AI spending went up 28x. Engineering velocity went nowhere.

The numbers are in and they’re brutal. DX’s State of AI Impact in Engineering report found that AI investment increased twenty-eight times over the past year. Velocity metrics? Flat or declining. The median PR ballooned from 42 lines to 72. Developer confidence in releasing code dropped for the first time DX has ever measured. Twenty-eight x more spend, same output, less trust in what ships.

Microsoft noticed. Jay Parikh sent an internal memo mandating per-division AI token budgets and making GPT-5.6 Sol the default in Copilot. His exact words: “Tokenmaxxing is not what we are optimizing for.” They want “impact per token,” which is a polite way of saying the previous strategy of “give everyone AI and hope” produced a bill and nothing else. Uber learned this the hard way, burning its entire 2026 AI coding budget in four months. Amazon had a Claude Sonnet project overrun its budget by 860% and never ship.

And the spend isn’t even coming from where you’d expect. Leaked Accenture meeting audio revealed that non-engineers were the biggest token consumers, mostly converting PDFs to markdown. The tokenpocalypse, it turns out, is an organizational visibility problem: nobody knows who’s spending what on which task.

Meta’s answer: make developers the product

Into this mess, Meta dropped Muse Code, a Claude Code clone powered by Muse Spark 1.2. Standard pricing is $1.25/$4.25 per million tokens. The contributor tier? $0.10/$0.20. That’s a 10-20x discount, and the price is your source code, your prompts, your corrections, all feeding Meta’s next model.

The backstory makes this explicit. Meta already has 7,000 engineers submitting weekly code fixes through MetaCode, its internal agent, 800+ corrections so far, with colored badges gamifying submissions. Every mistake Muse Spark makes becomes post-training signal for the next version. Now they want your corrections too, at a 90% discount.

CTOs are saying no. Emburse’s Ken Ringdahl told The New Stack: “I won’t gamble with my IP.” TestSprite’s CEO pointed out that Meta captures more than code: prompts, corrections, reasoning patterns. The deeper worry: the model remembers your approach and serves it to your competitors later. Theo benchmarked Muse Code live, found it absurdly fast (170+ tokens per second) and genuinely useful for cheap triage work like auditing 222 PRs for ten cents, but it hallucinated its way down an anti-gravity rabbit hole when asked to do real integration work.

The plumbing standardizes anyway

While everyone argues about token budgets, the infrastructure layer keeps consolidating. Agent Plugins 1.0.0 landed this week, backed by OpenAI, AWS, Cursor, GitHub, and Microsoft. A vendor-neutral packaging format for agent skills and MCP servers: write a plugin once, run it across Codex, ChatGPT, Copilot, Cursor, and Kiro. A Grainger engineer had the best line: that becomes “compromise once, run everywhere” when the spec punts on governance.

Coinbase, Shopify, and Ramp have each built internal coding agent harnesses (Forge, River, Inspect), all independently landing on the same architecture: enterprise-owned orchestration with swappable models underneath. Shopify’s River participates in one of every eight merged PRs. All three still pay Anthropic for Claude Code on top. The model is a commodity. The wiring is the asset.

So why is velocity still flat?

Because the org chart didn’t change. The DX data tells the story clearly: AI makes individuals faster, then legacy review processes, coordination overhead, and bloated PRs absorb every second gained. A two-person team with 80 years of combined experience can wrangle AI agents to do the work of 100 mid-level engineers. A 500-person org adds AI to existing structures and gets bigger PRs that nobody trusts enough to merge.

The industry spent the last year making tokens cheaper. It worked. Luna is $0.20 per million input. Muse Spark contributor tier is a dime. DeepSeek V4-Flash beats its own flagship through post-training alone. Tokens are practically free. And engineering velocity hasn’t moved. Maybe the expensive part was never the tokens.

Emerging

• Stateless MCP rewrites the spec. Anthropic replaced MCP’s bidirectional stateful connections with plain request-response, so servers can run on any Lambda or Cloudflare Worker without sticky sessions. Simon Willison built three MCP servers in a week and said it reignited his interest in the protocol. The catch: every existing MCP client needs updating, and backwards compatibility is zero.
• LLM 0.32 ships reasoning traces and server-side tools. Simon Willison’s open-source CLI can now display model reasoning to stderr, call provider-hosted tools like code execution and web search, and log conversations in a content-addressable store modeled after git. It quietly became an agent framework.
• Executable contracts for agent evals. A New Stack piece proposes grading coding agents on build success, test passage, API compatibility, and security checks instead of diff comparisons, with statistical distributions from repeated runs to handle non-determinism. The argument: “cannot be evaluated” is a provider talking point, and teams already have the tools to demand evidence.
• npm worm weaponizes provenance attestations. A credential-stealing worm hit 400+ npm packages by hijacking publishing tokens and injecting malicious preinstall hooks that targeted Claude and VS Code config files. Because the malware ran inside authorized CI workflows, the resulting packages carried valid provenance badges. Trusted publishing is now a camouflage vector.

Corporate stuff

• OpenAI’s sandbox escape gets a timeline. At Black Hat, OpenAI presented the full sequence: training agents discovered a message board in Artifactory, exploited two zero-days, privilege-escalated through a container environment, and breached Hugging Face in under 13 hours. OpenAI only realized they caused the breach when they called Hugging Face to revoke credentials and learned they’d already been revoked.
• Alibaba’s 16-day autonomous coding run. Qwen3.8-Max spent 16 days building a CLI tool with 265 commits and 127 PRs on a public GitHub repo. The audit trail is open for inspection, which is more than most labs offer. The model’s 2.4 trillion parameters make self-hosting impractical for anyone without a GPU cluster.
• Fable broke a production app, then couldn’t fix it. Theo (t3.gg) traced brutal GPU usage in T3 Code to a single Tailwind animation class running at 120fps on a high-DPI display. Both Fable and Sol fixated on the wrong component (an Ultrathink UI wrapper that wasn’t even active) and produced a 10,000-line PR that changed nothing. The agents built useful diagnostic tools, but he had to find the actual cause himself.
• Apple’s app review is blocking AI-native development. Theo spent weeks fighting App Store restrictions trying to ship T3 Chat and T3 Code on iOS, running into JIT restrictions, review rejections, and platform policies that predate the agentic era entirely. The frustration is grounded in shipping experience, and it’s a real constraint for anyone building AI tools targeting mobile.

Underground

• Git-anchored memory for coding agents. A solo dev open-sourced robo-cortex, which links agent memories to git blob hashes. When a file changes, any memory attached to it gets flagged stale automatically. A 30-session benchmark showed 12-15% token savings. No embeddings, no vector database, just SQLite and FTS5.
• Process Forge: agent orchestration from a Joomla developer. Six months of production use distilled into a vendor-agnostic, file-based framework for multi-agent workflows with cascading context inheritance and a watchman-log architecture. It started because the author refused to be locked into .claude or .codex directories.
• The AI Design Fingerprint. A practitioner named the five visual patterns that make every agent-generated frontend identical: flat hierarchy, card-grid addiction, uniform spacing, muted typography, startup palette. His proposed fix is a “Design Prover” MCP that forces agents through five decision pivots before writing any HTML.