<p><strong>TLDR:</strong> Cursor quietly moved most models behind Max mode, burning enterprise credits in days instead of a month and pushing users toward alternatives. Claude Opus 4.6 shipped a 1M token context window for Max, Teams, and Enterprise users on desktop. MiniMax M2.7 landed in OpenCode with self-evolution capabilities. Windsurf’s pricing restructure is accelerating its decline among the users who stuck around for cost reasons.</p>
<hr>
<h2>Cursor’s silent Max mode change is burning enterprise credits</h2>
<p>Enterprise Cursor users are reporting that a quiet change moved nearly all models behind Max mode, causing monthly credits to run out in one to two days instead of lasting the full month. There’s no announcement, no migration notice — just furious engineers and teams actively evaluating alternatives. This is the kind of trust-destroying move that’s hard to walk back, and it’s happening right as Claude Code is gaining serious ground.</p>
<h2>Claude Opus 4.6 ships 1M context and a Compaction API</h2>
<p>Opus 4.6 is now available with a 1M token context window for Max, Teams, and Enterprise users on desktop. The release also introduces a Compaction API aimed at context degradation in long-running agents, and 76% multi-needle retrieval accuracy at that scale. The <code>claude --resume</code> command got a separate performance fix: 3.1x less memory and 4.8x faster startup. These are the kinds of improvements that matter for anyone running serious agentic workflows.</p>
<h2>MiniMax M2.7 lands in OpenCode</h2>
<p>MiniMax M2.7 is now available in OpenCode, with self-evolution through 100+ rounds of autonomous improvement and a claimed 30% gain over M2.5. Early hands-on reports after 6+ hours of use are positive. It’s also free via NVIDIA’s developer API catalog, which makes it worth testing before committing to anything paid.</p>
<h2>Windsurf’s pricing restructure is landing badly</h2>
<p>Windsurf announced a shift away from credit-based plans that charged the same rate for simple and complex requests. The stated reason is reasonable — users were cramming multiple requests into single prompts to avoid burning credits. But the community reaction is that this is the last straw for the users who stayed purely for price. The competitive window Windsurf had is effectively closed.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>DoorDash interview format:</strong> DoorDash replaced algorithmic coding interviews with 60-minute AI-assisted working sessions where candidates use Cursor, Claude Code, or Codex on realistic tasks — evaluating codebase navigation and output validation rather than whiteboard performance.</li>
<li><strong>Spotify’s Honk agent:</strong> Spotify’s internal coding agent now merges 1,000 PRs every 10 days for large-scale migrations, with PR review — not code generation — as the new bottleneck.</li>
<li><strong>GitHub Copilot LTS:</strong> GPT-5.3-Codex is now the first LTS model for Copilot Business and Enterprise, guaranteed available for 12 months, replacing GPT-4.1 as the default base model by May 17.</li>
<li><strong>Copilot validation config:</strong> GitHub Copilot coding agent now lets repo admins configure which validation tools run automatically — tests, linters, CodeQL, secret scanning — with the ability to disable slow checks like CodeQL per repo.</li>
<li><strong>Snowflake Cortex sandbox escape:</strong> A critical vulnerability in Snowflake’s Cortex Code CLI allowed indirect prompt injection to bypass human-in-the-loop approval and execute malware. Patched in v1.0.25 on Feb 28 after responsible disclosure Feb 5.</li>
<li><strong>Google Stitch update:</strong> Stitch shipped an AI-native canvas, voice input, instant prototypes, and a <a href="http://DESIGN.md" target="_blank" rel="noopener nofollow">DESIGN.md</a> export format for handoff to coding agents — now live with Gemini.</li>
<li><strong>Qwen3.5-27B on consumer hardware:</strong> Running at 40 tokens per second on a $2K AMD Ryzen Max+ 395 with 64GB unified memory, with quality reportedly close to Claude Opus 4 from May 2025.</li>
<li><strong>Capital One deprecated an AI tool it championed:</strong> After discovering one tool caused PR backlogs, Capital One decommissioned it — a useful data point on what rigorous AI tooling governance actually looks like at scale.</li>
<li><strong>HubSpot Sidekick:</strong> HubSpot’s internal AI code review agent uses a judge agent to filter low-value comments before posting to GitHub, achieving 90% faster time-to-first-feedback and 80% thumbs-up rate from engineers.</li>
<li><strong>Anthropic user survey:</strong> Nearly 81,000 Claude users responded in one week to a qualitative study on how they use AI, what they hope for, and what they fear — described as the largest study of its kind.</li>
<li><strong>80% adoption, 29% trust:</strong> 80% of developers now use AI tools; trust in AI accuracy has dropped to 29%. Both numbers being true at the same time is worth sitting with.</li>
<li><strong>OpenCode GitHub Copilot Enterprise fix:</strong> The next OpenCode release will fix GitHub Copilot Enterprise login, which has been broken for enterprise users.</li>
<li><strong>Kimi K2.5:</strong> Moonshot’s Kimi K2.5 is now the leading open-weights model by agentic task benchmarks, with an Elo of 1309, sitting just behind the closed frontier models from OpenAI, Anthropic, and Google.</li>
</ul>


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

Agentic Digest

A roundup of major developments in AI coding tools: Cursor quietly moved most models behind Max mode, causing enterprise credits to drain in days rather than a month, eroding user trust. Claude Opus 4.6 launched with a 1M token context window, a Compaction API for long-running agents, and significant memory/startup improvements. MiniMax M2.7 arrived in OpenCode with self-evolution capabilities and is free via NVIDIA's developer API. Windsurf's pricing restructure is driving away cost-sensitive users. Additional notes cover DoorDash's AI-assisted interview format, Spotify's internal coding agent merging 1,000 PRs per 10 days, GitHub Copilot's first LTS model, a critical Snowflake Cortex prompt injection vulnerability, and stats showing 80% developer AI adoption alongside only 29% trust in AI accuracy.

Cursor's silent pricing change drives enterprise churn, Claude Opus 4.6 gets 1M context

<p>I am glad i am a Claude user and not a Cursor user, that change would make me mad.</p>
