We're watching a full-on crackdown on "tokenmaxxing" this week as companies realize how fast agentic coding loops can drain a budget — Meta just killed its own internal leaderboard after employees burned through 60+ trillion tokens in a month. But while corporate budgets tighten, the dev job market is doing something less expected: it's rebounding hard, just narrowly, for senior AI-fluent engineers.

This Week in AI Coding

• US developer job postings are up ~15% since Claude Code launched — but the gain is narrow. Indeed Hiring Lab's latest read puts software-dev postings at an index of ~114.6, roughly 15% above where they stood in February 2025. The catch: 71% of that increase came from senior roles, and 37% came from postings that put "AI" directly in the job title. If you're junior and don't have an AI-fluency signal on your resume, this recovery isn't really for you yet.
• Meta dismantled its internal AI token leaderboard after employees burned 60+ trillion tokens in 30 days. The "Claudeonomics" dashboard ranked ~85,000 employees by token consumption, complete with gamified titles like "Token Legend." It backfired — usage became the KPI instead of output — so Meta's replacing it with a centralized "AI Gateway" spend tracker and pushing engineers toward its own in-house coding assistant, MetaCode. A clean case study in Goodhart's Law for anyone setting AI usage incentives on their own team.
• OpenAI shipped GPT-5.6 (Sol/Terra/Luna) to everyone on July 9, after a first-of-its-kind government review. The family had been gated behind a restricted, ~20-organization preview since June 26 while the Commerce Department's AI Standards and Innovation Center vetted it — the first time a US lab held back a frontier model behind a state-approved roster. Pricing: Sol (flagship) $5/$30 per million tokens, Terra aims for GPT-5.5-level quality at roughly half the cost, Luna is the cheap high-volume tier.
• Google scrapped and rebuilt Gemini 3.5 Pro from scratch, now targeting a July 17 GA. Early testers reportedly found real gaps in recursive tool-calling and SVG generation, bad enough that Google threw out the base model and re-trained rather than patch it. A 2M-token context window is rumored but, as of this week, still unconfirmed by Google directly — treat the exact specs as a leak until there's a model card.
• Claude Code's July 17 update reworks how long-running sessions behave. /fork now spins your conversation into a full background session instead of a same-session subagent (that job moved to the new /subtask); there are now session-wide caps on WebSearch calls (default 200) and subagent spawns (default 200) to stop runaway loops; and MCP tool calls running past 2 minutes auto-background so the session doesn't hang. Useful if you've had a Claude Code session lock up on a slow MCP call.

What We Shipped

It was a week of hard cuts and pipeline debugging here at VeriduxLabs: - We killed SPARit (sparit.vc) on July 8 after two months as our lead launch product — zero paying customers, no funnel conversion. Focus is now fully consolidated on TypeWright (typewright.dev), our VS Code fork with a local 1.5B TypeScript model and a QLoRA fine-tuning pipeline, which we're building toward an exit. - Caught and fixed a 5-day-old bug in our YouTube content-backlog tracker: it was checking a stale legacy script directory, which made it look like we had 54 unscripted episodes queued up. The real number is 5. The script-writing side of the pipeline has quietly been on schedule the whole time. - That same audit found a real problem, though: our Remotion render/compose pipeline has had zero commits in 40 days. 57 fully-scripted episodes are sitting ready, waiting on rendering — unrelated to (but stacking on top of) our still-broken YouTube upload OAuth. Both are queued for a fix. - 4 new episode scripts drafted (EP174-177), keeping the content cadence going even with the render backlog.

Tool of the Week

Inkling, from Thinking Machines Lab — a 975B-parameter open-weights Mixture-of-Experts model (41B active params) that natively reasons over text, images, and audio with a 1M-token context window. The pitch isn't "strongest model available" — it's a strong, customizable open-weights base: efficient, controllable thinking effort, and available on Tinker for fine-tuning. Worth a look if you want an open-weight model you can actually shape instead of just prompt.

Quick Links

• Indeed Hiring Lab — AI and Job Postings: From Destruction to Creation?
• Thinking Machines Lab — Introducing Inkling
• MLQ News — Meta Caps Internal AI Token Spending
• Claude Code Changelog

endofcoding.com | vibecodingebook.com | llmhire.com | typewright.dev | YouTube @endofcoding