graviton compute and AI services, with Snowflake passing $7B in lifetime AWS Marketplace sales. The combined message is "AWS is the AI substrate." See the hardware summary page for the fabric breakdown.
Kilo Code: $1.65 webhook service with Grok Build 0.1 (@kilocode thread, article). The most concrete cost-of-coding-agent data point this week. xAI released Grok Build 0.1 at $1 per 1M input tokens and $2 per 1M output tokens with a 256K context window (versus GPT-5.5 at $5/$30 and Claude Opus 4.7 at $5/$25). Kilo Code ran it in two phases inside the Kilo Code VS Code extension: $0.17 to plan (Plan Mode), $1.48 to implement (Code Mode), $1.65 total for a working TypeScript + Bun + SQLite webhook delivery service with Drizzle migrations, AES-GCM encrypted secrets, SSRF guard, non-destructive replay, and a Stripe-style signature format. Throughput was roughly 120 tokens/sec with zero tool-calling failures across the run. The model opened with a web search (pulling Stripe's signature format, GitHub's retry behavior, the Standard Webhooks spec) and asked 9 clarifying questions before writing a line. Reviewer notes: the
get-by-idendpoint leaks the encrypted secret, signature comparison is not constant-time, integration coverage is thin. The take-away is not that Grok Build 0.1 ships production code; it is that the cost of a competent first pass on a backend service has fallen to the price of a coffee. Kilo Gateway already supports it with BYOK across 20 providers and 24 plans.Cursor: Opus 4.8 is live, input tokens dominate cost (@cursor_ai 4.8 launch, @cursor_ai cost note). Two posts. First, Opus 4.8 went live in Cursor within hours of the Anthropic announcement; CursorBench shows it working "much more efficiently than Opus 4.7" and being "more persistent on harder tasks." Second, a finding from the Cursor Developer Habits Report Spring 2026 (yesterday's evening slot, surfaced again): as agents consume more context, input tokens have become the majority of price-equivalent token costs. Combined with the Cursor report's other findings (cost-per-accepted-line varies 7x across model families, P99 power users dominate AI activity, mega-PRs growing), the financial structure of agentic coding is shifting in ways the prevailing per-token-output pricing model does not capture cleanly.
kipply (Anthropic) crowdsourcing Claude thinking calibration (@_sholtodouglas reposting @kipperrii). Sholto Douglas amplifies a one-liner ask from kipply: do users like how much Claude thinks on their tasks, and would they share examples of it thinking too much or too little. Combined with today's Opus 4.8 launch language ("more honesty about its own progress"), this is a public-facing telemetry channel into the same calibration improvement work that produced the four-times-fewer-bug-omissions number. Operational signal: Anthropic is investing in pre-launch user research on thinking-budget calibration for the 4.x line.
@bcherny on dynamic workflows for big migrations (thread, blog). Boris Cherny, creator of Claude Code, reframes Dynamic Workflows as the answer to a specific workflow problem: big migrations and refactors are some of a team's most important work and the easiest to push off, because they would tie up engineers for a quarter. With Dynamic Workflows, Claude can land that kind of work in days or weeks instead. Cherny advises defaulting to auto mode so Claude does not stop for permissions and to save Dynamic Workflows for "biggest jobs: migrations, refactors, perf optimization, batch bug fixes" given the token cost.
NVIDIA GTC Taipei 2026 + Colette Kress recognition (@nvidia Kress, Tesla FSD anecdote). NVIDIA's Colette Kress (CFO) recognized on Fortune Magazine's Most Powerful Women in Business for the third year. Operational signal only. The previously announced GTC Taipei keynote remains scheduled for 2026-06-01 at 11am Taipei time, focused on physical AI, scaling infrastructure, and science. Tesla posted an FSD anecdote about predicting bad-driving merge behavior from millions of recorded merges; no technical claim attached.
@Scobleizer: Frontier Tower demo day, Pika MCP, shift NYC (demo day, Pika MCP, shift NYC). Robert Scoble surfaces three signals from SF demo days. (a) Five international startups demoing at Frontier Tower; no startup names captured in the article-content fetch. (b) Pika labs launched a "Founder Starter Kit" of 4 skills (Build-a-Brand, App Screens, Product Sizzle, Founder Video) available via the Pika MCP for Claude. (c) shift in NYC offers free home cleanings in exchange for collecting cleaning data to train robots; the article-content body explains the data is anonymized before training but no protocol details. The robotics data-collection-as-a-service model is the operational analog to the Genesis World 1.0 simulation stack covered in yesterday's morning slot: train robots faster on simulation OR train them on harvested real-world data, both commercial approaches now visible.
Skip cluster (non-AI noise). @brivael ran a long thread on French political opinion, Mistral coverage, and an Apple iOS 27 third-party-AI-extensions repost (the MacRumors article itself is genuinely AI-relevant and is captured in today's digest Industry Pulse). @spencerpratt ran LA mayoral campaign content. @NICKIMINAJ and @WHFraudTF posts have no AI substance. The remaining items are personal commentary or political reposts and add nothing to the wiki.