GitHub Copilot Adds Its First Open-Weight Model, and It's Not Even Close on Price

GitHub Copilot Adds Its First Open-Weight Model, and It's Not Even Close on Price

Kimi K2.7 Code, Moonshot AI's trillion-parameter open-weights model, went generally available in GitHub Copilot on July 1 — the first open-weight model in Copilot's picker, available now to Pro/Pro+/Max users [4]. Business and Enterprise admins have to flip it on manually, which tells you GitHub is still being cautious about governance even while embracing openness [5].

The number that matters: 75% cheaper than Claude or GPT-class models, while matching Sonnet on coding benchmarks and supporting full agent mode [4]. Hacker News and Reddit threads spent less time debating benchmark scores and more time doing cost-per-token math for agentic workloads that burn through millions of tokens a day [6].

This is the quiet trend under the flashy headlines: open-weight models are no longer the budget option you tolerate — they're becoming the default for agentic, high-volume use, with frontier closed models reserved for the genuinely hard problems. If your engineering org is still paying frontier prices for routine coding agents, you're leaving margin on the table.

Anthropic's Fable and Mythos Models Are Back After an 18-Day US-Mandated Blackout

The US Commerce Department suspended access to Claude Fable 5 and Mythos 5 in mid-June over national security and jailbreak concerns — a suspension that hit everyone, including foreign nationals, not just adversarial actors [7]. Anthropic complied by disabling the models entirely for about 18 days. Access was restored around July 1-2, as part of a deal that gives the US government pre-release access to future Anthropic models and adds a new safety classifier layer [8].

For anyone running production workflows on these models, this was a masterclass in concentration risk. Eighteen days of downtime on infrastructure you don't control, for reasons entirely outside your business decisions, is now a documented possibility — not a hypothetical [9].

The relief on X was real, but so was the unease about what the new safety classifier means for advanced or edge-case use. If national governments can switch off frontier models with 18 days notice (or less), "which model should we build on" becomes a geopolitical question, not just a technical one.

MCP Has Quietly Become the Operating System for Agentic Work

Anthropic open-sourced the Model Context Protocol in late 2024 and handed it to the Linux Foundation; by mid-2026 it's the connective tissue behind thousands of MCP servers and SDKs across every major language [10]. The shift people are actually reporting isn't "better chat" — it's a move from prompting Claude to orchestrating it, with skills, connectors, hooks, and sub-agents running inbox triage, research briefs, and personal-OS-style setups [11].

The productivity claims — 5x gains in agentic mode versus simple chat — line up with what we're seeing internally. The unlock isn't a smarter model; it's dynamic tool discovery and parallel execution turning Claude into a coordinator of other systems rather than a single conversational endpoint [12].

Viral posts about multi-agent orchestration (Claude coordinating other models, custom MCP builders wiring up entire toolchains) aren't hype cycles — they're a preview of what "using AI" means once chat interfaces stop being the product and become the control panel.

European AI Models Face a Compliance Wall Just as the Market Moves Faster

Of 3,000+ global LLMs, roughly 30 are European. Zero are fully compliant with the EU AI Act as of July 2026, even with GPAI rules in force since August 2025 and full applicability landing this August [13]. Mistral is publishing compliance playbooks and legal documentation to help providers navigate risk assessments and governance requirements, which is a good sign of pragmatism — but it's also a tacit admission that compliance is currently a full-time job layered on top of actually building models [14].

The timing is brutal. While xAI ships voice agents in two minutes and Moonshot undercuts frontier pricing by 75%, Europe's own model builders are spending cycles on documentation instead of shipping. The calls for deregulation aren't just industry whining — they reflect a real gap: global AI is moving at API-release speed, and EU compliance is moving at directive speed.