GitHub Copilot's first full month of metered billing just closed, and some developers are staring at bills ten times higher than before. Meta says its new brain implant-free system can now read full sentences straight from a brain scan. Claude just landed inside Microsoft's cloud and got a half-price deal with California's government in the same week. Here is what matters today, and what to do about it.

GitHub Copilot completed its first full 30-day cycle under usage-based "AI Credits" billing on June 30, the system that replaced flat-rate plans on June 1. Developers running agentic coding sessions report monthly costs jumping from $29 to $750 and from $50 to $3,000, a 10x to 50x increase, since every input, output, and cached token now draws from a fixed monthly credit pool with no spending ceiling unless one is set manually.

Why it matters: The end of flat-rate AI coding tools shows the industry's era of subsidized, all-you-can-use access is closing, and teams that do not actively manage model choice and usage caps could see costs spiral overnight.

What to do:

Source: Tech Times

A new OpenAI research paper published today tracked Codex usage inside the company and found Research-team output on the tool grew 56 times since November 2025, Customer Support 32 times, and Engineering 27 times. Codex now generates 99.8% of OpenAI employees' AI output tokens versus ChatGPT, and Legal, Finance, and Recruiting crossed over to majority Codex use around April 2026, months after engineering adopted it first.

Why it matters: This is one of the clearest internal datasets yet showing agentic tools overtaking chat interfaces for real work, and non-engineering teams adopting fastest suggests coding agents are becoming general workplace tools, not just developer tools.

What to do:

Source: OpenAI

Meta's FAIR lab released Brain2Qwerty v2 on June 29, decoding complete typed sentences from non-invasive magnetoencephalography brain scans at 61% average word accuracy across nine volunteers, and 78% for the best participant, up from about 8% for prior non-invasive methods. The model trained on roughly 22,000 sentences collected over 10-hour sessions per volunteer, and Meta has open-sourced the training code for both this version and its 2025 predecessor.

Why it matters: Reaching accuracy once thought exclusive to surgically implanted brain-computer interfaces, without surgery, points to a realistic path toward communication tools for people who have lost the ability to speak.

What to do:

Source: Meta AI

Suno launched Spark on June 25, an incubator offering unsigned artists grants worth thousands to tens of thousands of dollars, mentorship, and marketing support in exchange for making music with Suno. Buried in the fine print is a "Good Vibes Only" clause barring participants from ever saying anything negative about Suno, plus a 60-day ban on working with rival AI music platforms after their final post goes live.

Why it's interesting: It is a strikingly candid example of a company using financial support to buy silence from the community it says it wants to empower, especially while facing an active lawsuit over how its models were trained.

Key takeaway: Read the fine print on any AI grant or partnership program before signing away your right to criticize the company later.

Source: Music Business Worldwide

Anthropic's latest Economic Index report, built on hourly usage data and a survey of 9,700 Claude users, found personal use of Claude climbs from about 35% of conversations on weekdays to nearly 50% on weekends, with recipe questions spiking in the evening and sleep advice clustering before dawn. The report also found that people who delegate the most work to Claude are the most optimistic about their job security and pay over the next year.

Why it's interesting: It captures AI quietly becoming part of ordinary daily rhythms rather than staying a strictly work-only tool, and challenges the assumption that heavy AI users are the most anxious about their jobs.

Key takeaway: The heaviest AI users on your team may be your best gauge of what full adoption looks like, not your most at-risk employees.

Source: Anthropic

Context: With GitHub Copilot's usage-based billing now in its second cycle and other coding tools following suit, the fastest way to cut costs without hurting output is matching each task to the cheapest model or mode that can actually do it well.

Prompt: I run [team or project] and my current AI coding spend is [tool and monthly cost]. List my main coding tasks (for example autocomplete, agentic refactors, PR review, debugging, documentation). For each, tell me whether it needs a frontier model or can run on a cheaper model or plain autocomplete, and explain the quality risk of moving it down a tier. Then give me a migration order ranked by the biggest savings for the least risk.

Do not: Do not move complex, multi-file agentic tasks to a cheaper model without testing output quality on a real task first.

If/Then: If a task touches production code with no code review step, keep it on the strongest available model.

Example: A five-person dev shop, Northwind Labs, ran this prompt, moved routine autocomplete and doc generation to a cheaper mode, kept agentic refactors on the frontier model, and cut its monthly Copilot bill by roughly 35% with no drop in shipped code quality.

Claude Opus 4.8 and Claude Haiku 4.5 became generally available in Microsoft Foundry on June 29, giving enterprise teams native Azure billing and identity management for the first time. Days later, California Governor Gavin Newsom announced a deal making Claude the first AI productivity tool available to every state agency and local government at a 50% discount, alongside free workforce training from Anthropic.

Why it's important: Landing inside Azure's billing system and inside a state government's procurement portal in the same week shows Anthropic pushing on both enterprise and public-sector distribution at once, even as it remains in a separate legal dispute with the federal government over defense contracts.

Business takeaway: If your organization already runs on Azure, Claude now shows up on your existing invoice with no separate contract needed.

Source: Office of Governor Gavin Newsom

Rocket Lab agreed to acquire satellite operator Iridium Communications for about $8 billion on June 29, combining its own launch vehicles with Iridium's 66-satellite network and 2.5 million subscribers to build a SpaceX-Starlink-style combined launch-and-communications business. Rocket Lab is funding the cash portion with a $3.6 billion bridge loan and says it also plans to build orbital AI computing infrastructure.

Why it's important: It is the clearest sign yet that space companies see owning both the rockets and the satellites, not just one or the other, as the winning model, and it puts Rocket Lab in more direct competition with SpaceX.

Business takeaway: Watch for more vertical-integration deals across the space and satellite-communications sector as AI compute demand extends into orbital infrastructure.

Source: SpaceNews

SK Hynix, which holds about 60% of the market for high-bandwidth memory chips used in AI accelerators, filed to raise up to $29.4 billion through a Nasdaq listing of American depositary receipts, with trading tentatively starting July 10. Every dollar raised is earmarked for new fabrication and packaging capacity, including its Yongin cluster in South Korea, as HBM capacity stays sold out into 2027.

Why it's important: A memory chip supplier converting a scorching AI-driven valuation into a record US listing signals just how tight the supply chain behind every AI accelerator remains, which keeps upward pressure on compute costs.

Business takeaway: Budget for continued tightness in AI hardware supply chains rather than expecting memory and chip costs to ease soon.

Source: CNBC

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas