Kimi’s official products page lists its online workspace alongside Kimi Work and Kimi Code, and dedicated Slides, Docs, Sheets and Deep Research features. Those are meaningful routes if the desired output is a presentation, an editable document, a spreadsheet or a sourced research report. Our Kimi Slides, Docs, Sheets and Deep Research guides describe narrower hands-on workflows and their observed limits.

DeepSeek’s official site directs visitors to its chat and developer platform. If your job is to ask questions, inspect cited web results or work with a PDF in chat, follow the official route and our DeepSeek Search and PDF guides for tested steps. This comparison does not claim DeepSeek lacks features that Kimi lists, or that one app’s interface is better because an API response was better.

For either service, use a short non-sensitive sample first, check the output against the source, and then review the account’s data and plan settings. Access and quotas may change.

The API study used ten fictional text prompts across the configurations below. Read the provider, mode and output cap together: they affect which results can be compared. The app row shows where the separate web-app trials fit.

DeepSeek’s change log, rechecked September 26, 2026, says Pro API service continues beyond September 14. Its current model table maps deepseek-v4-pro to V4-Pro-0813. This clarifies the earlier September 10 announcement of a planned switch to V4.1 Flash. Our September 24 response returned the Pro alias; it did not expose a verifiable checkpoint hash or establish that its serving stack matches Together’s. The official chat apps may use other model versions and tools.

We fixed two prompts per category before testing. The most interpretable text-answer comparison below uses Kimi K3 on Together with no-thinking requested and DeepSeek’s first-party alias with thinking disabled, both capped at 1,800 output tokens. Those are different providers; the table describes observed answers, not controlled model performance. We also publish the same-host Together lanes, including every blank final answer, in the full records.

On a narrow screen, swipe the table to compare all columns.

What would have changed a real user’s answer?

In the gallery notices, the prompt explicitly said to mark an absent contact email as “Not stated.” Kimi did not include that field, even though its other dated facts were right. This is a failure of completeness rather than a fabricated email. In the CSV repair, Kimi’s function accepted the malformed input "a"x,b; the frozen task requirement was to reject characters after a closing quote. First-party DeepSeek’s function rejected it. These are inspectable differences, unlike a vague impression that one answer “sounds smarter.”

The two invoice prompts are deliberately fictional and have independently calculated outcomes. Both APIs reached the same correct totals, but the catering fee’s third decimal did not distinguish half-up from half-even rounding; that task should not be cited as evidence of rounding-mode mastery. The notices and translations likewise are small first-answer examples, not general accuracy rates.

Our first frozen lane asked the Together-hosted DeepSeek V4 Pro 0813 and Together-hosted Kimi K3 to answer all ten prompts at high reasoning with a 3,000-output-token cap. Kimi returned final text on all ten. DeepSeek returned final text on seven, while the shuttle and both code tasks exhausted the cap without any final answer. We kept those three empty finals in the published records.

After seeing that, we froze a second plan and reran all ten prompts for both Together endpoints, this time requesting no-thinking and allowing 1,800 output tokens. Kimi returned final text on all ten again. Together’s DeepSeek endpoint still reported substantial reasoning-token use and no final text on the same three tasks. Its English-to-Chinese answer also stopped at the cap, but contained partial final text. We then froze a third plan and ran the ten prompts on DeepSeek’s first-party API with thinking disabled. That service returned final text on all ten. These were post-hoc diagnostic lanes, not pre-planned repeated trials; the second lane changed both requested mode and cap, so it does not isolate one cause.

The actionable lesson is to test the actual host and configuration you intend to deploy. A blank final answer at a token cap can reflect how reasoning, mode controls and output budgets interact. It does not establish that a model cannot solve the task. Conversely, one complete answer from a different host does not prove every future call will complete.

On September 24, 2026, Together listed moonshotai/Kimi-K3 at $3 per million uncached input tokens and $15 per million output tokens. It listed deepseek-ai/DeepSeek-V4-Pro-0813 at $1.32 and $3.96 respectively. At those list rates, a hypothetical one million uncached input tokens plus one million output tokens would be $18.00 for Kimi and $5.28 for DeepSeek. This is simple rate arithmetic, not our actual test bill or a comparison of consumer subscriptions. Cached inputs, provider discounts, internal reasoning output and changed rates can alter a real invoice. Check the live Kimi and DeepSeek cards before buying capacity.

Together lists Kimi K3 with text and image input; it lists this DeepSeek V4 Pro endpoint as text input. That is a provider-endpoint distinction, not a claim about everything either company’s apps or model family can do. DeepSeek’s first-party API rates and peak/off-peak rules are separate from Together’s pricing. See our DeepSeek API, Kimi API, DeepSeek pricing and Kimi pricing guides before choosing a service.

If you are evaluating your own workload, decide in advance how you will judge completeness, factual support, format, edge cases and spend. Test the same prompt and output requirement on the exact provider configuration you would ship. Keep blank or incomplete results, and allow enough output tokens for a reasoning model.

We froze ten prompts and success criteria before the first request: two source-update tasks, two invoices, two Python repairs, two short public notices and two translations. Every prompt used fictional material. Across the API configurations, we preserved 50 first responses, with one request per prompt in each configuration, no tools and no quality retries. The two supplemental test plans were chosen after the initial results; see how the host, mode and output cap changed.

We independently ran the returned Python functions against local checks written during evaluation, after the first requests; the functional requirements in the prompts were frozen beforehand, but the executable checker itself was not preregistered. Reviewers separately inspected the source facts, arithmetic, word limits and bilingual meaning. These AI-assisted reviews are our internal checks, not a human panel or external expert endorsement. The publicly downloadable package contains exact prompts, full final answers, empty finals, reported model names, settings, finish reasons, token usage and code-check results. Private request identifiers, credentials and reasoning traces are excluded.

Download all 50 first responses and protocol (.zip)

Limits that matter: Ten distinct fictional text prompts with one first response per configuration are too few to calculate a general accuracy or reliability rate. The lanes differ in host, mode and output cap; the first-party alias did not expose the exact backend checkpoint. The separate app trials above tested one source check with browsing, one document revision and one chart image. They did not test PDFs, long-context retrieval, a real repository, safety, privacy, uptime or speed; neither the app nor API samples establish general performance. No result here predicts performance on your private files or production traffic. We did not regenerate a failure to make the table look better.

For the wider context, visit the DeepSeek overview, Kimi overview or other comparisons. We will update this page when a material provider mode, model route, feature or price changes, or when a reproducible correction affects a finding. Send a correction with the affected prompt and evidence.