Three of your coding subscriptions changed shape this year, and the cheapest model you can run is now a one-click install
Monday, September 7, 2026·8 min read·4 stories
Four stories today, one thread: the cost of a unit of work keeps moving, and mostly not where the announcements point. Three of the assistants you may still be paying for quietly changed their billing or their end date this year. A free installer now reads your machine and picks the largest local model it can actually fit. A research lab shrank the standard computer-use benchmark environment by more than four times. And Meta FAIR published something more useful than another agent: a way to decide which experiment is not worth the GPU hours at all.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Three coding assistants changed their billing or their end date this year — check what you are actually on
If you set up a coding assistant in 2025 and have not looked at the invoice since, three things happened while you were shipping. AWS stopped accepting new signups for Q Developer IDE plugins and paid subscriptions on May 15, 2026, with full end of support on April 30, 2027 — a date worth putting in a calendar rather than discovering. Google closed individual-tier Gemini Code Assist on June 18, 2026; Standard and Enterprise tiers for organizations are still running, so a team account and a personal one now behave differently. And every GitHub Copilot plan moved to AI-credit billing.
That last one is the sneaky one, because the headline price got better. Copilot Pro is $10 a month and includes 1,500 credits, described as $15 of value, with code completions still unlimited on all paid plans. So the completion habit stays free-ish and the agent habit is now metered — which is a different bill for the same person depending on how they work. At the $20 tier, Cursor Pro, Claude Code and Kiro Pro all sit at the same number, verified against official pricing pages on September 3, 2026. The pricing pages agree on the numbers; they just do not agree on what a unit is.
For builders
Do one thing today: open the billing page of every AI coding tool you pay for, and write down two numbers — the monthly price, and what runs out. On Copilot Pro that is 1,500 credits against unlimited completions, which means the honest comparison against a $20 flat-rate tool is not price against price, it is your own agent-invocation count per month. If you are on Amazon Q Developer, the migration deadline is April 30, 2027, and the cheapest time to move a team off a tool is before the tool sets the date for you.
Story II
Nous shipped a local-model installer that reads your hardware first and tells you what will not fit
Hermes Desktop is a free, MIT-licensed desktop app built on the open Hermes Agent framework, and the part worth your attention is the setup flow. It reads your hardware, picks a model that fits it, downloads the weights, and configures the inference runtime — no account required for local operation. It ships a bundled llama.cpp build covering CUDA, Metal, Vulkan, HIP and plain CPU, so the usual afternoon of matching a backend to a GPU is gone.
The honest touch is the catalog. Every model is evaluated against your specific machine before you download it and gets a memory verdict: green means it fits in GPU memory, amber means it will spill into system RAM, red means do not. Quantization picks the highest-quality build that fits your GPU, with a hard floor at 4-bit — it will refuse to make a model fit by wrecking it. Recommended models are guaranteed at least a 64K context window, which is the difference between a toy and something you can point at a file. It runs on macOS 12 and up, Windows 10 and 11, and any Linux distribution, from the desktop page at hermes-agent.nousresearch.com.
For builders
Install it this week for one specific reason: an offline fallback with a known context budget. Run the setup, note which model comes back green on your machine and what its context window is, and you now have a measured answer to "what can I do with no network and no API key?" — instead of a guess. It is also the fastest way to sanity-check whether a cheap hosted model is worth its per-token price for your dullest, highest-volume task.
Story III
Berkeley's CUA-Lite makes the computer-use benchmark cheap enough to actually run
Most people never evaluate their computer-use agents, and the reason is boring: the environments are heavy. CUA-Lite, out of UC Berkeley, puts agents, environments, traces, and the training and evaluation framework into one system with a single action space across desktop, browser and mobile, a unified LiteSample data schema in parquet plus images, and a lite.gym environment interface. It ships more than 10 built-in agents — GPT, Claude, Gemini, Qwen3-VL and others — against more than 15 integrated benchmarks.
The numbers that matter are the resource ones. Lite.OSWorld uses 0.9 GB of memory against OSWorld's 4.1 GB, and cold-starts in 23.8 seconds against 29.9, which works out to roughly 4.6 times more parallel instances on the same box — and scores matched across the 13 models tested, so the shrink is not bought with fidelity. Around it sit more than 30,000 verifiable tasks and 20-plus datasets on Hugging Face, including Aguvis, OpenCUA, ScaleCUA and GUI-360. Setup is Python 3.12 and any Docker host, with no /dev/kvm or nested virtualization needed, installed via uv sync --all-extras. One caveat before you build on it: the repo carries no explicit license yet, so check the terms before any commercial use.
For builders
If you have an agent clicking through a UI in production and no regression suite behind it, this is the cheapest week you will ever get to build one. Start narrow: pick the two or three benchmark tasks from the 15 that resemble your actual workflow, run your current agent, and record the score as a baseline. That single number turns your next model swap from a vibe into a comparison. Just resolve the licensing question before it becomes part of a paid product.
Story IV
Meta FAIR published the part of an AI research agent that decides what not to run
Meta FAIR's Research Preference Models invert the usual framing. An RPM does not predict how well an experiment will score; it ranks unexecuted candidates against each other and picks one to actually run. Inside an evolutionary tree search, the agent generates 15 candidates in parallel and the RPM compares them pairwise in a knockout tournament, using validation scores from already-explored nodes as context, and only the winner burns compute. There are two variants: an inference-only LLM judge optimized with MIPROv2, at 57.7 to 59.0 percent offline accuracy, and an agentic version that also runs sandboxed pilot experiments capped at 30 with a 60-second threshold.
On AIRS-Bench — 20 public text and tabular tasks, 24 hours on a single H200 per task, 10 seeds — the baseline with no RPM scores 0.684, the inference-only variant 0.711, and the agentic variant 0.729. The number I would actually quote in a planning doc is the other one: both variants reached the 24-hour baseline score in roughly 15 hours, a 1.5 to 1.6 times speedup on wall clock. Note the offline accuracy again — 57.7 percent is barely better than a coin flip on any single pair, and it still pays off across a search. The scaffold, AIRA-dojo, is Apache 2.0 and the AIRS-Bench benchmark is open source, with Qwen3.6-27B as the open-weights backbone.
For builders
You do not need a research agent to steal the pattern. If you are running hyperparameter sweeps, prompt variants, or eval configurations on rented GPUs, add a cheap ranking pass before execution: generate your candidate set, have a small model run a pairwise knockout, and only execute the winner. A selector that is right 58 percent of the time on individual pairs still cut wall clock by about a third here, because the cost of the wrong pick is one skipped run, not a wrong answer.