Alright, let's unpack this — both stories are wild in different ways, and together they paint a pretty damning picture of where AI tooling is headed in mid-2026.
1. Microsoft Internal Teams Using Claude Code Instead of Their Own Copilot
Yeah, this is real, and it's as ironic as it sounds.
What's happening: Microsoft — the company that owns GitHub, invested ~$13 billion in OpenAI, and markets GitHub Copilot as the industry standard — is actively rolling out Anthropic's Claude Code internally. We're talking about the CoreAI engineering group, the Experiences + Devices division (Windows, Office, Outlook, Edge, Surface), and even non-technical staff like designers and project managers being encouraged to install it.
Why Claude Code is winning inside Microsoft:
- Agentic depth: Claude Code doesn't just autocomplete lines — it analyzes entire repos, executes terminal commands, and proposes multi-file fixes. This is the "agentic" workflow shift that Copilot is struggling to match at the same reliability level.
- Accessibility: Even designers and PMs are using it for rapid prototyping without deep coding knowledge.
- Precision: Anthropic's Claude 4 series (Sonnet 4.5/4.6, Opus 4.1/4.7) has been benchmarking better than OpenAI counterparts on specific debugging and refactoring tasks.
The scale of it: According to reports, Microsoft is spending close to $500 million annually on Anthropic products, making them one of Anthropic's top customers. This isn't a rogue engineer thing — it's a coordinated pilot program.
Microsoft's official line: Communications chief Frank Shaw said, "OpenAI continues to be our primary partner and model provider. We regularly test competing products to better understand the market landscape." Which is corporate speak for "yes, we're using the competitor's tool because it's better for the job."
The deeper signal: Even the creator of Copilot recognizes that a single AI model monopoly doesn't work. Different tasks need different "brains." The fact that Microsoft is running what amounts to a "coding referendum" between Copilot and Claude Code internally tells you everything about their confidence level in their own product's supremacy.
2. GitHub Copilot's Pricing Model Change: The June 1, 2026 Backlash
This one just went live today (June 1, 2026), and developers are absolutely losing it.
What changed: GitHub Copilot switched from a flat "premium request" model to token-based billing using "GitHub AI Credits" (1 credit = $0.01). Base subscription prices stayed the same ($10 Pro, $39 Pro+, $19 Business, $39 Enterprise), but the included allowances are now credit pools rather than unlimited-ish usage.
The math that broke people:
- Copilot Pro: $10/month now includes 1,000 AI Credits
- A single agentic session with a frontier model (like Claude Opus 4.7) can burn 30–40 credits
- Heavy users who were paying ~$29/month are now seeing projections of $750/month
- Some power users estimate jumps from $50 to around $3,000/month
The kicker: Under the old system, if you exhausted your premium requests, Copilot would fall back to a cheaper model and let you keep working. That fallback is gone. When your credits hit zero, premium features just stop until next billing cycle or until you buy more.
Why GitHub says they did it: In their own words, "Copilot is not the same product it was a year ago" — it's now an agentic platform running long, multi-step sessions that cost significantly more compute. They also admitted in internal docs (leaked by Ed Zitron) that the week-over-week cost of running Copilot had nearly doubled since January 2026.
The community response: The official GitHub community FAQ thread had 904 downvotes and 22 upvotes as of this morning — one of the most lopsided reactions in GitHub forum history.
Annual plan holders getting squeezed too: If you're locked into an annual plan, you're still on the old PRU system — but model multipliers just got jacked up. Claude Opus 4.7 went from a 7.5x multiplier to 27x. GPT-5.4 went from 1x to 6x. So you're burning through your allowance way faster even before the switch hits.
What this means practically:
- Tab-completion users (the majority) → probably fine, completions remain free/unlimited
- Agentic power users (the ones GitHub spent 2 years promoting) → 10x to 50x cost increases
- Code review now also burns GitHub Actions minutes on top of AI Credits, so teams with frequent automated reviews get hit on two billing tracks
The Reality Check: Connecting Both Stories
Here's the uncomfortable truth these two stories reveal together:
Microsoft can't afford its own AI at scale, and neither can its customers.
Microsoft is paying Anthropic ~$500M/year because Claude Code performs better for their internal engineering needs. Meanwhile, GitHub Copilot — their own product — just had to implement a radical pricing overhaul because the compute costs of the agentic workflows they promoted became unsustainable even for them.
The leaked internal docs said Copilot's week-over-week costs nearly doubled in 5 months. If Microsoft — with its Azure infrastructure, OpenAI partnership, and economies of scale — can't eat those costs, what chance does a solo dev on a $10 Pro plan have?
The alternatives people are jumping to:
- Cursor ($20/month) and Windsurf ($15/month) are the direct competitors seeing an influx
- Open-source options (Continue.dev, Cline, Aider) let you bring your own API keys and pay inference costs directly without platform markup
- And of course... Claude Code itself, which Microsoft is already using internally
The broader trend: Satya Nadella said on Microsoft's latest earnings call that any per-user business at Microsoft — productivity, coding, security — is becoming a "per-user and usage business." This isn't just GitHub. Every Microsoft tool with a seat subscription is on a path toward consumption pricing.
So yeah — the company that built Copilot uses Claude. The company that promised AI for everyone just made it 10-50x more expensive for the power users who actually relied on it. And the "free" tier is basically a demo now. Welcome to AI in 2026.
Top comments (0)