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
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 $0.554, 9. 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.
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.
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.
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.
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.
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.
The model facts that actually affect a large-corpus question. Tool features are left out, since they change with the app around the model.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.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.
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.
A starting point, not a rule. Run the placed-question test on your own corpus.
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 prose1, 10.
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 check4, 5.
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 hypothesis12, 13. 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.
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 other2, 12, 13. 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.
Upgrade as needed, and only pay for what you actually use
Save ~17% with the annual plan
Pro
Perfect for small and medium teams
Unlimited users & infinite memory
Multi-LLM chats
Unlimited use of DeepSeek V4 Flash
US & EU data residency
Enterprise
Get in touch
For organizations with advanced needs
Unlimited users & SSO
Priority Support
Unlimited use of DeepSeek V4 Flash
US & EU data residency
30-days money back guarantee