The one-line version: this week Alibaba's Qwen3.7 Flash listed a 1M-token vision-language window at roughly $0.03/$0.13 per million tokens on OpenRouter. The tempting conclusion — stop compacting, just paste everything in — is half right. Cheap context fixes your bill. It does nothing for the rot. You were never only buying tokens; you were buying attention, and attention still doesn't scale with the window.
If you build agents alone, here's the line that decides whether cheap long context lets you delete your context plumbing — or quietly makes your answers worse.
What actually got cheap#
Long context used to be expensive twice: once on the sticker (multimodal, long-window models charged a premium per token) and once on the engineering (you built RAG, compaction, or a memory tool to avoid paying it). Qwen3.7 Flash-class pricing kills the first cost. At three cents per million input tokens, a 200k-token window is about six-tenths of a cent a call. For a lot of teams, that's below the threshold where it's worth writing retrieval code at all.
So the honest first move is: re-price the features you shelved on cost. If you cut a screen-reading, document-parsing, or "read this whole thread and decide" feature because the token math didn't work, run it again — the math changed this week. That part of the hype is real.
The half they're not pricing in#
Here's what a cheaper token doesn't buy you: reliability at length. Models degrade non-uniformly as their input grows, and they start well before the window is full.
Chroma's Context Rot study is the cleanest evidence. Across 18 state-of-the-art models — GPT-4.1, Claude 4, Gemini 2.5, Qwen3 among them — accuracy fell as inputs lengthened on tasks as simple as retrieval and text replication, sometimes 30–50% below the short-context baseline, and crucially not at some clean cliff near the token limit (Chroma). The mechanism is attention: a transformer has a fixed attention budget and spreads it across every token you hand it, so each extra token of context dilutes focus on the ones that matter. This is the whole argument of context rot — a bigger window is not a cleaner one.
A 1M-token window is not a million usable tokens. Cheap tokens make the far-away ones affordable. They don't make them reliable.
Why the advertised number lies to you#
You'll see Qwen3.7 Flash, like every long-context model, quote near-perfect needle-in-a-haystack recall out to its full window. Believe the number and misread it. Needle recall measures match-and-copy — find a verbatim fact, quote it back — which is the single operation that doesn't dilute with length. It's the one corner of the task space where the architecture's strongest primitive does all the work.
Change the test and it falls apart. NoLiMa swaps the verbatim needle for one that matches the question by meaning instead of shared words; models that report >99% recall at a million tokens can fail it by 32k (NoLiMa). RULER adds multi-hop tracing and aggregation and shows most models peak at their shortest setting (RULER). Advertised capacity and effective robustness are different properties — and real agent work (association, multi-hop, aggregation) lives entirely in the gap the needle test skips.
The decision, by task not by budget#
The price change moved the boundary between "dump it in" and "curate it." It didn't erase it. Sort your agent by what it has to do with the far-away tokens:
Dump it in (cheap context wins):
- Single-fact retrieval — locate and quote one span. Match-and-copy doesn't dilute; a big cheap window is simpler than a retrieval pipeline and now costs about the same.
- Short-to-medium windows, simple Q&A — you're below the rot cliff. The compaction engineering costs more than the tokens it saves.
Keep curating (accuracy wins):
- Multi-hop reasoning and aggregation — the degradation is largest exactly here. A full window silently lowers answer quality, and the bill was never the point.
- Long-horizon agents — hours of self-accumulated tool output is the worst-case rot scenario. This is where context editing vs compaction still earns its keep: clear stale tool results, compact the transcript, or offload to a memory tool so the working window stays small even as total history grows.
Note what did not change: this is orthogonal to caching. Prompt caching vs context editing still cut different costs — caching makes the tokens you resend cheaper, editing deletes them so you stop resending — and cheap per-token pricing doesn't retire either. And the RAG vs long context call is now genuinely closer for copy-shaped tasks and unchanged for reasoning-shaped ones.
What to actually do this week#
- Re-price shelved features. Any multimodal or long-context feature you cut on cost gets a fresh look. The floor dropped.
- Don't rip out curation to save tokens. You were buying attention, not just tokens, and attention doesn't get cheaper when the window does.
- Measure your own rot before you trust a window. Vendor needle numbers won't tell you where your agent breaks — a short eval will. We wrote the 20-minute version: how to tell if your agent has context rot.
Cheap 1M context is a real gift for the copy-shaped half of your workload. For the reasoning-shaped half, the oldest rule in context engineering still holds: the lever is a cleaner window, not a bigger one.
