TLDR: Multiple users are reporting a noticeable drop in Claude Opus 4.6’s reasoning quality since early April, and Anthropic quietly reduced its thinking budget from 100 to 25. Meanwhile, Grok 4.20 is claiming the top spot on BridgeBench, beating GPT-5.4 and Opus 4.6 across reasoning and hallucination metrics. Vercel open-sourced a reference platform for cloud coding agents, and Anthropic’s unreleased Mythos model is getting restricted access through Project Glasswing due to its ability to autonomously discover zero-day vulnerabilities.
<hr>
<h2>Claude Opus 4.6 thinking budget cut by 67%</h2>
Several users noticed Claude Opus 4.6’s planning depth and code quality dropped in the first week of April. The explanation: Anthropic quietly reduced the model’s thinking budget from 100 to 25, a 67% cut, affecting both $20 and $200/month subscribers. API users were apparently unaffected. The timing is bad — Opus 4.6 is already hitting rate limits hard enough that some users are seriously considering switching to Codex Pro, and GitHub Copilot just retired the Opus 4.6 Fast model for Pro+ users, citing infrastructure strain from high concurrency.
<h2>Grok 4.20 leads BridgeBench</h2>
xAI’s Grok 4.20 is claiming the number one spot on BridgeBench, reportedly beating GPT-5.4, Claude Opus 4.6, Gemini, and Qwen across speed, reasoning, and hallucination control. Take benchmark claims with the usual skepticism — BridgeBench isn’t ARC-AGI — but the signal is real enough that it’s worth watching. Grok’s real-time X integration remains a genuine differentiator that none of the other frontier models have matched.
<h2>Anthropic’s Mythos model: restricted access only</h2>
Anthropic released Claude Mythos Preview, described as a tier above Opus, but it won’t be publicly available. Access is restricted to a consortium including AWS, Apple, Google, and Microsoft through Project Glasswing, backed by $100M in usage credits. The reason: during internal red-teaming, Mythos autonomously discovered and exploited zero-day vulnerabilities across major operating systems and browsers, including a 27-year-old OpenBSD bug. SWE-bench verified scores hit 93.9%, up from 80.8% for Opus 4.6. Bruce Schneier notes a security firm replicated similar findings using older public models, which suggests the threat isn’t waiting for Mythos to ship.
<h2>Vercel open-sources a reference platform for coding agents</h2>
Vercel open-sourced Open Agents, a reference platform for building cloud-based coding agents, pointing to the trend of companies like Stripe, Ramp, Spotify, and Block building internal “AI software factories.” The argument is that off-the-shelf coding agents don’t handle large monorepos or institutional knowledge well, and the competitive moat is shifting from the code you wrote to the means of producing it. The stack includes Fluid for agent runtime, Workflow for long-running durability, Sandbox for secure code execution, and AI Gateway for multi-model support. Worth looking at even if you’re not hosting on Vercel.
<hr>
<h2>Also notable</h2>
<ul>
<li>OpenCode 2.0 data model breakthrough: The opencode team announced a significant redesign of the data model for version 2.0, teasing that it will be highly embeddable and easy to build on top of.</li>
<li>Cursor agent splitting: Cursor 3 now lets you split agents for multi-tasking directly in the editor, similar to terminal splitting — a small but genuinely useful workflow improvement for anyone running parallel tasks.</li>
<li>OpenCode Go capacity crunch: OpenCode Go is growing faster than its GPU providers can supply capacity, with the team scrambling to secure more compute; expect some instability.</li>
<li>GitHub Copilot data residency: Copilot now offers US and EU data residency with FedRAMP Moderate authorization, supporting GPT-5.4, Claude Sonnet 4.6, and Claude Opus 4.6 at launch (Gemini excluded due to GCP limitations), with a 10% model multiplier increase.</li>
<li>GitHub Copilot merge conflict fix: A new “Fix with Copilot” button on <a href="http://github.com" target="_blank" rel="noopener nofollow">github.com</a> resolves merge conflicts in three clicks using the cloud agent, which also handles failing Actions workflows and code review comments.</li>
<li>Copilot remote CLI sessions: GitHub launched <code>copilot --remote</code> in public preview, letting developers monitor and steer a running Copilot CLI session from a browser or GitHub Mobile via QR code.</li>
<li>Microsoft MEMENTO research: Microsoft Research found that when reasoning models compress and delete their chain-of-thought tokens, the deleted content persists in KV cache hidden states and keeps influencing downstream computation — a property of transformer attention, not a learned behavior.</li>
<li>OctoGent open-sourced: An experimental Claude Code orchestration dashboard that spins up 4–10 specialized Claude Code instances per codebase section, with inter-agent communication, a graph canvas view, and token usage heatmaps.</li>
<li>Sandcastle 0.4.1: Added support for OpenCode, Pi, and Codex agents plus Podman, Daytona, and Vercel environments, positioning as a simple sandboxed runner for any agent.</li>
<li>Cloudflare Sandboxes GA: Cloudflare’s agent sandboxes are now generally available with persistent Python/JS interpreter contexts, PTY support, credential injection via egress proxy, and pricing shifted to active CPU usage only — Figma uses it for Figma Make.</li>
<li>Linux kernel AI policy: Linux 7.0 ships with an official policy requiring human accountability via Developer Certificate of Origin, prohibiting AI agents from adding Signed-off-by tags, and mandating an “Assisted-by” disclosure tag for AI-assisted contributions.</li>
<li>Stanford “mirage effect”: Researchers removed images from six major vision benchmarks and found GPT-5, Gemini, and Claude still scored 70–80% accuracy, suggesting models are pattern-matching on text context rather than actually processing images.</li>
<li>Google I/O leak: Google appears to be testing an agent feature called “Agent” for Gemini Enterprise with a Tasks UI including goal, connected apps, files, and a “require human review” toggle suggesting desktop task automation is coming.</li>
</ul>

AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Claude Opus 4.6's thinking budget was quietly cut 67% (from 100 to 25), causing noticeable drops in reasoning quality for subscribers. xAI's Grok 4.20 now leads BridgeBench over GPT-5.4 and Opus 4.6. Anthropic's unreleased Mythos model — capable of autonomously discovering zero-day vulnerabilities and scoring 93.9% on SWE-bench — is restricted to a consortium of AWS, Apple, Google, and Microsoft via Project Glasswing. Vercel open-sourced Open Agents, a reference platform for cloud-based coding agents. Additional updates include Cursor 3 agent splitting, GitHub Copilot data residency and merge conflict fixes, a Microsoft MEMENTO research finding on KV cache persistence, Cloudflare Sandboxes GA, and a Stanford study showing frontier models score 70–80% on vision benchmarks even without images.

Claude Opus 4.6 gets quietly nerfed, Grok 4.20 tops BridgeBench

Regarding LLMs and benchmark pattern matching:
We finally have it in writing. what many of us suspected is true. These models are tuned specifically for these tests.
With all these AGI and SWE-bench tests, the models are not really so smart that they fully understand and solve them. They often look good because the teams behind them tuned them for the benchmarks.
It is like studying the answers before a school test.
Most of us suspected it. Now we know.
And this matters for the new ARC-AGI-3 too. If current LLMs are scoring around 0.2% - 0.3%, and then a provider releases a new model that suddenly gets 50% and calls it a “leap in intelligence,” do not believe it. The model did not just 200x previous intelligence, they were just tuned for the next benchmark.

I tested Opencode Go for a month. But tbh, not a single model is really in the same league with GPT/Claude.
I test a bit more since the recent add of Mimo V2 PRO &amp; Qwen 3.6 Plus (The qwen family rarely deceive).
I still think it is an interesting offer. And nicely complement $20 subscriptions. Just dont expect to do complex SWE with it (and I include non trivial web app in this category).

Is there any official announcement about Opus 4.6’s thinking budget change? Is there enough pushback that they are going to change it back? I love Claude Code but this really is making me think about switching to Codex as my daily driver…