Skip to content
TEN Brief Ten verified stories a day 2026.08.14 KO

이 기사는 한국어로도 읽을 수 있습니다 →

Tech · 4 min read · Reference

What is a context window — a model does not read all million tokens

A context window is the maximum number of tokens a language model can handle in a single request, counting input and output together — and the advertised length is not the same as the length across which the model can accurately retrieve information

A sunlit library reading room, long tables and a reader seen from behind

The three lines

  • Context window = total tokens visible in one request, input and output combined
  • Most leading 2026 models converged on 1 million tokens, but retrieval accuracy degrades markedly between the 200K and 1M ranges
  • The distinction to hold onto is advertised context length versus effective context length

Key questions

What is a context window?
The maximum amount of text a language model can consider at once in a single request, measured in tokens — roughly word-fragment sized chunks. The important detail is that the limit covers input and output together. If a model has a 1M-token window and you have already supplied 990,000 tokens, only 10,000 remain for the answer. In a conversation, the entire prior exchange is resubmitted as input on every turn, so the window fills continuously.
How much text is a million tokens?
It depends on the tokenizer and the language. English runs roughly four characters per token; languages such as Korean typically consume more tokens for the same content, because tokenizers are usually optimised on English-heavy corpora. So a 'million tokens equals N books' conversion changes substantially by language. In practice, reading the token usage returned by the API is more reliable than estimating from character counts.
Is a bigger window always better?
It carries three costs. Price: most APIs bill by input token, so filling the window raises the cost of each request. Latency: longer inputs increase time to first response. And accuracy: a model does not attend evenly across a large window. Retrieval of specific facts degrades toward the far end of the advertised length, a pattern observed consistently across frontier models.
How do you check the effective context length?
Through benchmarks. The best-known family is needle-in-a-haystack testing, which plants a specific sentence inside a long document and measures whether the model finds it. Single-needle tests score generously, so multi-needle retrieval — typically eight planted facts — is closer to production reality. Suites such as RULER measure not only retrieval but reasoning over long context. The vendor's published context length is a specification; these benchmarks are the measurement.

One number appears in every model announcement now. One million tokens of context.

By 2026 it has effectively become the standard. Most leading models arrived at the same figure, and some went past it.

The benchmarks tell a different story. Accepting a million tokens and finding a specific sentence inside a million tokens are separate capabilities.

1. The definition

A context window is the maximum amount of text a model can consider at once in a single request, measured in tokens — roughly word-fragment sized chunks.

The most frequently missed detail:

The window is not an input allowance. It is the total for input and output combined.

Supply 990,000 tokens to a 1M-token model and 10,000 remain for the answer. Ask for a long document and a long response, and one of them gets cut.

Conversational products add a second effect. The model does not remember previous turns. Each new question resubmits the entire conversation as input. The window fills as the conversation grows, and once the limit is reached, the earliest content falls out. That is usually what is happening when a chatbot "forgets what you said."

2. Tokens are counted differently by language

LanguageRough behaviour
English~4 characters per token
Korean, and many non-Latin scriptsMore tokens for the same content

Tokenizers are generally built on English-heavy corpora, so text in other scripts fragments into more tokens.

Two practical consequences. The same document fills the window faster in Korean. And the same task costs more in Korean, since most APIs bill per token.

The exact multiplier differs by tokenizer. Reading the usage figures the API returns beats estimating from character counts.

3. Three costs of a bigger window

CostDetail
PriceBilled per input token — filling the window raises per-request cost
LatencyLonger input increases time to first token
AccuracyRetrieval degrades toward the far end of the window

The first two are predictable. The third is the problem.

4. Advertised length versus effective length

The consistent 2026 finding is this: accuracy holds well over the early portion of an advertised window and falls substantially as inputs approach the upper limit, particularly between the 200K and 1M ranges.

One benchmark summary put the drop at 30 to 60 points. Per-model source data was not cross-checked, so treat that figure as indicative. That a drop exists is agreed across sources.

Two terms are worth separating.

TermMeaningSource
Context lengthMaximum tokens the model acceptsVendor specification
Effective context lengthTokens the model reliably usesBenchmark measurement

Vendors publish the first. The second has to be measured.

The hard problem in 2026 is no longer capacity. Building a model that accepts ten million tokens is far easier than making it locate one specific fact inside them without inventing an answer.

5. Measuring it — three benchmark types

BenchmarkWhat it measuresCaveat
Needle in a Haystack (NIAH)One planted sentence inside a long documentSingle-needle scores are generous
Multi-needle retrievalTypically eight planted factsMuch closer to production
RULER-style suitesRetrieval and reasoning over long contextTests whether the model can use what it found

The middle row matters most. Single-needle scores mislead because finding one sentence in one document is not a real task. The real task is loading an entire codebase and correcting all eight related functions — and there, models fail by finding some and reporting the rest as absent.

The silence is the danger. The model does not say it missed anything; it produces a plausible answer from the subset it retrieved.

6. Working practice

Do not fill the window because you can. Irrelevant material raises cost and latency while lowering accuracy.

Put critical instructions at the beginning or the end. Information placed in the middle of a long input is the most frequently ignored, an observation reproduced repeatedly.

Keep retrieval in front of the model. Selecting the relevant passages instead of dumping everything into the window — the RAG pattern — did not become unnecessary when windows grew. It became cheaper to justify.

Summarise long conversations as you go. Since the entire history is resubmitted each turn, periodic compression improves both cost and accuracy.

7. The 2026 landscape

Model tierContext (mid-2026)
Most leading models1,000,000 tokens
Some models500,000 tokens
Some experimental modelsBeyond

Convergence on one figure means context length has stopped being a differentiator. DeepSeek V4-Pro, which reached general availability on August 13, 2026, also lists 1 million tokens.

The competition has moved to the next question: who retrieves most accurately within the same million.

8. What is unresolved

The 30–60 point retrieval drop cited above comes from a single benchmark summary; per-model source data was not checked. The claim that one particular model sustains accuracy across its full window comes from the same summary, with no independent reproduction found.

No token multiplier is given for non-English languages, because it varies by tokenizer.

Context lengths quoted here reflect mid-2026 and change with each version. What stays useful in this piece is not the table but the habit of separating advertised length from effective length.

Sources

  1. Morph — LLM context window comparison (2026): 20 models from 200K to 10M tokens
  2. Digital Applied — Long-context retrieval 2026: needle-in-haystack test
  3. DataNorth — LLM context length & context window explained (2026)
  4. Zylos Research — LLM context window management and long-context strategies 2026
  5. DevTk.AI — LLM context windows explained: 4K to 1M tokens (2026)

Verification

Published
Last modified
Cross-check
Checked against 5 independent sources.
Unverified
  • The figure of a 30–60 point retrieval drop between 200K and 1M tokens comes from one benchmark summary and was not cross-checked against per-model source data
  • The claim that one particular model maintains retrieval across the full window is from the same summary and no independent reproduction was found
  • No specific token-per-character multiplier is given for non-English languages, as this varies by tokenizer
  • Model context lengths cited reflect mid-2026 and change with version updates
Authoring
Reviewed by a person before publication. The full process is described in the Editorial.

Ten stories, once each morning

We send the three-line summaries only; the full pieces stay on the site. One-click unsubscribe, any time.

Related