Quick Links for Devs: Week 27, 2026

Tokenmaxxing is dead, long live tokenmaxxing

theahura at 12 Grams of Carbon flips the dominant narrative on its head. The popular take on "tokenmaxxing" — Meta and others accidentally incentivizing employees to burn tokens on nothing — misses the point. This was a blunt-force strategy to break through organizational resistance to AI adoption. Senior engineers and respected holdouts wouldn't touch AI tools, so executives tied token usage to performance reviews. Yes, people ran two agents chatting with each other all day. Yes, money was burned. But it worked: now everyone uses AI at least a little. The problem is what comes next. Tokenmaxxing created volume, but OpenAI and Anthropic are now stripping subsidies as they prep for IPOs — API prices are rising, subscription limits tightening, and suddenly all that token-hungry behavior hits the bottom line. The companies that figure out how to move from "everyone uses AI" to "using AI effectively" are the ones that pull away. The race to efficiency starts now.

Introducing Claude Sonnet 5

Anthropic dropped Claude Sonnet 5 — "the most agentic Sonnet model yet" — narrowing the gap with Opus 4.8 at Sonnet-class prices. It's a strict improvement over Sonnet 4.6 across reasoning, tool use, coding, and knowledge work, and at higher effort levels it can match Opus 4.8 on some tasks. The pricing is aggressive: intro rates of $2/$10 per million tokens (input/output) through August, then $3/$15 after. That undercuts OpenAI's GPT-5.6 Terra ($2.50/$15) and demolishes Sol ($5/$30). Safety-wise, it shows lower hallucination and sycophancy than its predecessor, better prompt injection resistance, and cyber safeguards enabled by default — though it scored somewhat higher on misaligned behavior vs Opus 4.8 and Mythos Preview. If you've been following the pricing war from week 26, this is Anthropic's counter-move: use Sonnet to compete on cost while keeping Opus/Mythos as the premium tier. The frontier model pricing compression continues.

The Winners and Losers of the AI Revolution

Tyler Cowen at ARC 2026 delivers a deliberately sober forecast: AI won't cause mass unemployment, but it will radically remix social status. The biggest losers? High-credential, rule-following professionals — the Manhattan partners on $2M salaries who did everything "right." The winners? People with initiative who figure out how AI and agents work and do something different. Cowen's economic projection is modest — AI adds ~0.5% to growth, slowed by organizational and regulatory drag — but that's still enough to rescue the US from its $38T national debt. The uncomfortable truth: countries without their own AI risk losing sovereignty, and the world will choose between American and Chinese systems. Pairs with the Borretti piece from week 26 — Borretti says nobody escapes, Cowen says the reshuffle is already underway and initiative is the only hedge.

Working With AI: A Concrete Example

Carson Gross (htmx creator) walks through a real bug in the hyperscript parser — an as keyword binding conflict — and shows exactly where AI helps and where it hurts. Claude nailed the diagnosis in minutes, finding a root cause in a refactor that accidentally expanded the grammar after fetch. But when it came to the fix, Claude proposed a hack that only solved the reported case, then an over-engineered general solution. Carson wrote the actual fix himself: two lines. The takeaway: "AI is very good at helping you understand things. It is less good at helping you fix things." The Sorcerer's Apprentice problem is real — accept AI's fixes without understanding the system and you ship bugs that break other cases. Use AI for diagnosis, apply your own judgment for the fix.

Qwen 3.6 27B is the sweet spot for local development

Piotr Migdał at Quesma makes the case that Qwen 3.6 27B (dense) is the first local model that genuinely works for development. It wrote a hexagonal minesweeper in React from a single prompt — first try, proper Node package — and built a responsive landing page from a short prompt. The MoE variant (35B A3B) is faster at ~105 tokens/s but less precise, ignoring instructions. The dense 27B is slower at ~18 tokens/s but more reliable. Running via llama.cpp with 8-bit quantization on a MacBook Max M5 128GB, it uses ~41GB RAM. Migdał recommends against Ollama on ethical grounds and points out it works with OpenCode and Hermes as a local agent backend. If you read the O'Claire piece from week 26 on the 50x pricing gap between open-weight and frontier models, this is the proof — local models are no longer toys. They're practical dev tools.