Long-Context Model Selection: What 1M Tokens Really Mean

Short answer. A 1,048,576-token context window is now table stakes: eleven of the twenty-one tracked models record it, spanning five of six vendors and both flagship and flash tiers. But the input window is the wrong number to optimize — the output ceiling is the honest economic parameter. Output limits range from 131,072 (GLM, Qwen) to 1,048,576 (Kimi K3), a spread that decides whether a model can generate a long document or merely read one. For long-document analysis, any 1M model works and the flash trio (DeepSeek/Qwen/GLM at $0.15) is the cost-optimal choice; for generating output longer than ~100K tokens, only DeepSeek (393K) and Kimi K3 (1M) publish enough headroom.

Decision criteria

Criteria	Relevance / Notes
Context window	1M on 11 models; 256K mid tier (Doubao Turbo, Kimi K2.x, Qwen A95B); 200K legacy (MiniMax M2.7)
Maximum output	DeepSeek 393,216; Doubao 262,144; GLM/Qwen 131,072; Kimi K3 1,048,576 — the real differentiator
Cache-hit pricing	$0.003 (DeepSeek) to $0.30 (Kimi K3) — rewards stateful re-reading of a large context
Open-weight 1M option	Only GLM-5.3 and GLM-5.3-Flash — for self-hosters
Flash-tier 1M at $0.15	DeepSeek V4.1-Flash, Qwen3.8-Flash, GLM-5.3-Flash — long context no longer costs a premium
Input/output asymmetry	A 1M-input model with 131K output is a long-input analysis tool, not a long-form generator

The 1M-context lineup

Model	Input	Output	Price (input/output)	Open?
Kimi K3	1,048,576	1,048,576	$3.00 / $15.00	Yes
DeepSeek-V4.1-Flash	1,048,576	393,216	$0.15 / $0.60	Yes
DeepSeek-V4-Pro	1,048,576	393,216	$0.66 / $1.98	Yes
Doubao Seed 2.1 Pro	1,048,576	262,144	¥6 / ¥30 (CNY)	No
GLM-5.2	1,048,576	163,840	$1.40 / $4.40	Yes
GLM-5.3	1,048,576	131,072	$1.40 / $4.40	Yes
GLM-5.3-Flash	1,048,576	131,072	$0.15 / $0.50	Yes
MiniMax-M3	1,048,576	not published	$0.30 / $1.20	Yes
Qwen3.8-Max	1,048,576	131,072	$2.00 / $6.00	No
Qwen3.8-Flash	1,048,576	131,072	$0.15 / $0.47	No

Entity routing

Route by whether the deliverable is a summary or a generated document.

Scenario	Best-documented fit	Why
Long-document analysis at minimum cost	DeepSeek-V4.1-Flash	1M/393K, $0.15, $0.003 cache
Long-form generation (only 1M output)	Kimi K3	1M/1M, $3/$15
High-output generation below 1M	DeepSeek-V4-Pro or Doubao Seed 2.1 Pro	393K / 262K output
Self-hosted long context	GLM-5.3 or GLM-5.3-Flash	Open-weight 1M, Apache-2.0
Agent workloads, re-read-heavy context	Flash trio	1M + cache discounts = agent-economics sweet spot
Latency-sensitive chat, context tradeable	Doubao Seed 2.1 Turbo	256K low-latency tier

Relevant comparisons: DeepSeek-V4-Pro vs Kimi K3, Kimi K3 vs Qwen3.8-Max, Qwen3.8-Max vs GLM-5.3.

What the evidence shows

The 1M-input norm and the 131K–1M output spread encode an economic truth: input context is cheap to advertise, output tokens are expensive to serve. Vendors converge on 1M input because the marketing threshold is now there; they diverge on output because output tokens bill at 2–5x input rates and directly consume inference capacity. The output limit is therefore the honest parameter, and Kimi K3’s 1M/1M pairing is unique in the collection — it reframes the $3/$15 price, since per-token output pricing on a 1M-output model implies a different cost ceiling for a single request than a 131K-output competitor.

Context tiering now maps to product tiering, not vendor capability. The 256K/200K models (Doubao Turbo, Kimi K2.x, MiniMax M2.7, Qwen A95B) are the cheap or legacy tiers of vendors whose flagships are 1M — not capability failures. Only ByteDance maintains a current-generation model (Turbo) below 1M, deliberately, as its low-latency tier — context and latency are still a trade in production, even if the spec sheet says otherwise.

The open-weight long-context list is short by design: GLM-5.3 and GLM-5.3-Flash are the only open-weight 1M-context models, while Alibaba’s open A95B remains at 256K, so the open long-context crown currently sits with Zhipu. China AI Hub analysis indicates the flash-tier 1M wave — DeepSeek V4.1-Flash, Qwen3.8-Flash and GLM-5.3-Flash all shipping 1M at the $0.15 floor — is the more consequential shift: long context no longer costs extra at the entry tier, which is why agent tooling has standardized on these three models.

Selection procedure

Work through these steps in order.

1. Classify the deliverable first. Is the output a summary/analysis or a generated document? If the deliverable is a summary, any of the ten 1M models works and the 131K output tier is ample. If it is a generated report longer than ~100K tokens, only DeepSeek (393K) and Kimi K3 (1M) publish enough headroom.
2. Decide cost versus output. The flash trio (DeepSeek-V4.1-Flash, Qwen3.8-Flash, GLM-5.3-Flash) delivers 1M context at $0.15 input; Kimi K3’s 1M output costs $3/$15. Match the spend to the generation requirement, not the input requirement.
3. If the workload re-reads a large context, optimize cache. Agent workloads that hold a 1M window across many turns should favor DeepSeek’s $0.003 cache or the flash trio generally — the cached portion of input is nearly free.
4. If you must self-host, restrict to Zhipu. GLM-5.3 and GLM-5.3-Flash are the only open-weight 1M-context models.
5. Do not assume 1M means usable 1M. The vendor claim is not independently stress-tested; models can degrade at the far end of a window, so validate effective context on your own documents.

How context and output interact in practice

The relationship between the input window and the output ceiling is the single most misunderstood point in long-context selection. A model with a 1M input and a 131K output — GLM-5.3, Qwen3.8-Max, GLM-5.3-Flash, Qwen3.8-Flash — can hold roughly 700,000–800,000 English words of working material in one request, but can only emit about a fifth of that back. That is a long-input analysis tool: ideal for multi-document review, whole-codebase understanding, or long agent traces where the output is a summary or a decision. It is not a long-form generation tool.

The distinction is economic, not just semantic. Output tokens bill at 2–5x input rates, so a model whose ceiling is 131K caps the cost of any single request at a lower bound than a 1M-output model — but also caps the deliverable. Kimi K3’s 1M/1M pairing is the only one in the collection that can generate a million tokens in one request, and its $3/$15 pricing implies a genuinely different per-request cost ceiling: a fully-fledged 1M-token generation on K3 costs multiples of what the same request on a 131K model would, because the 131K model simply cannot do it. China AI Hub analysis indicates this is why the output limit, not the input window, is the honest parameter to put in a decision matrix — the input window is now a commodity, while output headroom remains the scarce, priced resource.

Limitations

Maximum output is recorded for ten of the models; nine do not publish an output-token limit, which the database records as absent rather than unlimited. The “1M token” input is the vendor’s documented claim and is not independently stress-tested for effective usable context (models can degrade at the far end of a window). Cache-hit economics depend on workload statefulness, which varies. ByteDance’s long-context models are CNY-priced and cn-beijing-served, so international buyers must account for currency and region. MiniMax-M3 lists 1M context but does not publish an output limit, which constrains long-form generation planning.

Sources

• China AI Hub — Models database
• DeepSeek API docs — Models & Pricing
• Kimi API — pricing (chat)

Labels used above: Official fact (context windows, output limits and pricing from primary sources), Vendor-reported claim (the 1M-token context statements published by vendors), and China AI Hub analysis (our synthesis, introduced as such). No third-party evaluation evidence is currently recorded for effective long-context performance.