Gawk · 2026-05-22
Gawk — 2026-05-22 · 4 tool incidents
Five verifiable things that moved in the AI ecosystem in the last 24h. Every number traces to a public source.
What moved
- · claude-opus-4-6-thinking holds #1 on LMArena for the 30th consecutive snapshot.
- · ai downloads grew for the 8th consecutive snapshot.
- · openai downloads grew for the 8th consecutive snapshot.
Benchmark movers
#No rank changes in the LMArena top 3
Why this matters · Public benchmarks are gameable, but rank shuffles still hint at where the frontier is genuinely moving.
#1 claude-opus-4-6-thinking
anthropic · 1500 Elo
#2 claude-opus-4-6
anthropic · 1498 Elo
#3 gemini-3.5-flash
google · 1486 Elo
Source: lmarena.ai
Tool Health
#4 incidents in the last 24h (vs 2 yesterday)
Why this matters · Provider outages and degradations cause retry storms upstream. Tracking the 7-day shape catches flapping providers before they page you.
Elevated errors on Claude Opus 4.7
started 10:17 UTC · resolved 10:31 UTC · minor impact
Elevated error rate on multiple models
started 04:16 UTC · resolved 08:50 UTC · major impact
Elevated latency and error rates for ChatGPT 5.5 Thinking
started 22:05 UTC · resolved 23:22 UTC · minor impact
Elevated error rates on ChatGPT paid plans
started 13:52 UTC · resolved 15:23 UTC · minor impact
claude-code: Operational → Degraded
Source: status.anthropic.com, status.openai.com, status.claude-code.com
Top HN stories
#Top 5 on HN in the last 24h
Why this matters · What developers debate on Hacker News often previews which models, frameworks, and patterns they'll actually adopt next.
We should get rid of average CPU utilization
24 points · 19 comments
Multi-Stream LLMs: new paper on parallelizing/separating prompts, thinking, I/O
22 points · 1 comments
CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs
15 points · 0 comments
Show HN: Spec-Driven Development Workflow for Claude Code
15 points · 2 comments
Show HN: Let agents run any analysis with Mixpanel data, no UI required
14 points · 1 comments
Source: news.ycombinator.com
SDK adoption
#6 notable shifts across the six registries
Why this matters · Package install volume is where developers place real bets — distinct from where labs are shipping marketing.
PyPI: anthropic
−123.0k 24h downloads day-over-day
pypistats.org is a third-party aggregator — counts include mirror hits, CI builds, and pip install retries.
npm: @anthropic-ai/sdk
−32.7k 24h downloads day-over-day
crates.io: ort
+34.8k 90d downloads day-over-day
Docker Hub: ollama/ollama
+329.4k all-time pulls day-over-day
Homebrew: ollama
−546 30d downloads day-over-day
VS Code Marketplace: GitHub.copilot
+15.2k cumulative installs day-over-day
Source: pypistats.org, npmjs.com, crates.io, hub.docker.com, formulae.brew.sh, marketplace.visualstudio.com
Agent frameworks
#2 agent frameworks moved >10% in the last week
Why this matters · Agent frameworks rise and fall on weekly download velocity and open-issue trajectory, not GitHub stars.
CrewAI · +47%
4.2M weekly downloads · 52k stars
Pydantic AI · -29%
8.3M weekly downloads · 17k stars
pypistats.org · View on Gawk →
Tracks the `pydantic-ai` meta-package; the lean `pydantic-ai-slim` variant is roughly 4× larger but users typically compare the meta number.
Source: pypistats.org, pypistats.org
Notable AI Lab activity
#8 labs moved in the last 24h
Why this matters · GitHub event volume on a lab's own repos is the cleanest publicly-verifiable proxy for engineering activity we have.
LangChain
−24 events (now 172) · San Francisco, US
Hugging Face
+16 events (now 122) · New York, US
Weights & Biases
−8 events (now 72) · San Francisco, US
DeepSeek
+16 events (now 67) · Hangzhou, CN
New mover: AMD
55 events · Santa Clara, US
New mover: Intel AI
27 events · Santa Clara, US
Dropped off: Google Research
was 72 events · Mountain View, US
Dropped off: Replicate
was 52 events · San Francisco, US
Source: gawk.dev
Model Usage
#deepseek/deepseek-v4-flash holds #1 · biggest mover: openai/gpt-5-mini +12 to #30 · biggest drop: qwen/qwen3.5-397b-a17b -12
Why this matters · OpenRouter rankings reflect API-first developer spend; direct customers like consumer ChatGPT are invisible by construction.
openai/gpt-5-mini climbed +12 to #30
was #42 on 2026-05-15
OpenRouter request volume reflects developer API spending, not end-user adoption. Biased toward API-first workflows that route through OpenRouter — direct OpenAI / Anthropic / Google customers who never use OpenRouter are invisible.
qwen/qwen3.5-397b-a17b slipped -12 to #43
was #31 on 2026-05-15
#1 deepseek/deepseek-v4-flash
#2 tencent/hy3-preview
#3 anthropic/claude-sonnet-4.6
Source: openrouter.ai