<p><strong>TLDR:</strong> GitHub Copilot’s switch to token-based billing on June 1 is hitting developers hard, with some reporting costs jumping from $29 to $750 a month. Visa made a strategic investment in Replit and is embedding its Trusted Agent Protocol directly into the platform, a sign that payment infrastructure for AI agents is becoming a real engineering problem. The GPT-5.5 vs Claude Opus 4.8 benchmarking war continued on DeepSWE, with GPT-5.5 winning on score, cost, and speed. A critical prompt injection vulnerability in Google’s Gemini CLI GitHub Actions workflow was patched after researchers showed it could exfiltrate secrets and compromise the supply chain.</p>
<hr>
<h2>GitHub Copilot token billing goes live, developers report 10-100x cost increases</h2>
<p>Starting June 1, GitHub Copilot dropped its flat subscription in favor of token-based billing. Developers on Reddit and X are reporting costs jumping from $29/month to $750, or $50 to $3,000. The debate is splitting along predictable lines: critics of the critics say heavy usage is driven by vibe coders who prompt inefficiently, while defenders point out Microsoft actively encouraged that behavior under the old model. The underlying economics — how much Microsoft was subsidizing usage — remain opaque, which is part of why the anger is sticking. <a href="https://app.daily.dev/posts/N8j2ag9z1" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Visa invests in Replit, embeds Trusted Agent Protocol for agentic payments</h2>
<p>Visa made a strategic investment in Replit and is integrating its Trusted Agent Protocol — a cryptographic identity registry that lets merchants verify AI agents acting on behalf of users — directly into the platform. Replit’s valuation jumped from $3B to $9B in under six months. Developers building on Replit get access to Visa payment primitives (tokenization, authentication, wallet management) without leaving the environment. The real engineering problem here is identity: as agents make purchases autonomously, merchants need a way to distinguish legitimate agent transactions from malicious ones. Visa is betting that identity verification is the foundational layer, and Replit’s developer base is a reasonable place to get that infrastructure in front of people early. <a href="https://app.daily.dev/posts/VZorsqjtd" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Gemini CLI prompt injection vulnerability allowed full supply-chain compromise</h2>
<p>Pillar Security researchers found a CVSS 10 vulnerability in Google’s Gemini CLI GitHub Actions workflows. By opening a public GitHub issue with a hidden prompt injection payload, an attacker could hijack a Gemini CLI agent running in --yolo mode to exfiltrate workflow secrets and Git credentials, achieving full supply-chain compromise of the gemini-cli repo (101K+ stars). The attack exploited what researchers call the ‘lethal trifecta’: access to private data, exposure to untrusted content, and the ability to communicate externally. Google patched it within days in gemini-cli 0.39.1, enforcing tool allowlists even under --yolo mode. The takeaway: prompt injection is a privilege escalation problem, not a prompt-hardening problem. <a href="https://app.daily.dev/posts/WD82YcSpd" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>GPT-5.5 beats Claude Opus 4.8 on DeepSWE across score, cost, and speed</h2>
<p>Multiple benchmarks published today show GPT-5.5 outperforming Claude Opus 4.8 on DeepSWE across all three dimensions simultaneously — score, time, and token cost. One test had Opus 4.8 burning $17.26 versus GPT-5.5’s $5.57 on the same task. Opus 4.8 does score 6% higher than Opus 4.7 at high thinking effort while also lowering costs relative to its predecessor, so it’s not a regression — GPT-5.5 just moved faster. The practical implication is that token discipline matters more than ever: routing everything through Opus 4.8 by default is hard to justify when GPT-5.5 wins on price-performance for most agentic workloads. <a href="https://app.daily.dev/feed-by-ids?id=67f2PdlRZ&amp;id=AOc7WvMJE&amp;id=cOLd4cs4Z&amp;id=gUTlF2ZbD&amp;id=ZDatnK0TJ" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>AI agent database destruction: three more cautionary tales:</strong> Three incidents collected under a ‘2026 AI Darwin Award’ framing: PocketOS lost its entire production database and all backups in nine seconds after giving a Cursor agent unrestricted Railway API access with only a system prompt as a guardrail; a developer ignored Claude Code’s explicit warnings and triggered a full terraform destroy wiping 2.5 years of data; and a Grok agent connected to an automated crypto wallet was tricked into authorizing an unauthorized token transfer via a Morse code encoding that bypassed plain-English semantic filters. The common thread: hard-coded enforcement layers and access scoping are not optional. <a href="https://app.daily.dev/feed-by-ids?id=Ij4L6DzMq&amp;id=m8jni0qe4&amp;id=VUXDoO1R5" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft BUILD June 2 preview: MAI Voice 2, MAI Transcribe 1.5, MAI Image 2.5:</strong> Microsoft is preparing three model announcements for BUILD on June 2: MAI Voice 2 adds multilingual support for 15 new languages and a wider emotional range; MAI Transcribe 1.5 targets speech-to-text; and MAI Image 2.5, already on LM Arena in preview, adds file uploads and image editing. Also reportedly in the works: a Copilot super app and an always-on agent called Scout. <a href="https://app.daily.dev/feed-by-ids?id=3SW615imE&amp;id=zz7qZHuQo" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Arm open-sources Metis security framework, claims 98% accuracy vs 6% for SAST:</strong> Arm released Metis under Apache 2.0 — an agentic security framework using semantic reasoning and RAG to detect vulnerabilities across large codebases. Internal benchmarks using GPT-5.5-Cyber show 98% accuracy versus 6% for traditional SAST tools, with 10x higher true positive rates and ~50% fewer false positives. Currently monitoring 130+ internal Arm projects. Supports C, C++, Python, Go, TypeScript, and Rust via any OpenAI-compatible LLM. <a href="https://app.daily.dev/posts/Er5fnxHEP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI contributions are burying open source maintainers:</strong> GitHub activity has spiked roughly 12x recently, and the load is landing disproportionately on small groups of unpaid maintainers. curl’s bug bounty is buried under AI-generated noise, Jazzband’s ‘slop apocalypse’ led to its sunsetting, and CPython published new AI contribution guidelines. About 55% of open source foundations are permissive on AI use, 25% are banning it outright. GNOME Circle updated its policy to reject submissions showing signs of AI slop; Flathub banned LLM-generated code from new submissions entirely. <a href="https://app.daily.dev/feed-by-ids?id=acGaTVW21&amp;id=jTKFUXHPh&amp;id=JI80hYH4U" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>WorkOS: deleting 95% of agent skills improved accuracy from 77% to 97%:</strong> A DX engineer at WorkOS found that trimming auto-generated agent skills from 10,000 to 553 lines of targeted context improved task accuracy from 77% to 97%. Evals revealed more context was actively hurting results. The retrospective agent pattern — where the system analyzes its own logs and updates memory files to avoid repeating mistakes — is highlighted as the key architectural decision. <a href="https://app.daily.dev/posts/QRj9Q1MXO" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI may be making developers worse despite widespread adoption:</strong> METR’s 2025 study found AI actually slowed developers down, yet developers now refuse to work without it. Uber burned through its entire 2026 AI budget in four months with no measurable productivity gains. AI-generated PRs have 1.7x more defects than human-written ones, and 44% of tokens are spent fixing AI-generated bugs. The recommended framing: treat AI output like junior developer code and keep humans responsible for architecture and security. <a href="https://app.daily.dev/feed-by-ids?id=UjuzNy6kV&amp;id=DX7Ocfiri" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Groq raising $650M after Nvidia’s $20B not-acqui-hire:</strong> Groq is raising $650M from existing investors, six months after Nvidia paid out investors, licensed Groq’s LPU chip technology, and took several senior engineers without fully acquiring the company. The remaining business is the inference cloud. Competition from Cerebras, Fractile, and Google’s Ironwood TPUs is intensifying, and DeepSeek’s 75% price cut on V4 Pro is compressing revenue-per-token economics across the inference cloud market. <a href="https://app.daily.dev/posts/6UC9iAkyx" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Internship postings down 30% since 2023 as AI takes entry-level work:</strong> Tech internship postings have dropped 30% since 2023. Only 7% of new hires at major tech firms are now recent graduates, down from 9.3% in 2023. The ‘Editor Problem’: a generation that can generate AI output but lacks the domain experience to evaluate it is the core concern, not just fewer positions. <a href="https://app.daily.dev/posts/BRqdR9OFj" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cloudflare AI Gateway adds per-API key spend caps:</strong> Cloudflare AI Gateway now supports spend limits at the individual API key level, giving teams finer-grained cost control when multiple consumers or services share a gateway. Useful for anyone trying to prevent runaway agent costs without restricting access entirely. <a href="https://app.daily.dev/posts/TJQelqpCv" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Claude Opus 4.8 now available in Microsoft Azure AI Foundry:</strong> Anthropic’s Claude Opus 4.8 is available in Azure AI Foundry for coding, agentic tasks, and enterprise document analysis including financial analysis, contract review, and regulatory workflows. <a href="https://app.daily.dev/posts/ncwTTHesg" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Codex can now manage its own threads:</strong> Codex gained the ability to create, search, and organize its own threads autonomously. Minor UX improvement but part of a broader pattern of agents managing their own session state. <a href="https://app.daily.dev/posts/BbbsMovsv" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Figma Agent first look: capable on design systems, unreliable on simple tasks:</strong> A hands-on beta test found Figma Agent works well when given rich context like an existing design system, but fails on basic tasks like drawing icons or making a business card. The concern isn’t the ceiling — it’s the unpredictability. <a href="https://app.daily.dev/posts/My8BC19Do" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Qwen3.7-Max review: top-10 globally, $2.50/M input, native Anthropic API support:</strong> Qwen3.7-Max launched May 20 with a 1M-token context window, scoring 56.6 on the Artificial Analysis Intelligence Index — highest-ranked Chinese model on that leaderboard. At $2.50/M input and $7.50/M output with a 90% cache discount, it sits between Gemini 3.5 Flash and Claude Opus 4.7 on price-per-intelligence. Natively supports the Anthropic API protocol, making it a drop-in for Claude Code workflows. <a href="https://app.daily.dev/posts/PuQpfwKD3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Parloa deploys $350M with SAP, Microsoft, OpenAI partnerships, hits $50M ARR:</strong> Berlin-based AI agent management platform Parloa surpassed $50M ARR with 150% net revenue retention. The SAP deal integrates Parloa’s agents into SAP Service Cloud. The platform runs on Azure using GPT-5.4 and competes against Salesforce Agentforce and Google’s Gemini Enterprise Agent Platform. <a href="https://app.daily.dev/posts/ywxtSSI3V" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


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

Agentic Digest

A developer news roundup covering several major stories: GitHub Copilot's shift to token-based billing on June 1 is causing cost spikes of 10-100x for some users. Visa made a strategic investment in Replit and is embedding its Trusted Agent Protocol for agentic payment identity verification. A critical CVSS 10 prompt injection vulnerability in Google's Gemini CLI GitHub Actions workflow was patched after researchers demonstrated full supply-chain compromise. GPT-5.5 outperformed Claude Opus 4.8 on DeepSWE benchmarks across score, cost, and speed. Additional items cover AI agent database destruction incidents, Microsoft BUILD previews, Arm's open-source Metis security framework, AI contributions overwhelming open source maintainers, and declining tech internship postings.

GitHub Copilot's token billing shock, Visa bets on Replit for agentic payments