Effective Context
The Number on the Box
Qwen 3.8 Max markets an extensible 1M-token window. It scores 92.3% at 128K and 29.7% at 512K on Context Arena's retrieval test. The window is real; recall at range is a separate question.
Marketing collapses two things into one number. Advertised context is how many tokens a model accepts. Effective context is how much of that input it still uses reliably under load.
So compare degradation curves, not window sizes. The curves below come from one independent benchmark: Context Arena runs every model through the same long-context test with one scoring function. Models with no independent coverage aren't plotted.
LLM mechanics
Effective context is smaller than advertised context
Effective-context retrieval by model, measured on Context Arena (MRCR v2, 8-needle).
Top contenders
Second tier
Reading the Curves
Recall is a property of the whole system, not a model-card number. Architecture and training determine what a model can distinguish; your harness sets the conditions: thinking effort, context length, and how much similar material it must resolve. The chart shows the result of that combination.
The position-based attention curve in Context Engineering is the mechanism — how the model's focus is spread. This recall curve is the measured outcome: it hides an answer somewhere in the window and demands the model find it and reproduce it, so it probes the whole range rather than the two bright ends, and it falls as more content lands where attention is thin. Reproduction is strict — a model that can't quote a passage exactly can still summarize or extract it, so a low score is a floor on one capability, not proof it can't use its context.
The shape is the signal. A flat line sustains precise retrieval; a slope exposes its limit; a cliff marks where context engineering must keep the task in a reliable range—or where the model belongs on less recall-dependent work. Qwen 3.8 Max measures 92.3% at 128K, 67.7% at 256K, then 29.7% at 512K. A low start means the skill is unproven. A line ending at 512K, as with GPT-5.6 Sol and Claude Opus 4.6, is unmeasured at 1M—not a failed score.
Models are not databases: they must preserve identity, position, and relationships well enough to answer later. That is why giving a model more context can reduce recall instead of improving it. Context Arena makes those outcomes comparable under one task and scoring rule. An agent retrieving an old decision, API contract, tool result, file detail, or user constraint needs a curve that holds at its operating length. That is its effective context limit—not the API maximum. It is one input to Choosing a Production Model, alongside task quality, latency, and cost.
The Test: Eight Poems About Penguins
MRCR v2 builds a long synthetic conversation: "Write a poem about penguins" appears eight times, each answered with a different poem among near-identical distractors. Then:
Prepend the string AKJSs89sal to the fifth poem about penguins. Do not include any other text.
The model finds all eight, orders them, selects the fifth, and reproduces it, scored by string similarity with partial credit. Miss the hash prefix and it's zero. This isn't needle-in-a-haystack: NIAH hides one conspicuous fact in filler; MRCR hides eight near-identical candidates and demands ordered discrimination — closer to agentic context, where the hard part is picking the right one.
Here's what that score actually means. The model isn't solving anything — it's echoing back text it was just handed, and copying is the easiest job you can give an LLM. It still never gets it perfectly right, because an LLM doesn't really copy: it guesses, word by word, what probably comes next, and reproducing a poem is a long chain of those guesses — every single one has to be right (local prediction, global consequences). One wrong guess anywhere changes the answer, and the hash prefix is exactly where a wrong guess is most likely: a short, odd, meaningless string with nothing to anchor it. That's not a flaw in the test; it's the cost of working with a machine that predicts instead of recalls. The test only looks pristine because it runs in a clean room — nothing else in the window, so the model gets the front, the strongest part of its attention. Real windows are never that tidy: by the time the system prompt, tool definitions, and project notes have claimed the front, your task sits further back in the weaker middle. You can't keep the machine from guessing; you can only set up the room so the guess lands — which is what context engineering is for.
Who Grades the Homework
Everyone runs MRCR differently — their own effort modes, anchors, averaging — so a cumulative average up to 128K is a different quantity from a pointwise score at 128K. Several labs publish no length-stratified eval at all. MRCR is one corner of a landscape — OAI-MRCR, GraphWalks, RULER, LongMemEval ask other questions — so pick the benchmark that matches the workload.
For MRCR, Context Arena runs every model through the same DeepMind MRCR v2 harness and publishes the curves and code. One harness, one scoring function: what's left in the data is the models. Read vendor cards the same way: a vendor number often isn't even the same quantity as the measured one. Google's reproduces — 84.9% cumulative @128K and 26.3% @1M come back as 85.1% and 25.9% — while Anthropic reports 91.9% @256K where the arena measures 69.1%. Treat numbers that exist only under vendor conventions as claims, and select on measured curves.
The Takeaway
Every model has an effective context limit, and it's a curve, not a number. If your workload routinely reaches 512K–1M, measured evidence narrows the shortlist; below 128K, compare the measured curve with task quality, latency, and cost — several compact and specialized models already fail there. None of it makes context management obsolete: Grounding covers retrieval and RAG as system levers; Context Engineering covers the attention-shape problem operationally.