# Opus 4.8, Anthropic’s Scale Surge, and the New Contest in AI Infrastructure

*By AI High Signal Digest • May 29, 2026*

Anthropic dominated the day with Claude Opus 4.8 and a huge financing and revenue update, while infrastructure and workflow tools advanced across training, RL, Office integration, and agent deployment. The brief also covers key research breakthroughs, enterprise AI spending, and a notable state-level AI policy move.

## Top Stories

*Why it matters: the day’s biggest signals were frontier model quality, commercial scale, and deeper investment in AI systems infrastructure.*

- **Anthropic launched Claude Opus 4.8.** Anthropic says 4.8 improves judgment, honesty about its own progress, and long-horizon autonomy at the same price as 4.7 [^1]. Third-party tracking put it at #1 on the Artificial Analysis Intelligence Index at 61.4 and #1 on GDPval-AA at 1,890 Elo, with an implied ~67% win rate over GPT-5.5 xhigh [^2]. Anthropic also reported 69.2% on SWE-bench Pro and about 4x fewer unremarked code flaws than 4.7 [^3].

- **Anthropic paired the release with a massive financing update.** The company said it raised $65B in Series H funding at a $965B post-money valuation, and that run-rate revenue crossed $47B earlier this month, driven by Claude deployments in core operations and everyday work [^4][^5]. That makes this both a model launch and a scale signal.

- **SpaceX says it is nearing a custom training stack for very large clusters.** Musk said SpaceX has almost finished a C-based training stack that exact-maps to 220k GB300s with 800G NICs, uses heavy pipeline parallelism, and could deliver more than an order-of-magnitude speed improvement versus JAX on large training runs [^6].

## Research & Innovation

*Why it matters: the most useful technical advances today focused on cheaper RL, faster visual grounding, and stronger search-based reasoning.*

- **Hugging Face cut async RL weight-sync bandwidth by roughly 100x.** The key observation is that about 99% of bf16 weights stay bit-identical between RL steps; HF therefore transfers only sparse deltas. On Qwen3-0.6B, per-step payload fell from 1.2 GB to 20–35 MB, and the team demonstrated fully disaggregated RL over HTTPS and a single Hub bucket [^7].

- **NVIDIA Research released LocateAnything for faster object localization.** The vision-language detector was trained on 138M high-quality samples and decodes bounding boxes in parallel instead of one coordinate at a time, improving localization accuracy and throughput for grounding and detection [^8].

- **Harvard and MIT introduced Bidirectional Evolutionary Search (BES).** BES combines forward search, backward decomposition into checkable sub-goals, and evolution-style recombination. It improved Llama-3.2-3B-Instruct on MuSiQue from 4.0% to 7.0% accuracy and beat other open-source evolutionary frameworks on circle packing and Heilbronn optimization [^9].

## Products & Launches

*Why it matters: agents are moving closer to real work by plugging into codebases, Office apps, and open deployment stacks.*

- **Claude Code added Dynamic Workflows in research preview.** Claude can now write an orchestration script on the fly, spin up hundreds of coordinated subagents, and verify results before returning them. Anthropic says this is for tasks like large migrations and service-wide investigations [^10][^11][^12].

- **Perplexity Computer is now inside Microsoft Office.** It is available in Excel, Word, PowerPoint, and Outlook, where users can draft documents, model, build decks, and handle email from a side panel. Perplexity says the product uses its enterprise security layer, including SAML SSO, audit logs, and granular admin controls [^13][^14].

- **StepFun released Step 3.7 Flash as an open-weight agent model.** The Apache 2.0 release is a 198B sparse MoE with ~11B active parameters, 256K context, 400 TPS, tool-use support, and benchmark wins including #1 on ClawEval-1.1 and SimpleVQA Search [^15].

## Industry Moves

*Why it matters: enterprise AI spending is concentrating around context, infrastructure, and specialized new labs.*

- **Glean crossed $300M ARR.** The company said it reached the milestone five months after $200M ARR and argued that enterprise AI’s moat is a strong context layer grounded in a company’s workflows, permissions, and systems. It said more than 85% of customers use Glean across five or more job functions [^16].

- **Hyperscaler AI capex remains on the same steep curve.** Epoch said Q1 2026 spending came in at $156.1B, close to its $155.1B trendline, keeping projections of $770B in 2026 and more than $1T in 2027 intact [^17][^18].

- **Inherent launched with a $50M seed round.** The new London-based Public Benefit Corporation says it is building AI agents that discover new knowledge and is explicitly organizing around recursive self-improvement of the research organization [^19].

## Policy & Regulation

*Why it matters: state-level AI governance is starting to turn into concrete compliance requirements.*

- A cited report said **Illinois SB315 passed 110-0**, with provisions for third-party audits, transparency reports, risk frameworks, and whistleblower protections for frontier labs; the same report said OpenAI endorsed it [^20].

## Quick Takes

*Why it matters: several smaller updates sharpened the picture on deployment, voice, and agent commerce.*

- OpenAI shipped a new **GPT-5.5 instant** with improvements to sycophancy, factuality, and multilingual performance [^21].
- **Cartesia Ink-2** topped the new streaming STT benchmark for final accuracy at 3.59% WER and 0.21s latency [^22].
- **Elicit** launched an MCP server so agents can search **138M+ papers** and run full research reports inside Claude, ChatGPT, Copilot, Gemini, and other MCP tools [^23].
- Google expanded **Universal Commerce Protocol** toward hotels, food ordering, and YouTube in the U.S. [^24].

---

### Sources

[^1]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2060042702150930686)
[^2]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2060117582120976868)
[^3]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2060043209833951575)
[^4]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2060061347522433422)
[^5]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2060061348818518493)
[^6]: [𝕏 post by @elonmusk](https://x.com/elonmusk/status/2059884150187053488)
[^7]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2059989047947260203)
[^8]: [𝕏 post by @NVIDIAAI](https://x.com/NVIDIAAI/status/2060058563544801787)
[^9]: [𝕏 post by @TheTuringPost](https://x.com/TheTuringPost/status/2060194173505155358)
[^10]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2060044853279617150)
[^11]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2060044856114942328)
[^12]: [𝕏 post by @claudeai](https://x.com/claudeai/status/2060042710753382816)
[^13]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2060013442720010598)
[^14]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2060013454761836992)
[^15]: [𝕏 post by @StepFun_ai](https://x.com/StepFun_ai/status/2060149124117475791)
[^16]: [𝕏 post by @jainarvind](https://x.com/jainarvind/status/2060077511212445874)
[^17]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2060076222873526506)
[^18]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2060076235322282321)
[^19]: [𝕏 post by @inherent_labs](https://x.com/inherent_labs/status/2060119235372752924)
[^20]: [𝕏 post by @thursdai_pod](https://x.com/thursdai_pod/status/2060176911704457649)
[^21]: [𝕏 post by @michpokrass](https://x.com/michpokrass/status/2060219759682330970)
[^22]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2060021901234458958)
[^23]: [𝕏 post by @elicitorg](https://x.com/elicitorg/status/2059999270078726627)
[^24]: [𝕏 post by @Google](https://x.com/Google/status/2060044912171549083)