Sep 7, 2026 / Headless Commerce
The Shopify community spent 2026 arguing about a word. Hydrogen, Shopify's framework for custom storefronts, is marketed as headless commerce, but a growing number of developers push back that it is something narrower: frontend flexible, Shopify native commerce. The distinction is not pedantry, it decides whether a Hydrogen build gives you what you actually wanted. The part most people miss is that going headless with Shopify does not decouple you from Shopify. Checkout, payments, and the commerce engine stay exactly where they were. What you buy is control over the frontend, not platform independence. In 2026 Hydrogen runs on React Router v7 with Vite, and Shopify has even opened it to run on other stacks including Next.js, which sharpens the question rather than settling it. This post separates what Hydrogen headless genuinely delivers from what the label implies, and gives you a decision framework for when that trade is worth making on a Plus build.
Sep 7, 2026 / AI Infrastructure
MiniMax M2.5 is the open weight coding model that made the closed frontier explain itself. It is a 230B mixture of experts with only 10B active parameters per token, and it scores 80.2 on SWE-Bench Verified, which lands within 0.6 points of Claude Opus 4.6 at 80.8. It also takes first place on Multi-SWE-Bench at 51.3, so the multilingual result is not a fluke. The weights are on Hugging Face, and hosted access runs around fifteen cents per million tokens, roughly a tenth of what a frontier closed model costs to run at volume. That combination, near frontier coding at open weights and a fraction of the price, is the reason M2.5 keeps showing up in build versus buy conversations. This post reads the benchmarks honestly, explains what the 10B active design does to your serving costs, and lays out where an open weight model like this belongs in a real stack and where it does not.
Sep 7, 2026 / AI Infrastructure
Perplexity open sourced Lily, the local inference engine it built for hybrid compute in Perplexity Computer, and the numbers are worth a close look. Lily is a single process runtime: a Rust layer loads the checkpoint and drives the generation loop, hand written Metal kernels execute the model, and an OpenAI compatible endpoint streams tokens. It is specialized, not general. It runs one model family, Qwen3.6-35B-A3B, and it treats prompt prefill and token by token decode as two different workloads with different bottlenecks. On an M5 Max MacBook Pro, Perplexity reports 1.23x higher prefill throughput and 1.35x higher decode throughput than MLX-LM, with output quality effectively unchanged. That is a real gain from specialization on the same hardware. This deep dive reads the benchmark, explains why splitting prefill and decode buys that much, and shows where a narrow engine like Lily belongs next to a general runtime like MLX-LM.
Sep 7, 2026 / AI Tooling
SWE-Bench asks a model to produce a patch that passes tests. Terminal-Bench asks an agent to sit at a terminal and actually get the job done: compile the code, train the model, fix the broken environment, verify the result. That difference in framing is why the two benchmarks rank agents differently, and why Terminal-Bench has become the number people quote when they argue about which coding agent to run. On Terminal-Bench 2.0 the frontier now clears the high 80s and low 90s, but the score hides most of what matters. The benchmark measures a scaffold plus a model, not a model alone, so the same model can move ten points depending on the harness around it. This post explains the task design, reads the current 2026 leaderboard, and shows how to weight the number against your own stack rather than treating it as a single ranking of raw model strength.
Sep 6, 2026 / AI Infrastructure
Both Claude Code and Cursor now let you run several coding agents at once, and the marketing makes them sound interchangeable. They are not. Cursor 3 spawns up to eight agents, each in its own git worktree on a separate branch, so the isolation happens at the filesystem. Claude Code Agent Teams runs multiple model instances that a lead session coordinates, assigning subtasks and merging results, with isolation happening in each agent's context window. That difference decides which work parallelizes cleanly on each tool, how you avoid agents clobbering each other, and where the coordination overhead eats your throughput gains. This post breaks down the two parallelism models, the honest speedups you can expect, and a rule for when parallel agents are worth the orchestration cost at all.
Sep 6, 2026 / AI Infrastructure
On June 2, 2026, Cognition renamed Windsurf to Devin Desktop and pushed it as an over the air update, no reinstall, no migration wizard, settings ported automatically. The surface change is a name and a logo. The real change is strategic: Windsurf was a standalone agentic IDE, and Devin Desktop is now the local half of a lineup whose other half is the cloud Devin agent. That framing, one desktop IDE plus one autonomous cloud engineer sharing a brand and increasingly a protocol, is what the rebrand is actually about. This post separates what changed from what did not, explains how the Agent Client Protocol fits, and places Devin Desktop against Cursor and Claude Code so you can decide whether the rename means anything for your team.
Sep 6, 2026 / AI Infrastructure
Meta released Muse Spark 1.3 on September 2, 2026, and it lands in the top ten of the Artificial Analysis Intelligence Index while charging 1.25 dollars per million input tokens and 4.25 per million output. That is a small fraction of what GPT-6 Astra and Claude Opus 5 cost, and the model posts near perfect retrieval on a million token context where most competitors fade into the 60s and 70s. The catch is that a headline intelligence rank and a cheap sticker price do not tell you where a model belongs in a production stack. This post works through the real numbers: the pricing tiers including the contributor discount that trades your data for a lower rate, the long context retrieval results, and the agent benchmarks where Opus 5 still leads. The point is not that Muse Spark 1.3 is the best model. It is that at this price it changes which jobs you should stop paying a frontier rate for.
Sep 6, 2026 / Headless Commerce
The important B2B news on Shopify in 2026 is not a new feature, it is a pricing move. Company accounts, custom pricing catalogs, net payment terms, self serve ordering, and Shopify Flow automations, all of which used to be Plus only, are now available on every paid Shopify plan at no extra cost. The Winter Editions release layered Sidekick on top, letting merchants spin up a B2B company profile from a natural language prompt. Together those changes reset the decision most wholesale brands face: when is native Shopify B2B enough, when do you still need Plus, and when does the requirement actually call for a headless build. This post maps the native feature set after the 2026 changes, the real reasons to stay on Plus, and the specific signals that mean your wholesale channel has outgrown the native storefront.
Sep 6, 2026 / AI Infrastructure
A model can post a 95 on SWE-Bench Verified and land near half that on SWE-Bench Pro, and both numbers are real. Verified is a 500 task human validated subset of open GitHub issues, and the frontier has largely saturated it in 2026 with scores in the mid 90s. SWE-Bench Pro, built by Scale AI, runs harder tasks across more languages and, crucially, includes private codebases that model trainers cannot legally access, so contamination is prevented by construction. The result is a benchmark that separates models Verified no longer can. If you are picking a coding model on benchmark numbers, reading these two scores as if they measure the same thing is how you end up disappointed in production. This post explains what each one measures, why the gap is structural rather than noise, and how to weight them.
Sep 4, 2026 / AI Infrastructure
The four most used AI coding tools in 2026 price on four different models, which makes a straight sticker comparison almost useless. GitHub Copilot sells seats with monthly AI credits. Cursor sells seats with usage allowances. Claude Code meters against rolling time windows under flat Max tiers. OpenAI Codex bills token metered credits that ride whatever ChatGPT plan you already pay for. The list prices span 10 dollars a month to 200, but the number that decides your actual bill is not the sticker, it is how heavily you run agents, because daily autonomous agent use commonly pushes real spend toward 60 to 200 dollars per developer regardless of which logo is on the plan. This post lays out the September 2026 tiers for each tool, explains the four pricing models, and gives you a way to budget that survives contact with real usage.
Sep 4, 2026 / AI Infrastructure
Anthropic released Claude Fable 5.1 on September 1, 2026, keeping the headline rates at 10 dollars per million input tokens and 50 per million output but cutting the cache read price from 1.00 to 0.25 dollars per million, a 75 percent drop. That single line is more consequential for anyone running agents than the benchmark gains, because agentic workloads re-send a large, stable context on every turn, and cache reads are how you avoid paying full input price for it over and over. At 0.25 per million, a cache read is 2.5 percent of the normal 10 dollar input rate, which changes what a long context agent actually costs to operate. This post explains what cache reads are, why the cut lands hardest on agent loops, what else changed in 5.1, and how the restricted access Mythos 5.1 sibling fits.
Sep 4, 2026 / AI Infrastructure
On September 1, 2026, Anthropic cancelled a planned price increase for Claude Sonnet 5 that would have moved it from 2 dollars per million input tokens and 10 per million output to 3 and 15. The rate stays at 2 and 10. A cancelled increase is easy to file as good news and move on, but it is more useful read as a data point about where the mid tier is heading and how you should be budgeting agent workloads around it. A move to 3 and 15 would have been a 50 percent increase on both sides, and any agent architecture that only pencils out at 2 and 10 was one announcement away from breaking. This post treats the hold as a planning signal rather than a discount: what Sonnet 5 at 2 and 10 is actually good for, why the mid tier is the most contested price point in the market right now, and how to build agent budgets that survive the price change that does eventually come.