Anthropic's unannounced test shrank Claude Code's 'high' effort to 'low'

Anthropic's Claude Code team spent the weekend responding to users who noticed "high" effort quietly behaving like the old "low" — an unannounced test, confirmed on Hacker News, with credits promised to anyone who reports a real regression. A security researcher's fingerprinting tool linked the mystery Ox Alpha coding model to Zhipu AI, a lab under U.S. sanctions, the same day its free OpenRouter access is set to end and its early record benchmark score got walked back to ordinary. The Model Context Protocol's maintainers published their first roadmap since March, prioritizing agent identity over more tool-calling primitives. OpenAI cut GPT-5.6 Sol's API pricing by up to a third, then spent the weekend chasing Codex quotas that drained too fast and reset every paid subscription's usage. And a 27-billion-parameter open model reverse-engineered a commercial license check in 30 minutes, running entirely offline.

What you're actually running

Anthropic's unannounced test shrank Claude Code's 'high' effort to 'low'

Claude Code 2.1.237 quietly remapped what "high" effort means for Claude Fable 5 sessions: the model now reads "high" as the numeric value 10, exactly what "low" used to mean, with nothing said in the changelog. Developer argofowl found it after an afternoon debugging his own app before suspecting Claude. Thariq Shihipar of the Claude Code team confirmed it on Hacker News: Anthropic is running a new API serving-config test, live now, that changes how the effort number maps. "The scale isn't 0-100, the number isn't meaningful on its own, and the effort you selected is the effort you're getting," he wrote, adding that internal evals found no performance change. Commenters weren't reassured — reports of 40-plus-minute waits on simple tasks and unusually verbose output piled up under the thread. If Fable 5 in Claude Code feels off this week, check your version and file /feedback with the session ID; Anthropic says it will credit confirmed regressions.

Ox Alpha's tokenizer fingerprint links it to sanctioned Zhipu AI

Security researcher unclecode fingerprinted Ox Alpha, the anonymous free coding model on OpenRouter. Its tokenizer counts match Zhipu AI's GLM-5.3 on all four tests; every other lab matched two. He calls that shared infrastructure, not proof of identity. The U.S. sanctioned Zhipu in 2025. Ben Davis's early 80% score on DeepSWE, a coding benchmark, put it ahead of GPT-5.6 Sol and Fable 5; a full run corrected that to 63%. Free OpenRouter access ends today.

Nobody knows who built AI coding model Ox Alpha or where the code goes - SiliconANGLE

Nobody knows who built AI coding model Ox Alpha or where the code goes - SiliconANGLE

SiliconANGLEDuncan Riley

Protocol and pricing moves

MCP's new roadmap prioritizes agent identity over more tool-calling features

The Model Context Protocol's maintainers published their first roadmap update since March. The headline shift is agentic messaging: server-initiated events, Tasks maturing into a first-class primitive instead of an extension, and unified HTTP-native transport, so a remote MCP server behaves like any other web workload. Their own words: "a remote MCP server is now no different from any other HTTP workload, making it easy to host and operate one on any infrastructure." If you're building or wiring up MCP servers, that's the direction to build toward.

The New MCP Roadmap

An update on the Model Context Protocol roadmap and focus areas for upcoming specification releases.

Model Context Protocol BlogDavid Soria Parra (Lead Maintainer) and Den Delimarsky (Lead Maintainer)

OpenAI cuts GPT-5.6 Sol prices up to 33% through November

OpenAI dropped GPT-5.6 Sol's API pricing Friday. Input tokens fall from $5 to $4 per million, a 20% cut. Output falls from $30 to $20, a 33% cut. The lower rate applies to the pay-as-you-go API, Codex credits, and eligible ChatGPT Work plans, not to Pro, Plus, or Business subscriptions, and holds through at least November 21. If your agent calls Sol directly, the bill just got smaller with no code changes needed.

OpenAI resets every paid Codex quota after finding three usage-draining bugs

Codex quotas were draining faster than they should all week, and the OpenAI engineer who posts Codex's rate-limit updates spent the weekend saying so. Thibault Sottiaux named three causes: image inefficiencies in long sessions, a power-hungry Computer History feature, and a conversation-title generator quietly eating quota. Fixes shipped with a full usage reset for every paid subscription, propagated by Sunday night. His thread drew 3.4 million views. If your week's quota vanished mid-task, check your account; it starts fresh.

Open and local

Qwen 3.8 27B cracks a license check in 30 minutes, offline

An XDA Developers writer handed Qwen 3.8 27B a commercial app's license-verification binary, running the model locally with no cloud calls. In 30 minutes it disassembled the ARM64 binary, recovered a deliberately obscured RSA public key, caught its own mistake mid-task, and produced a working bypass proof-of-concept. No API key, no GPU cluster, no execution of the app itself. Six months ago this was frontier-only work; now it runs on the machine next to you.

I gave Qwen 3.8 27B a reverse-engineering job I assumed needed a frontier model, and it finished in 30 minutes

Qwen 3.8 27B genuinely shocked me with what it achieved here.

XDAAdam Conway

Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.

Also worth your time

• LLMs are changing which languages and problems developers pick up at all — Armin Ronacher argues language choice matters less now that an agent can rewrite in another language on request, which opens previously gatekept domains (eBPF, DWARF, crypto) to more developers, for better and worse.
• A week of running Codex more than Claude Code, logged in detail — Claude built more abstractions and Sorbet signatures than asked for; Codex stayed literal and did less, including creating a branch pointed at another branch that produced a 4,000-plus-line PR after a bad rebase.
• A developer wrapped Claude Code, Copilot, and Grok into "clones" that work while you don't — Munder Difflin runs locally and messages between teammates' agent clones to handle reviews and docs. Early and unverified — engagement, not endorsement; nobody outside the builder has reported running it yet.
• Building an almost-fully self-hosted, sandboxed agentic software factory — one engineer's write-up of wiring coding agents into an isolated pipeline end to end, including where it still needed a human.
• Shopify's CEO implemented Cursor's "Git at Scale" design over the weekend — Tobi Lütke called Cursor's systems post one of the most interesting he'd read in a while and shipped walgit, a working open-source Rust implementation, 742 stars in under a day.

The New Way is human-curated — a person picks every story. The summaries are written with AI (Claude) and reviewed before we hit send.