<p><strong>TLDR:</strong> Alibaba’s Qwen 3.7 Max is drawing serious attention as a frontier-level model that matches or beats Claude Opus 4.7 on agentic benchmarks at roughly half the input cost and a quarter of the output cost. Meanwhile, OpenAI shipped a cluster of Codex improvements — side conversations, diff summaries, token stats, computer use on Windows, and self-managing threads. The enterprise AI spending problem got sharper: a mystery company burned $500M on Claude in a single month, Amazon shut down its internal tokenmaxxing leaderboard, and GitHub published a playbook for cutting agent token costs by up to 62%. Anthropic’s $36B compute debt deal and the broader AI bubble math are also worth watching.</p>
<hr>
<h2>Qwen 3.7 Max arrives as a serious Claude alternative on agentic benchmarks</h2>
<p>Alibaba’s Qwen 3.7 Max scores within striking distance of Claude Opus 4.7 on the Artificial Analysis Intelligence Index while costing roughly half the input price and a quarter of the output price. On agentic benchmarks — Terminal Bench 2.0, SWE-bench Pro, MCP Atlas — it’s competitive or ahead. The model ran autonomously for 35 hours on a GPU kernel optimization task, making 1,158 tool calls to achieve a 10x speedup on Alibaba’s custom chip. It natively supports the Anthropic API spec, making it a drop-in replacement in most agent frameworks. Caveats: high token verbosity (roughly 4x median output volume) and a 48% attempt rate on factual questions due to conservative abstention. The broader signal is that Chinese labs are now serious competitors on both performance and price, not just budget alternatives. <a href="https://app.daily.dev/feed-by-ids?id=Nn1Lv8Sqx&amp;id=9uPFbfYuX&amp;id=8sFeTKBt7&amp;id=Yfhd4lQwq" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Codex ships a cluster of UX and capability updates</h2>
<p>OpenAI pushed several Codex improvements in quick succession: side conversations let you ask questions without breaking the main thread, end-of-turn diff summaries show what changed after each step, a new Profile tab surfaces detailed token consumption stats, and self-managing threads mean Codex can now create and search its own sessions. Computer use and remote control from ChatGPT landed on Windows. The desktop app may have dropped Electron in favor of OpenAI’s own OWL web layer, based on hints in the UI. One developer’s 24-hour comparison found Codex still wins as a daily driver for UX and auto-testing consistency, while Claude Opus 4.8 wins on raw planning quality for hard problems — the recommended workflow is to use both. <a href="https://app.daily.dev/feed-by-ids?id=8NE4O9Qwl&amp;id=8mh8jFcT9&amp;id=gEmiWOQZZ&amp;id=0QjSiCyvw&amp;id=qIP5Jo5fo&amp;id=OQe72dqZD&amp;id=xCGAQqq5T" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Enterprise AI cost spiral: $500M in a month, leaderboards shut down, GitHub cuts 62%</h2>
<p>An unnamed large corporation reportedly spent $500M on Claude in a single month after failing to set usage limits on employee licenses — agentic tools consuming up to 1000x more tokens than standard queries are cited as a factor. Amazon shut down Kirorank, its internal Kiro usage leaderboard, after employees spun up agents to perform pointless tasks just to boost scores; Meta faced the same problem with its Claudeconomics leaderboard in April. On the other side of the ledger, GitHub published a concrete playbook: pruning unused MCP tool schemas, replacing MCP calls with gh CLI commands, and running daily Auditor and Optimizer agents cut token spend by up to 62% in production CI workflows. Engineers are reportedly getting bonuses for this kind of cost work. <a href="https://app.daily.dev/feed-by-ids?id=maWMLDVwV&amp;id=RiC9EzkdX&amp;id=MZxEGEsce&amp;id=63Sft4CV0&amp;id=DfFkBTj8Q" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Apollo and Blackstone arrange $36B compute debt deal for Anthropic</h2>
<p>Apollo Global Management and Blackstone are structuring a roughly $36 billion debt deal through a special-purpose vehicle to purchase Google TPUs and lease them to Anthropic, keeping the hardware debt off Anthropic’s balance sheet. Broadcom is providing a residual-value support agreement on about $31 billion of the senior debt, effectively backstopping chip resale value if Anthropic defaults. This is one of the largest private-credit transactions ever built around a single company’s compute needs. A separate financial analysis argues the math behind frontier AI valuations is increasingly strained: OpenAI posted a negative 122% operating margin in Q1 2026, and both OpenAI and Anthropic have made compute commitments that require revenue growth to levels that look hard to hit. <a href="https://app.daily.dev/feed-by-ids?id=SdTxTb8ja&amp;id=Rka5HMph7" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Salesforce: 231-day migration shipped in 13 days with Claude Code:</strong> Salesforce published a detailed writeup on agentic development with Claude Code. A migration scoped at 231 days completed in 13, and a single PR delivered 21 endpoints at 100% test coverage. Separately, Benioff confirmed Salesforce made zero engineering or service agent hires in FY2026 while growing sales headcount 20%. <a href="https://app.daily.dev/feed-by-ids?id=zzUavEwor&amp;id=Xj1BFKSgz" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Qwen 3.7 and open models: 1 in 3 AI teams now running open-weights:</strong> LangSmith Signal data shows 1 in 3 AI teams ran an open-weights model in their workflows recently. Qwen 3.7 Max’s price-performance ratio is accelerating this trend, with multiple developers reporting it as a viable drop-in for Claude in agentic pipelines. <a href="https://app.daily.dev/feed-by-ids?id=Aiy1vCUla&amp;id=bcXQK3NHs&amp;id=RwzuzWvWW" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cursor adds auto-review mode for fewer approval prompts:</strong> Cursor shipped auto-review mode, letting agents execute tool calls with fewer interruptions while maintaining execution guardrails. Reduces friction in agentic workflows without going fully hands-off. <a href="https://app.daily.dev/posts/C015GyFgv" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cloudflare supports Claude Managed Agents:</strong> Cloudflare now lets developers run and manage Claude agents directly within its platform, with connections to private systems, selectable runtime environments, and native monitoring tooling. <a href="https://app.daily.dev/posts/wxja636yF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>xAI releases grok-build-0.1 on API in public beta:</strong> xAI made grok-build-0.1, the model behind the Grok Build CLI, available via API at $1/M input and $2/M output tokens. Early impressions highlight strong agentic capabilities and multi-tool calling for interface design tasks. <a href="https://app.daily.dev/feed-by-ids?id=yZ2l0sf7P&amp;id=GGpDTcV4j" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Flathub bans AI-generated app submissions:</strong> Flathub updated its policy to explicitly ban LLM usage in both the app submission process and the apps themselves, driven by a surge in low-effort AI-generated submissions. Already-published vibe-coded apps won’t be removed retroactively. <a href="https://app.daily.dev/posts/d0Vhmytfe" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>SQLite adds <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> rejecting agentic code contributions:</strong> SQLite added an <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> explicitly refusing AI agent code contributions, accepting only agentic bug reports that include reproducible test cases. The policy was recently hardened by removing the word ‘currently’ from the rejection statement. <a href="https://app.daily.dev/posts/fAioh3cWv" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ChatGPT share links abused to host fake outage malware pages:</strong> Attackers are using <a href="http://chatgpt.com/s/" target="_blank" rel="noopener nofollow">chatgpt.com/s/</a> share links to render convincing fake OpenAI outage pages via Google Ads, then redirecting victims to sites serving macOS and Windows infostealers. Similar abuse has been observed on Claude Artifacts and Grok. <a href="https://app.daily.dev/posts/h2GQ8yzz2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Liquid AI releases LFM2.5-8B-A1B edge MoE model:</strong> Liquid AI’s new on-device MoE model hits 253 tokens/s on an M5 Max laptop under 6GB memory, with a 128K context window and 38T token pretraining. Day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. Available open-weight on Hugging Face. <a href="https://app.daily.dev/posts/N9FwBiQAp" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Kilo Code: lessons from 3M downloads and 40 trillion tokens:</strong> The open-source Claude Code fork shares production lessons: task size should be bounded by reviewability, teams need governance controls before more autonomy, and the product had to evolve from a VS Code sidebar into a full plan→scope→run→verify→review→merge workflow. <a href="https://app.daily.dev/posts/mxPHp8cAR" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub Copilot usage API adds AI adoption phase cohorts:</strong> The Copilot metrics API now classifies each engaged user into one of four adoption phases based on 28-day rolling usage, from code completion through multi-agent workflows. Enables admins to track developer maturity beyond simple active-user counts. <a href="https://app.daily.dev/posts/9ebAWKhOd" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic internally rejected a token leaderboard idea:</strong> An Anthropic insider says someone proposed building a token leaderboard two months ago. After a heated internal debate, the idea was killed — several employees anticipated the tokenmaxxing behavior that Amazon and Meta subsequently experienced firsthand. <a href="https://app.daily.dev/posts/O0zsePUhn" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Gemini’s subscription value questioned after usage cap changes:</strong> Google shifted Gemini from fixed daily requests to compute-based caps with five-hour and weekly limits, and removed 1,000 monthly AI credits from AI Pro. The author concludes other $20/month subscriptions now offer better value, with NotebookLM as the one standout Google AI product. <a href="https://app.daily.dev/posts/xaHPDEmoF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Asana acquires Stack AI for $75M:</strong> Asana bought the Y Combinator-backed no-code AI agent builder to add cross-system workflow execution across Salesforce, Slack, and Google Workspace. Q1 revenue was $205.1M, up 9.5% YoY, but the stock is down over 53% since January. <a href="https://app.daily.dev/posts/MD8LFMbEz" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>JetBrains used AlphaEvolve to cut IntelliJ indexing time 4.6%:</strong> JetBrains ran Google DeepMind’s AlphaEvolve on its B-tree implementation, generating algorithm variants across 50+ iteration sessions. One candidate reduced end-to-end indexing from 17.4s to 16.6s in full IDE integration tests on a mature, heavily hand-tuned codebase. <a href="https://app.daily.dev/posts/TO4ZNWYRn" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linus Torvalds pushes back on ‘99% of code is AI’ claims:</strong> At Open Source Summit, Torvalds compared AI to compilers — useful tools, not replacements for understanding. He noted a ~20% increase in Linux kernel submissions tied to AI use and raised concerns about AI-generated drive-by bug reports burning out smaller project maintainers. <a href="https://app.daily.dev/posts/Oupb9F5aS" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Writing ADRs for agents improves their codebase awareness:</strong> Architecture Decision Records capture non-obvious decisions that code alone can’t express, making every agent touching the codebase more informed. Described as the thinnest documentation layer with the highest signal-to-noise ratio for agent context. <a href="https://app.daily.dev/posts/h1ybOX32o" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>DNS-AID proposed for agent discovery via standard DNS:</strong> The Linux Foundation introduced DNS-AID, a DNS extension letting AI agents and MCP servers discover and verify each other using existing DNS infrastructure rather than proprietary registries. Initially developed by Infoblox with Deutsche Telekom and Amazon contributions. <a href="https://app.daily.dev/posts/vKvKMsoL6" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


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

Agentic Digest

A curated roundup of major AI coding and enterprise AI developments: Alibaba's Qwen 3.7 Max emerges as a cost-competitive alternative to Claude on agentic benchmarks (half the input cost, quarter the output cost), while OpenAI shipped multiple Codex UX improvements including side conversations, diff summaries, and Windows computer use. Enterprise AI spending is spiraling — one company burned $500M on Claude in a month — while GitHub published a playbook cutting agent token costs by 62%. Anthropic secured a $36B compute debt deal via Apollo and Blackstone. Notable items include Salesforce completing a 231-day migration in 13 days with Claude Code, Cursor adding auto-review mode, Flathub banning AI-generated app submissions, SQLite rejecting agentic code contributions, and Linus Torvalds pushing back on '99% of code is AI' claims.

Qwen 3.7 Max challenges Claude on price, Codex gets a wave of UX updates