GitHub Copilot Billing Flip | AI Field Notes #41

AnalysisCopilot's flat-rate billing ended June 1. Agentic workflows, code review, and multi-step refactor sessions now pull from a pool of GitHub AI Credits charged at the underlying model's API rate, while code completion and Next Edit Suggestions still run on the flat subscription. A Pro subscriber paying $10 per month gets roughly one full agentic session before hitting the credit ceiling. GitHub converts API rates into credits without publishing a simple per-task price, so a developer cannot know what an action costs until after it runs. The company expects heavy users to move to Pro+ ($39 per month) or Enterprise ($19 per seat), and the new billing applies to all plans immediately.

AnalysisClaude Mythos, Anthropic's most capable model, expanded from its original 50-partner cohort to roughly 200 organizations across 15 countries on June 2 under Project Glasswing, a joint vulnerability-research initiative. The new group includes Okta, Samsung, SK Hynix, NATO, and the EU's ENISA cybersecurity agency, covering power, water, healthcare, and communications infrastructure. Anthropic says Mythos can identify thousands of zero-day vulnerabilities in a given codebase over several weeks of autonomous scanning, and estimates a successful attack on one of these partner systems could affect more than 100 million people. A public rollout of Mythos-class models is expected within weeks, pending stronger cyber safeguards the company says it is developing rapidly.

AnalysisGitHub shipped the Copilot App in technical preview on June 2: a standalone desktop client for Windows, macOS, and Linux that runs each coding session in an isolated git worktree so parallel agents do not overwrite each other's edits. The standout feature is Agent Merge, which monitors CI checks, tracks required reviewers, addresses failing tests, and merges a pull request automatically once all conditions are satisfied. Canvases, a new bidirectional work surface, let a developer and an agent edit the same artifact simultaneously. The app is available in technical preview to all existing Copilot Pro, Pro+, Business, and Enterprise subscribers, with no additional cost on top of the new credit billing.

AnalysisThe GitHub Copilot SDK reached general availability on June 2 after five months in public preview, with Rust and Java added at the GA milestone to the existing TypeScript, Python, Go, and .NET support. The SDK exposes Copilot's agentic runtime, which handles planning, tool calls, file edits, and multi-turn sessions, through stable APIs in all six languages. Developers can supply their own API keys via BYOK for Anthropic, OpenAI, or Microsoft Foundry instead of routing through GitHub's credit infrastructure. During preview, teams used it to build CI/CD assistants and internal tools; the production-readiness commitment starts now.

AnalysisMicrosoft debuted two in-house models at Build 2026 on June 2. MAI-Thinking-1 is the company's first reasoning model trained entirely from scratch without distillation, with 35 billion active parameters, a 128K-token context window, and a score on SWE-Bench Pro (a benchmark measuring whether a model can fix real software bugs) that matches Claude Opus 4.6 on the coding track; it is currently in private preview via Microsoft Foundry. MAI-Code-1, a separate inference-efficient model tuned specifically for GitHub workloads, uses 60% fewer tokens than comparable models on hard coding tasks and is already live in GitHub Copilot and Visual Studio Code. Both are being opened to third-party platforms including Fireworks AI, Baseten, and OpenRouter.

AnalysisMicrosoft shipped Microsoft IQ at Build 2026 on June 2: a shared intelligence layer that feeds enterprise context to AI agents across GitHub Copilot, Foundry, and Copilot Studio in a single integration step. IQ has four components: Work IQ (email, documents, meetings, and people), Foundry IQ (knowledge bases and organizational rules), Fabric IQ (live business data from real-time intelligence signals), and Web IQ, which provides fast web grounding and was announced at Build. Work IQ APIs reach general availability on June 16; the rest are available today. The pitch addresses a specific problem: AI agents built on fragmented cloud services keep spawning isolated data silos, and IQ is Microsoft's attempt to solve that with a single context layer.

AnalysisAnthropic filed a confidential S-1 with the SEC on June 1, setting up a potential IPO near a $965 billion valuation after a $65 billion Series H raised in late May. Run-rate revenue reached $47 billion in May 2026, more than doubling OpenAI's last reported pace, with over 1,000 enterprise customers spending at least $1 million annually on an annualized basis. No share count or price has been set; an October 2026 debut is the working assumption if markets cooperate. The filing also signals a structural shift: enterprise hiring has overtaken research hiring at Anthropic for the first time, visible in LinkedIn headcount data cited in analyst coverage of the S-1 disclosure.

AnalysisMiniMax released M3 on June 1 as the first open-weight model to combine a 1-million-token context window, native image and video input, desktop computer operation, and frontier-tier coding in a single package. The Chinese company, which listed on the Hong Kong Stock Exchange in January 2026, reports 59.0% on SWE-Bench Pro, the highest open-weight score on record and within range of Claude Opus 4.7 territory, though the results were run on MiniMax's own infrastructure with agent scaffolding and have not been independently verified. API access starts at $0.60 per million input tokens; open weights and a technical report are expected on Hugging Face within roughly ten days.

AnalysisMicrosoft released ASSERT at Build 2026 on June 2: a developer tool that reads an agent's specification written in plain English and automatically generates an evaluation suite for that agent, then runs those evals against the deployed version before release. It pairs with the Agent Control Specification, a portable governance format that defines what an AI agent is allowed to do in a machine-readable file compatible with any platform. The combination addresses a gap that has widened as agent deployment has outpaced agent testing: organizations are now shipping agents whose behavior they cannot systematically measure. Both tools are available to preview today.

AnalysisCognition raised $1 billion at a $26 billion valuation in late May 2026, led by Lux Capital, General Catalyst, and 8VC, on revenue that grew from $37 million to $492 million in annualized run-rate over twelve months, a 1,230% increase. Enterprise customers including Mercedes-Benz, NASA, Goldman Sachs, and Santander drove tenfold usage growth since January. Cognition itself now attributes more than 90% of its own code to Devin, its autonomous coding agent. The raise explicitly frames Devin as the agent-native alternative to IDE-based coding tools, betting that fully autonomous end-to-end agents will outcompete copilots that hand control back to a human between steps.

AnalysisMicrosoft and Mayo Clinic announced a multi-year collaboration on June 2 to build a frontier AI model trained on Mayo's de-identified clinical data and longitudinal patient records, combining a dataset that spans over 150 years of clinical history with Azure compute and model-training infrastructure. The stated goal is a model capable of supporting the broadest scope of clinical reasoning and healthcare use cases. First outputs go to Mayo clinicians for testing before any broader release. A 'many years' training timeline signals this is long-cycle infrastructure, not a near-term product. The partnership gives Microsoft a healthcare-specific training data agreement that neither Anthropic nor OpenAI currently holds.

AnalysisSpaceX opened investor presentations for its IPO on June 4, targeting a $1.75 trillion valuation and a June 12 first day of trading in what would be the largest equity listing in history. The company absorbed Elon Musk's xAI, owner of the Grok AI model and the X social platform, in a February 2026 all-stock deal at a combined $1.25 trillion valuation, posting $18.67 billion in 2025 revenue and a $4.94 billion net loss from the integration. A debut at $1.75 trillion would exceed Saudi Aramco's 2019 record and would make 2026 the year three AI-adjacent companies (SpaceX-xAI, Anthropic, OpenAI) attempted the three largest equity listings simultaneously.

AnalysisMicrosoft announced Azure Cobalt 200 virtual machines at Build 2026, delivering 50% better compute performance than the prior generation for agentic AI workloads. The ARM-based Cobalt 200 is tuned for the bursty, context-heavy inference patterns that multi-agent systems generate, which differ from the steady throughput of single-model API calls. Alongside it, Microsoft launched Azure Container Apps Sandboxes: secure isolated execution environments designed for AI agent code that runs with lower trust guarantees than a full VM. GPU-Accelerated Fabric Data Warehouse, running up to 7x faster than comparable cloud data warehouses, was also announced, targeting the data retrieval layer that feeds agent context.