A strong week for pricing news and platform moves. Token economics are forcing real decisions across the AI coding stack. Here's what mattered.
Grok 4.5 undercuts Anthropic and OpenAI on coding agent pricing
SpaceXAI shipped Grok 4.5 this week — the first model built jointly with Cursor since the acquisition — and the pitch is direct: Opus-class coding performance at a fraction of the cost. Grok 4.5 is priced at $2 per million input tokens and $6 per million output, compared to $5 and $25 for Anthropic's Opus 4.8. Independent benchmarks put it fourth on overall intelligence rankings, trailing Fable 5, GPT-5.5, and Opus 4.8 — but token efficiency is where it separates itself. Cursor's own data shows Grok 4.5 using roughly 4.2 times fewer output tokens than Opus 4.8 on equivalent coding-agent tasks.
One disclosure worth noting: an earlier snapshot of the Cursor codebase accidentally made it into Grok 4.5's training data, giving it an edge on CursorBench that Cursor says can't be fully quantified. The data has been removed from future runs. For teams evaluating the model, that's worth factoring in before treating benchmark charts as the final word.
Read more
Copilot bills hit $800 — Visual Studio responds with usage alerts and MCP trust checks
When GitHub moved Copilot from request-based to token-based billing on June 1, some heavy agentic users saw monthly costs climb from around $39 to over $800. The complaint was consistent: nobody saw it coming. Visual Studio's June update addresses that directly with a real-time usage window that shows token consumption, configurable warning thresholds, and alerts when overage billing kicks in. It's the visibility that should have shipped alongside the billing model itself.
The same update also adds a two-stage trust check for MCP servers. Before startup, the current config is compared against a previously trusted baseline. After startup, the tool and instruction fingerprint is checked again. If anything changed, you get a dialog — not a silent failure. For teams that added MCP servers early and haven't looked at them since, this is worth enabling now.
Read more
Microsoft's AutoGen reaches 1.0 with a manager-led orchestration model
AutoGen 1.0 ships with a opinionated bet on how multi-agent systems should work: a manager agent coordinates a team of specialized agents, each scoped to a specific role. The framework now has a stable API, making it a viable foundation for production agent systems rather than just research prototypes. If you're architecting multi-agent workflows, this one is worth a closer look.
Read more
Kore.ai runs AI at 6,500 commits a month — here's the infrastructure behind it
This is one of the more concrete production accounts of agentic coding at scale published this year. Kore.ai built a harness that lets AI agents write 6,500 commits per month — and the piece details exactly what it takes to get there: the guardrails, the review workflows, the failure modes. If you're building toward anything resembling this, it's required reading.
Read more
How Intuit is rethinking reliability for AI-generated code
AI-generated code introduces reliability risks that traditional testing frameworks weren't built to catch. Intuit's approach is worth understanding — not as a case study to copy, but as a template for what serious reliability engineering looks like when agents are writing production code.
Read more
Also worth reading this week
The through-line this week: token economics are no longer a background consideration. Pricing, visibility, and efficiency are now first-class decisions for every team running AI coding tools at scale.
See you next Friday.
Developer Weekly Briefing is published every Friday on Coder Legion. Written by Tom Smith.