Long documents

Kimi K3 vs Claude Opus 5
for long document questions

This page compares two current models on one job: answering questions across a large set of documents. It looks at retrieval deep in the context, multi-hop reasoning, source attribution, throughput and cost per run, and a fair way to test both on your corpus.

Jul 29, 2026 · 11 min read

The bottom line
Opus 5 defends and K3 scales

Claude Opus 5 is the safer default for high-stakes questions over a large document set, because of documented source-grounding support and stronger current independent reasoning evidence. Kimi K3 is the stronger cost-first choice, especially when the same corpus gets queried again and again.

The money is concrete. One request with a million input tokens and a short answer runs about $3.03 on K3 against $5.05 on Opus 5, and a cached repeat is about $0.33 against $0.5549. At a few hundred questions a month that difference is a budget line, not a rounding error.

The quality claim is softer, and worth naming as soft. The K3 long-context results that look strongest were reported by its own vendor against Claude Opus 4.8 rather than Opus 52, and Anthropic's consistency claim for Opus 5 comes without a published retrieval curve8. For staged work, use K3 for broad extraction and candidate evidence, then Opus 5 for the final synthesis and source check, as long as the handoff carries document IDs and the quoted evidence rather than a summary.

Who this is for
Which document work this fits

Start with Opus 501

Legal and due diligence

A missed qualification or a citation that supports half a sentence is the expensive failure. The API that returns the document and character index it used saves you building that audit trail by hand.

Check the tables02

Finance and reporting

Your questions mix prose with tables and need an exact number out the other side. Test that case specifically, because a model can retrieve the right page and still compute the wrong figure.

Start with K303

Research and policy

You query the same corpus continuously and a person reads every answer. Automatic prefix caching at $0.30 per million makes the hundredth question cost a fraction of the first.

Use both04

Support and operations

Volume is high and the stakes vary by ticket. Extract cheaply with one model, then send anything consequential to the other for a second read with the quotations attached.

What we compared
Retrieval not the app

This page compares the two models through their API in one neutral setup, not one model inside one document product against the other inside another.

The parts that matter are finding a fact buried deep in the context, reasoning across several documents, picking the controlling version when documents contradict each other, attaching an answer to the passage that supports it, and the cost of a large run. Throughput matters too once answers get long.

We left tools out of the spec table on purpose. Upload interfaces, retrieval layers and vendor research applications belong to the app around the model, so the same model behaves differently in a chat product, in the API or inside a workspace. Judging those here would compare wrappers, not reading.

Specs at a glance
What a million tokens costs

The model facts that actually affect a large-corpus question. Tool features are left out, since they change with the app around the model.

Spec
Kimi K3
Claude Opus 5
Why it matters
Context window
1,048,576 tokens
1,000,000 tokens
Both hold a very large corpus in one request38
Input price
$3 per million on a cache miss
$5 per million
The input side dominates the bill on a big document set49
Output price
$15 per million
$25 per million
Matters once answers carry quotations and reasoning49
Cached input
$0.30 per million, cached automatically on an unchanged prefix
$0.50 per million to read, $6.25 or $10 to write
Repeated questions over one corpus get much cheaper on both459
Long-context tier
None, one rate across the window
None, standard rates across the window
Neither model charges a premium for using the full window38
Source attribution
Quotes and IDs via a schema you define
Document, page, block and character indices with cited text
Opus 5 returns the location itself rather than writing it in prose310
Reasoning effort
Low, high or max
Adaptive thinking with adjustable effort
A per-call setting to tune by evaluation, not a fixed trait311

Figures from Moonshot AI and Anthropic documentation, checked July 2026. The cost examples above assume one million input tokens and a 2,000-token answer and exclude prompt, markup, reasoning and citation tokens.

Head to head
Defensibility against cost

The answer changes by what a wrong answer costs you. Read the evidence column closely here: several rows rest on cross-version results, because these two versions have barely been tested against each other.

Job
Better choice
Why the edge exists
Best evidence
Retrieving a buried fact
Too close to call
The best K3 long-context number was reported by its vendor against Claude Opus 4.8, which is a different model from the one on this page. Anthropic says Opus 5 holds up across the full window but has not published a retrieval curve that would settle it.
K3 scored 74.7 against Opus 4.8's 67.7 in vendor-reported results2
Multi-hop reasoning late in the context
Claude Opus 5, provisional
Independent work before Opus 5 found the Anthropic line relatively stable through 512,000 tokens and still solving some three-hop chains at one million. A separate long-context suite also ranked an earlier Opus first at the million-token scope. Both are cross-version, so they show a lineage rather than a result.
Cross-version studies put earlier Opus models ahead at 1M tokens1213
Questions over a real document collection
No demonstrated winner
A benchmark that hands the model a whole PDF corpus and requires parsing, retrieval and analysis put the two within a point of each other, but again against Opus 4.8 and with the vendor assembling the comparison.
K3 scored 63.3 against Opus 4.8's 63.9 on OfficeQA Pro214
Pointing at the right passage
Claude Opus 5
Its API returns document, page, block or character indices plus the cited text, and Anthropic states the pointer refers to the supplied document content. That is stronger than asking a model to write citation strings into prose and hoping they line up.
Anthropic documents the citation pointer format and behaviour10
General reasoning and answer judgment
Claude Opus 5
The one current exact-version independent comparison puts Opus 5 ahead on a broad index covering long-context reasoning, knowledge reliability, science, coding and agentic work. It is wider than document QA, and it is the only like-for-like number available.
Opus 5 at xhigh effort scored 60 against K3's 571
Cost of a fresh large run
Kimi K3
On published rates a million-token question with a short answer is about two dollars cheaper on K3, before any of the ancillary tokens a real workflow adds. At volume that is the whole argument for K3.
About $3.03 against $5.05 at published rates49
Repeated questions on one corpus
Kimi K3
A full-prefix cache hit plus a short answer lands near $0.33 on K3 against $0.55 on Opus 5. K3 caches automatically while the long prefix is unchanged, where Anthropic caching is configured explicitly and charges more to write the cache first.
Cache-read rates of $0.30 against $0.50 per million459
Output speed
Claude Opus 5
Measured throughput after generation starts was around half again as fast on Opus 5 in one independent measurement. Provider and load conditions move these numbers, so treat it as an observation rather than a specification.
About 55 output tokens per second against 341

Better-choice calls map to what the sources actually evaluated. Three rows rest on comparisons against Claude Opus 4.8 rather than Opus 5, and the rows say so, which is why no retrieval winner is declared.

How to test
A position-controlled test

The test that matters here is about where the evidence sits, not how clever the question is. Then judge what your team actually pays for: did it find every hop, did it pick the controlling document version, did it point at the right passage, and did it say when the documents do not answer the question.

Sample01

Build five placed questions

One fact buried at 90 to 95 percent depth, one needing evidence from three documents in different parts of the context, one conflict between an original policy and a later amendment, one whose correct answer is that the documents do not establish it, and one table-plus-prose question needing an exact calculation.

Prompt02

Demand evidence in the schema

Same ordered source material, same prompt and same effort level on both sides, with the same evidence demanded either way: document IDs, short supporting quotations and source positions. Claude's citation pointers cannot run in the same request as a forced JSON schema, so ask K3 for that evidence inside its JSON answer and ask Opus 5 for it in a short prose answer with citations turned on, then check both against the same list.

Setup03

Match effort and allowance

Set the same reasoning level and the same output allowance on each API, and run both where the team will deploy. API and chat results differ because system prompts, retrieval layers and context management differ.

Scoring04

Score each claim to its quote

Do not edit before scoring. Check every claim against the passage it cites, count missed hops, wrong document versions and unsupported statements, and note how often the model correctly abstained. Use blind review for anything commercial.

What the evidence shows
Usable context is not window size

The single most useful finding across these sources is that an advertised million tokens is not a million equally usable tokens. Here is what each source helps judge.

Source
What it measures
What it suggests
How to weigh it
AA exact-version comparison
A broad capability index plus measured throughput
Opus 5 ahead on the index at xhigh effort and faster on output
The only like-for-like source, and broader than document QA1
Moonshot benchmark table
Long-context reasoning and a full-corpus PDF benchmark
K3 strong on long context and level on the corpus benchmark
Vendor-assembled and measured against Opus 4.8, not Opus 52
Million-token multi-hop study
Retrieval against three-hop reasoning at increasing depth, on classical Chinese text
Needle retrieval looks solved while multi-hop falls away past 512K
A narrow text domain, so read the shape of the curve not the scores12
ATLAS long-context suite
How rankings change as input length grows
Retrieval strength does not always carry over to answering
Its models predate both of these, so it supplies method not verdict13
K3 technical report
How the long-context training was built
Training scattered required facts across the full sequence
Design evidence, and not proof of production retrieval accuracy7

K3 launched on July 16 and Opus 5 on July 24, 2026, so most relevant results compare one of them against an older model. Prices, throughput and leaderboard positions can all move quickly.

How to prompt each one
Demand evidence objects

The best prompt is not the same for both, and one rule holds for both: make the model return the evidence in a shape you can check automatically.

Kimi K3 does best with explicit boundaries, because Moonshot warns it can act too proactively, and with a rigid evidence schema. Its API expects the full reasoning history to be preserved between turns, so dropping the returned assistant message can destabilise later answers in a multi-turn session6. Ask for an answer plus document IDs, exact quotations, source locations, conflicts and a confidence value, and require an insufficient-evidence result when a hop has no support.

Claude Opus 5 does best with a tight scope and a stated answer length. Anthropic notes it can over-verify or widen the requested scope when a prompt carries redundant workflow instructions11. Ask it to identify the minimum evidence chain first, then answer within a word limit with a citation on every material claim, and to separate direct evidence from inference and from an unresolved conflict.

A Kimi K3 prompt: a strict evidence object per claim

Answer only from the supplied documents.

Return JSON with:
  answer
  document_ids
  exact_supporting_quotes
  source_locations
  reasoning_summary
  conflicts
  confidence

Resolve amendments by effective date.
If any hop lacks evidence, return insufficient_evidence.
Do not broaden the task.

A Claude Opus 5 prompt: minimum chain then a short answer

Answer using only these documents.

First identify the minimum evidence chain needed.

Then answer in no more than 200 words, with a citation
attached to every material claim.

Distinguish direct evidence, inference and unresolved
conflict. Do not investigate anything outside the question.

Weak spots
Where the answers break

The shared failure is the dangerous one: an answer that looks cited and is only partly supported. The useful question is what to change in the prompt or the pipeline.

Model
Weak spot
What it looks like
How to fix it
Kimi K3
Evidence is vendor-reported
The strongest public numbers come from its own launch material and are measured against an older Claude, so they may not describe how it behaves against Opus 5 on your corpus.
Verify late-window retrieval on your own position-controlled set before committing, and treat the launch table as a reason to test rather than a result27.
Kimi K3
Multi-turn state is fragile
Later answers drift or contradict earlier ones when the returned reasoning history is not passed back, and it can widen the task on its own.
Preserve the complete assistant message between turns, forbid unstated assumptions, and require one evidence object per claim6.
Claude Opus 5
Long answers cost twice
It over-verifies or produces a longer report than asked, and its higher input and output rates make that verbosity expensive rather than merely annoying.
State the answer length, the evidence count and the stopping condition. Drop duplicated double-check instructions and start at high effort rather than defaulting to maximum11.
Both
Citations cover part of a claim
A pointer that supports half of a sentence, or a confident synthesis built from incomplete evidence. Neither a big window nor a high reasoning setting prevents this.
Score each claim against its cited passage, add an abstention option, and run a second-pass citation verifier. Keep document IDs and version metadata inside every source block10.

Which one to choose
Start from the cost of error

One question first. What is more expensive, a missed qualification in the answer or the extra model cost? Then follow the branch that matches most of your work.

What does a wrong answer cost? Legal or financial consequence Hundreds of reviewed queries Need API source pointers Discovery before verification Three hops late in the context Claude Opus 5 Kimi K3 Claude Opus 5 Kimi K3 Test both from Opus 5 Retrieve passages first

A starting point, not a rule. Run the placed-question test on your own corpus.

Recommendations
Pick by risk then by volume

If a wrong answer carries legal, financial, scientific or compliance consequences, choose Claude Opus 5. The same applies when exact source pointers have to come back through a documented API mechanism rather than being written into prose110.

If the corpus gets queried hundreds of times and a person reviews the answers, choose Kimi K3 and let the automatic prefix cache do the work. It is also the better first stage for broad discovery or extraction that a separate verification step will check45.

If the question needs three or more hops spread across the final third of a near-million-token input, test both and start production evaluation with Opus 5 as the safer hypothesis1213. And if neither model reaches the recall you need, do not just raise the reasoning effort: retrieve candidate passages first and hand the model those passages plus their surrounding context.

One case sits outside all of this: if the same corpus has to be indexed once and then queried by many people repeatedly, with permissions per document, that is a retrieval system rather than a chat workspace. Playgram is the right place to work out which model reads your material well, before you build that around it.

Bottom line
Opus 5 for the final answer

Pick Claude Opus 5 when the defensibility of the answer matters most. Pick Kimi K3 when the cost per large-context query is the governing constraint. K3's cheaper run is real and meaningful at scale, and it does not by itself prove better usable context.

The evidence is uneven in a specific way worth remembering: most relevant results compare one of these models against Claude Opus 4.8, not against each other21213. Both launched in July 2026, vendor benchmark configurations differ, and prices, throughput and behaviour can all move. A controlled evaluation on your own documents is part of the buying decision here, not an optional last check.

The safest final step is to test the shape of your own corpus, not a generic prompt from the internet. A fair test needs the same setup for both models: the same source material, the same prompt and the same place to run them, so the result reflects the models and not the tool around them. In practice that is harder than it sounds, since most teams end up running one model in one app and the other in a different one, on two separate subscriptions, which tilts the comparison before the first answer comes back. The cleaner the setup, the more the difference you see is really Kimi K3 vs Claude Opus 5, and not just which one happened to be easier to reach that day.

Query one corpus twice
Right here inside Playgram

That's the practical case for the setup just described, and it is also what makes a verification stage cheap to run. When both models sit in one workspace, you can ask the same question of each, compare the two answers and their quoted evidence side by side, and hand a first-pass extraction to the other model for checking without loading the documents again.

Playgram lets you run that same comparison directly: upload the document set once, put it in front of the latest Kimi and Claude models, and keep the conversation going with either one without re-uploading anything or starting over for the second opinion.

The same memory carries across the team too, not just this one comparison, over the latest GPT, Claude, Gemini and Grok models and many more, all in one place15. The line-up is curated, so retired models are turned off and new ones are added as they ship.

Team memory

Shared across everyone and every model.

Project memory

Scoped to a campaign or document set.

Personal memory

Your own working style, kept private.

Fair pricing
Pay per usage, not per seat

Upgrade as needed, and only pay for what you actually use

Save ~17% with the annual plan

Pro

$50/ month

Perfect for small and medium teams

Unlimited users & infinite memory

Multi-LLM chats

Unlimited use of DeepSeek V4 Flash

US & EU data residency

Get started

Enterprise

Get in touch

Unlimited Credits

For organizations with advanced needs

Unlimited users & SSO

Priority Support

Unlimited use of DeepSeek V4 Flash

US & EU data residency

Book a call

30-days money back guarantee

Frequently asked
questions

At published rates, a single request with one million input tokens and a 2,000-token answer works out at roughly $3.03 on Kimi K3 and $5.05 on Claude Opus 5. If the same corpus prefix is cached, the repeat question drops to about $0.33 on K3 and $0.55 on Opus 5. Those figures exclude the extra tokens your prompt, document markup, reasoning and citations consume, so treat them as a floor rather than a quote.

Nobody can say yet, and that is the honest answer. The encouraging K3 long-context numbers were reported by Moonshot against Claude Opus 4.8, not Opus 5. Anthropic states that Opus 5 holds its reasoning and instruction following across the full window but has not published a depth-by-depth retrieval curve for it. Until someone runs a position-controlled test on both exact versions, build your own.

Claude Opus 5, and it is a design difference rather than a matter of degree. Its API can return the document, page, block or character index plus the cited text, so an auditable answer does not depend on the model writing citation strings into prose. On K3 you get the same effect by defining a strict output schema with document IDs and exact quotations, which works well but puts the verification burden on you.

No, and this is the most useful thing to know before you design the workflow. Independent long-context research finds that simple needle retrieval can look solved while multi-hop reasoning falls away sharply between 512,000 and one million tokens, and that model rankings change with input length. If your questions need three hops across the last third of a near-million-token input, expect to retrieve candidate passages first rather than feeding everything in.

That is the staged pattern, and it works if the handoff keeps the evidence. Use K3 for broad extraction and candidate evidence, then Opus 5 for the final synthesis and source verification, passing document IDs and the exact quotations through rather than a prose summary. Without that, the second model is just re-reading someone else's conclusion.

No. Run the same prompt on both and compare the answers, or switch between them mid-conversation. You choose after reading both answers instead of guessing up front.

Related comparisons

GPT-5.5 vs DeepSeek V4 Pro for summarizing documentsClaude Opus 4.8 vs GPT-5.5 for legal document reviewClaude Opus 4.8 vs Gemini 3.1 Pro for research reportsKimi K3 vs GPT-5.6 Sol for fact-checking drafts

One corpus and two answers
One place to check the source

Put the same document set in front of the latest Kimi and Claude models, keep the context in one place, and see which answer you would actually defend. Set it up in a minute.

Get startedSee the pricing