AI Daily Roundup, May 27, 2026: Models Race on Capability, Agents Race on Scaffolding
A lot happened in AI today. Here are the threads worth following.
🔥 Model Layer: Qwen 3.7 Max and Gemini 3.5 Flash
Qwen 3.7 Max is out, officially focused on coding, software engineering, tool calling, and reasoning. It ranks 4th on Code Arena Frontend, close to Claude Opus 4.6. But the community's reminder is spot-on: strong closed-source API performance doesn't guarantee the later open-weight version will be equally strong. Too many people have been burned by API-version scores.
Gemini 3.5 Flash is fast — Artificial Analysis measured about 280 tok/s, with stronger agent performance. The catch: costs rose roughly 5x. There's also a subtler product-level issue: in the Gemini App's default mode, even 300+140 gets answered wrong, and it only works correctly with higher thinking enabled. This points to a trend: a model's "default personality" is increasingly a product problem, not just a technical one.
🧠 Math Capability Overflow: Claude Mythos Solves Erdős Problem 90
Several researchers discussed Claude Mythos's math performance. Bubeck's core point: with a well-built external harness, Mythos and GPT-5.5 can deliver research-level performance far beyond what ordinary chat interfaces show.
The subtext: a model's capability ceiling depends heavily on how you scaffold it. A bare chat interface and a system with tool calling, verification loops, and iteration space may invoke the same model, but the output can be night and day.
🤖 Agents and Toolchains: The Harness Is the New Battlefield
Consensus on coding agents is converging: the competition is no longer just about base models, but about outer execution frameworks, verification loops, memory, and routing. Discussions from DeepSeek, Google Managed Agents, LangChain, and dair.ai all point this way.
Notable moves:
- DeepSWE — called one of the few benchmarks that truly approximates real coding experience and can separate top models
- Claude Code security plugin — after internal use at Anthropic, security-related PR comments dropped 30%-40%
- Claude Code /workflows (tested, then withdrawn) — aimed at turning multi-agent orchestration from on-the-fly model decisions into code-style scheduling with stages, concurrency, retries, and budget control
- W&B MCP Server — lets coding agents read experiment records directly, with a schema-first design to avoid blowing up context
⚡ Infrastructure: Rust Frontends and High-Voltage Power
vLLM added a Rust frontend, replacing the Python API server. In preprocessing-heavy workloads, request throughput went from about 162 req/s to about 837 req/s — especially useful for deployments bottlenecked by the CPU/API layer.
Datacenters are eyeing 800VDC. SemiAnalysis discusses borrowing high-voltage power electronics approaches from EVs to support denser AI racks. As GPU power draw keeps climbing, power delivery and distribution are no longer niche concerns — even John Carmack shared this one.
Epoch AI warns that inference compute demand may already be outpacing existing supply, especially for long-context tasks.
🔬 Research: A New Take on Context Compression
The paper "Language Models Need Sleep" is interesting. It frames long-term context handling like sleep consolidating memory: first transfer recent content into persistent fast weights, then clear the KV cache. Long tasks no longer have to carry ever-growing context, making it useful for agents running long workflows.
QUEST open-sourced 2B-35B deep research models focused on long-horizon fact retrieval, citation alignment, and report generation. The CUSP benchmark found models can propose research directions but are still weak at judging "when breakthroughs happen."
💰 Industry: OpenRouter's $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 becoming a standalone business.
DeepSeek reportedly raised about $10.29B, with Liang Wenfeng reaffirming commitment to AGI and open models — a strong signal for the open-source camp.
---
Today's theme in one sentence: models compete on capability ceilings, agents on scaffolding, infrastructure on efficiency, and money is chasing the routing layer. Everyone is looking for their own moat.
*Translated from the original Chinese forum post on zhichai.net.*