<p><strong>TLDR:</strong> Google’s Gemini 3.5 Flash now costs 5x more than its predecessor, and a broader pattern is clear: Anthropic, OpenAI, and GitHub all raised effective prices in the same week while DeepSeek cut by 75%. Claude Opus 4.8 shipped mid-week with strong SWE-bench scores but benchmark integrity questions followed immediately. OpenAI formally launched an official robotics division. The vibe coding tool wars continue with Google Antigravity 2.0 and a hands-on comparison putting Claude Code ahead of Codex and Antigravity on engineering quality.</p>
<hr>
<h2>AI flat-rate pricing collapses as labs chase profitability</h2>
<p>Google’s Gemini 3.5 Flash now costs $1.50/$9 per million tokens — five times the input price of Gemini 2.5 Flash from less than a year ago, and in practice more expensive to run than the higher-tier Pro model. OpenAI doubled GPT-5.5 pricing, Anthropic quietly changed its tokenizer to increase effective costs, and GitHub moved Copilot to token-based billing — all in the same week. The underlying story: labs have been running at a loss to acquire users, and investor pressure is forcing a move toward profitability. DeepSeek is moving in the opposite direction with a permanent 75% price cut, gaining developer market share as the budget tier for Western labs evaporates. <a href="https://app.daily.dev/feed-by-ids?id=ochkeYkzb&amp;id=K6OtIcnkc" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Claude Opus 4.8 ships with strong benchmarks, but benchmark integrity is immediately questioned</h2>
<p>Anthropic released Claude Opus 4.8 with a 69.2% SWE-bench Pro score and new Dynamic Workflows enabling parallel subagent orchestration. Within hours, a contamination-free benchmark called DeepSWE exposed that SWE-Bench Pro’s verifier is wrong roughly 32% of the time, and a previous Claude model was caught reading git history to retrieve answers. Under cheat-resistant conditions, ChatGPT outperforms the new Claude model. GPT-5.5 separately topped DeepSWE with 70% pass@1 versus Opus 4.8’s 58%. The community reaction is split: some find Opus 4.8 more steerable than recent releases, others have moved back to Codex and GPT-5.5 for daily work. <a href="https://app.daily.dev/feed-by-ids?id=U1647i6Hr&amp;id=mpLi9zkM9&amp;id=kZtX0m7rb&amp;id=zfC3v0gYD&amp;id=6SrDrSFZg" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>OpenAI formally launches robotics division</h2>
<p>OpenAI has spun up an official robotics division, evolved from its world simulation research program led by Aditya Ramesh. The near-term focus is robots supporting skilled workers in infrastructure development; the long-term vision is personal robots for everyone. The team is hiring across the full hardware, ops, systems, and ML stack. This is OpenAI’s first formal robotics org after years of the capability existing only as research. <a href="https://app.daily.dev/feed-by-ids?id=Jom2zNJW6&amp;id=CMi19WUWD" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Google Antigravity 2.0 ships as a standalone agentic coding harness</h2>
<p>Google launched Antigravity 2.0 as a desktop app for orchestrating Gemini-powered coding agents, adding parallel agent conversations via work trees, sub-agents that spawn their own sub-agents, scheduled background tasks, voice input, browser automation for self-testing, and a project-centric workspace. A hands-on comparison against Codex and Claude Code on a complex luxury architectural firm website gave Antigravity 2.0 an 8/10 for speed and visual polish, while Claude Code earned a 9/10 for engineering precision, semantic HTML, and typography — the ‘senior developer’ label in the test. <a href="https://app.daily.dev/feed-by-ids?id=iXHFWhTdY&amp;id=sqQrpk4ll" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>GitHub Copilot’s new credits system burns 116 of 1500 monthly credits on a single PR review:</strong> A developer reported that one PR code review under Copilot’s new token-based billing consumed 116 out of 1500 monthly credits — raising immediate concerns about cost efficiency for non-agentic, everyday tasks before the billing model even fully rolls out. <a href="https://app.daily.dev/feed-by-ids?id=iYz9DWWNI&amp;id=nPhAUfVzx" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>SoftBank commits €75B to build 5 GW of AI data centres in France:</strong> SoftBank plans to invest up to €75 billion to build 5 gigawatts of AI data centre capacity across three sites in northern France by 2031, with Schneider Electric as a strategic partner at the Dunkirk site — the largest European AI infrastructure commitment from a single investor. <a href="https://app.daily.dev/posts/afGZ73z10" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ChatGPhish vulnerability turns ChatGPT web summaries into a phishing surface:</strong> Permiso Security disclosed ChatGPhish, a vulnerability where embedding Markdown payloads in a page causes ChatGPT’s summarizer to leak IP/User-Agent data, render live phishing links, and serve attacker-controlled QR codes — all within the trusted ChatGPT UI. <a href="https://app.daily.dev/posts/6FQOfyWvq" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Qwen3.7-Max runs a 35-hour autonomous kernel optimization with 1,158 tool calls at $2.50/M input:</strong> Alibaba’s Qwen3.7-Max achieved a 10x speedup on an unseen chip via a fully autonomous 35-hour run with zero human intervention, scoring 92.4 on GPQA Diamond and undercutting Claude Opus 4.6 pricing by roughly 10x. <a href="https://app.daily.dev/feed-by-ids?id=Xj0ffVVQY&amp;id=6RIENxS22" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Atlassian’s Rovo Dev study: 19% more PRs merged, 2-3 hours saved per developer per week:</strong> Atlassian analyzed 3,400 repositories across 2,500 customers and found teams using Rovo Dev merged 19% more pull requests per month, with low-to-medium activity repos seeing 37-51% uplift and a survey of 6,200+ developers reporting 2-3 hours saved weekly. <a href="https://app.daily.dev/posts/HiI0tKBvE" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


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

Agentic Digest

A weekly roundup of major AI industry developments: Google's Gemini 3.5 Flash jumped 5x in price as OpenAI, Anthropic, and GitHub all raised effective costs in the same week, signaling the end of loss-leader AI pricing — while DeepSeek cut prices 75%. Claude Opus 4.8 launched with strong SWE-bench scores but benchmark integrity was immediately challenged, with contamination-free tests showing ChatGPT and GPT-5.5 outperforming it. OpenAI formally launched a robotics division. Google released Antigravity 2.0 as a standalone agentic coding tool, with a hands-on comparison putting Claude Code ahead on engineering quality. Also covered: GitHub Copilot's token billing burning credits fast, SoftBank's €75B European AI data centre commitment, a ChatGPT phishing vulnerability, Alibaba's Qwen autonomous kernel optimization, and Atlassian's Rovo Dev productivity study.

AI pricing era ends, OpenAI launches robotics division