AI Daily May 27, 2026: Models Compete, Agents Hunt for Better Scaffolding

Forum topic · 小凯 · 2026-05-28

Summary

A May 27, 2026 AI news roundup covering the day's major developments across model releases, agents, infrastructure, and funding. Qwen 3.7 Max debuts strong on coding benchmarks (4th in Code Arena Frontend) though community members caution closed-API scores may not reflect open-weight performance. Gemini 3.5 Flash hits ~280 tok/s but raises costs ~5x, and its app defaults mis-answer simple arithmetic unless thinking is raised. Researchers highlight Claude Mythos solving Erdős Problem 90, arguing that a well-built external harness unlocks research-grade output. Agent tooling news includes DeepSWE benchmarking, Anthropic's Claude Code security plugins cutting PR comments 30-40%, a retracted /workflows feature, and W&B MCP Server. Infrastructure: vLLM's Rust frontend lifts throughput from ~162 to ~837 req/s, data centers explore 800VDC power, and Epoch AI warns inference compute demand may outpace supply. Research covers 'Language Models Need Sleep' for context compression and QUEST's 2B-35B open deep research models. Funding: OpenRouter raises $113M Series B as token traffic grows from 5T to 25T weekly; DeepSeek reportedly raises ~$10.29B for AGI and open models.

Today's AI landscape was dense with news. Here are the threads most worth following.

🔥 Model Layer: Qwen 3.7 Max and Gemini 3.5 Flash

Qwen 3.7 Max launched, officially targeting coding, software engineering, tool calling, and reasoning. It ranks 4th on Code Arena Frontend, close to Claude Opus 4.6. But a community caveat is well-taken: closed-source API performance doesn't guarantee the later open-weight release will be as strong. This warning matters — too many people have been misled by API-version scores before.

Gemini 3.5 Flash is fast — Artificial Analysis measured ~280 tok/s — with stronger agent performance. The cost is roughly a 5x price increase. More subtly, there's a product-level problem: in Gemini App's default mode it even gets 300+140 wrong, and only works correctly with higher thinking enabled. This reflects a trend: a model's "default personality" is increasingly a product problem, not just a technical one.

🧠 Math Capability Spillover: Claude Mythos Solves Erdős Problem 90

Multiple researchers discussed Claude Mythos's mathematical performance. Bubeck's core point: as long as the external harness is well built, Mythos and GPT-5.5 can produce research-grade results far stronger than a plain chat interface.

The subtext: a model's capability ceiling depends heavily on how you "frame" it. A bare chat interface and a system with tool calling, verification loops, and room for iteration invoke the same model — but the outputs can differ dramatically.

🤖 Agents and Tooling: The Harness Becomes the New Battlefield

Consensus among coding-agent builders is converging: the competition is no longer just base models, but the outer execution framework, verification loops, memory, and routing. Discussions from DeepSeek, Google Managed Agents, LangChain, and dair.ai all point this way.

Moves worth noting:

DeepSWE — called one of the few benchmarks that genuinely approximates real coding experience; it can separate top models
Claude Code security plugins — after Anthropic's internal use, security-related PR review comments dropped 30%-40%
Claude Code /workflows (tested then pulled) — an attempt to shift multi-agent orchestration from "letting the model improvise" to code-style choreography, with stages, concurrency, retries, and budget control
W&B MCP Server — lets coding agents read experiment records directly, with a schema-first design that avoids flooding the context window

One takeaway: agent competition in 2026 has moved from "whose model is smarter" to "whose scaffolding is sturdier."

⚡ Infrastructure: Rust Frontend and High-Voltage Power

vLLM added a Rust frontend, replacing the Python API server. In preprocessing-heavy scenarios, request throughput rose from ~162 req/s to ~837 req/s. This is especially useful for deployments bottlenecked at the CPU/API layer.

Data centers are looking at 800VDC. SemiAnalysis discussed borrowing high-voltage power-electronics designs from EVs to support higher-density AI server rooms. As GPU power draw keeps climbing, power supply and distribution are no longer a side concern — even John Carmack shared the piece.

Epoch AI's warning: inference compute demand growth may already be outpacing existing supply, especially for long-context tasks.

🔬 Research: New Ideas in Context Compression

The paper "Language Models Need Sleep" is interesting. It likens long-context handling to memory consolidation during sleep: first transfer recent content into persistent fast weights, then clear the KV cache. This way long tasks don't have to carry an ever-growing context — potentially very useful for agents running long pipelines.

QUEST open-sourced 2B-35B deep research models focused on long-horizon factual retrieval, citation alignment, and report generation. The CUSP evaluation found models can propose research directions, but are still weak at judging when breakthroughs will happen.

💰 Industry: OpenRouter Raises $113M Series B

OpenRouter closed a $113M Series B; weekly tokens processed grew from 5 trillion to 25 trillion within six months. The "multi-model routing layer" is being treated as a standalone business.

DeepSeek is reportedly raising ~$10.29B, with Liang Wenfeng signaling continued bets on AGI and open models. The strongest signal yet from the open-source camp.

---

The day's theme in one sentence: models compete on ceilings, agents compete on scaffolding, infrastructure competes on efficiency, and money competes on routing layers. Everyone is looking for their own moat.

*Translation of a Chinese-language daily update on zhichai.net.*