Kimi K3 vs MiniMax M3

MiniMax M3 has been retired from our catalogue. Its figures below are the last published rates; calls to it are no longer served, while the other model in this comparison is.

vs

Which one, when

Both take text, image and video in and return text, with chat, code, reasoning and tool use on either side, so the real split is price and output headroom: kimi-k3 costs 10x more on input, 12.5x more on output and 5x more on cached reads than minimax-m3. Pick kimi-k3 when you need its 1048576-token context and matching 1048576-token max output for single-pass generation of very long artifacts, since minimax-m3 caps output at 524288. Pick minimax-m3 for cost-sensitive volume, its long-context flag, and the option to turn thinking off, which kimi-k3 does not allow.

Benchmarks

MiniMax M3: the vendor has not published benchmark scores.

Above averageNo peer higherKimi K352 / 6712 / 67
Kimi K3 MiniMax M3 other models measured peer average ★ no peer scored higher
DeepSWE 1.1
67.5%
N/A
OSWorld-Verified
84.8%
N/A
Cybergym
80%
N/A
Finance Agent v2
54.4%
N/A
Harvey Lab-AA
no peer scored higher 94.6%
N/A
GPQA Diamond
93.5%
N/A
BrowseComp
91.2%
N/A
Video-MME (w. sub)
no peer scored higher 90%
N/A

Vendor-published: Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Alibaba (Qwen) Anthropic Anthropic Anthropic Anthropic ByteDance DeepSeek DeepSeek Google Google Google Google MiniMax Moonshot Moonshot OpenAI OpenAI OpenAI OpenAI OpenAI Tencent Z.ai Z.ai Z.ai Z.ai

Pricing

Kimi K3 MiniMax M3 Δ
Input / 1M tokens $3 $0.3 10×
Output / 1M tokens $15 $1.2 13×
Cache read / 1M tokens $0.3 $0.06 5×
Cache write - no separate charge -

Rates from the live catalogue at build time; each model page carries the current rate card.

Where they sit · input price per 1M tokens across all 91 chat models on this billing unit (log scale)

Kimi K3 · $3 MiniMax M3 · $0.3
$0.05 · Qwen3 VL Flash $30 · GPT-5.4 Pro

Capabilities

Kimi K3 MiniMax M3
Tool calling yes yes
Thinking control always on configurable
Structured output yes -
Prompt caching implicit (automatic) implicit (automatic)
Cache lifetime not published not published
Minimum cached prefix not published 512 tokens

Specs

Kimi K3 MiniMax M3
Input modalities text image video text image video
Output modalities text text
Released 2026-07-16 2026-06-01
Context window 1,048,576 1,000,000
Max output 1,048,576 524,288
Thinking parameter reasoning_effort (top-level; the thinking object is not accepted)
  • thinking.type
  • reasoning_split
Accepted values
reasoning_effort
  • low
  • high
  • max
thinking.type
  • adaptive
  • disabled
reasoning_split
  • boolean
Default max adaptive: thinking on, with the model deciding when extra reasoning helps

Specs are transcribed from each vendor’s documentation; a row a vendor does not publish is left out rather than inferred. Full sources: Kimi K3 · MiniMax M3

One prompt, both models, measured through the gateway

PROMPT Explain to a CFO, in exactly three sentences, why our LLM bill doubled after we enabled reasoning. CHECK exactly 3 sentences

Kimi K3 passed · 3 sentences

Reasoning models don't just answer questions—they "think" first, generating long internal chains of step-by-step logic before producing a response. Those hidden thinking steps are billed as output tokens (the most expensive kind, typically 3–5x the price of input tokens), and a single query can generate thousands of them even when the visible answer is only a paragraph long. So you're paying for dramatically more compute per request: the bill doubled because the model does far more work behind the scenes, not because usage increased.

out 755 tok (+637 thinking) latency 20.8 s

MiniMax M3 passed · 3 sentences

Reasoning models don't just answer — they first generate extensive internal "thinking" tokens to work through the problem step-by-step, and you're billed for all of those tokens even though users never see them. A query that previously cost ~200 tokens may now cost 2,000–10,000 once reasoning is enabled, multiplying per-query spend by roughly 5–50x while the visible output stayed the same size — that's the doubling. The lever is tuning: cap reasoning effort on simple queries, cache the reasoning traces, and reserve the full reasoning model for only your hardest 10–20% of requests, which typically restores 60–80% of the savings.

out 764 tok (+629 thinking) latency 3.5 s

Instruction following (exactly three sentences, so countable), audience shaping (CFO register), and the hidden-thinking billing gap the token meters below expose.

PROMPT This function is supposed to return the longest run of consecutive equal items, but callers report it is off by one on single-element inputs and misses runs that end at the last index. Fix it and explain the bug in one paragraph. CHECK fix passes tests

Kimi K3 passed · 8/8 cases

```python def longest_run(xs): if not xs: return 0 best = 1 cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 best = max(best, cur) else: cur = 1 return best ``` The bug is a fencepost error: `cur` counts *matching adjacent pairs* rather than *items in the run*, because it starts at 0 and only increments when `xs[i] == xs[i - 1]`. A run of length L contains L−1 equal pairs, so every run was undercounted by exactly one. For a single-element list (a run of length 1) the loop never runs and the function returns 0 instead of 1; likewise any run ending at the last index — e.g. the two `2`s in `[1, 2, 2]` — reported 1 instead of 2, making it look like no run existed at all. […]

out 1837 tok (+1547 thinking) latency 47.2 s

MiniMax M3 passed · 8/8 cases

The bug is that the loop starts at index 1 and treats `cur` as the number of *equal transitions* rather than the run length, so a run of *k* consecutive equal items is reported as `k − 1`, and a single-element list (where the loop body never executes) returns 0 instead of 1. A correct version tracks the actual run length by seeding `cur = 1` and resetting it to 1 (not 0) on a mismatch: ```python def longest_run(xs): if not xs: return 0 best = cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 else: cur = 1 if cur > best: best = cur return best ``` Now `[5]` returns `1`, `[1,1,1]` returns `3`, and a run like `[1,2,2,2]` that ends at the last index is correctly captured as `3` because `cur` is updated before `best` is compared on every iteration, including the final one.

out 845 tok (+596 thinking) latency 3.5 s

Whether the fix is actually correct (runnable), explanation density, and token efficiency on a bounded task.

PROMPT Extract a JSON object with fields {company, quarter, revenue_usd_m, growth_pct, guidance} from this text. Use null for anything not stated; add no other fields. CHECK valid JSON, schema exact

Kimi K3 passed · 5/5 fields, guidance null

{ "company": "Meridian Robotics", "quarter": "September quarter", "revenue_usd_m": 412, "growth_pct": 11, "guidance": null }

out 924 tok (+863 thinking) latency 25.9 s

MiniMax M3 passed · 5/5 fields, guidance null

```json { "company": null, "quarter": "Q3", "revenue_usd_m": 412, "growth_pct": null, "guidance": null } ```

out 1002 tok (+955 thinking) latency 3.6 s

Schema obedience (no invented fields), hallucination pressure (guidance is explicitly withheld), and structured-output path differences.

PROMPT Write a 120-word product announcement for an API gateway feature that caches prompts across providers. Forbidden words: "seamless", "unlock", "game-changer", "revolutionize", "empower". Exactly one sentence must be a question. CHECK 120 words, 0 banned words

Kimi K3 passed · 120 words, 0 banned, 1 question

Meet Prompt Cache, a new API gateway capability that stores prompt responses and serves them across OpenAI, Anthropic, Google, and Azure endpoints. It matches requests by model, prompt hash, tools, temperature, and tenant policy, so repeated work returns fast while sensitive contexts stay isolated. Teams set TTLs, stale rules, encryption scopes, and bypass flags per route. Analytics show hit rate, latency saved, token spend avoided, and drift risk by provider. What changes for developers? Keep one integration, add cache headers, and watch fallback logic respect consent, residency, and audit needs. During rollout, canary keys compare fresh answers with cached copies before promotion. Prompt Cache cuts vendor calls, steadies p95 latency, and gives platform owners controls for cost, quality, and compliance.

out 1527 tok (+1354 thinking) latency 37.9 s

MiniMax M3 passed · 121 words, 0 banned, 1 question

Introducing PromptCache, our intelligent caching layer for the API gateway. Tired of paying twice for identical requests? The feature stores prompt-response pairs across every connected provider, returning cached results instead of hitting costly endpoints. Deployment requires only a single configuration flag, and semantic similarity detection happens automatically. Latency drops up to 70% on repeated traffic while costs follow the same downward curve. Developers access detailed cache hit metrics, eviction logs, and provider breakdowns through the dashboard. Whether routing to OpenAI, Anthropic, or local models, every cached token saves time and money. The cache respects TTL policies, supports manual purging, and encrypts stored content. Available today for Pro and Enterprise tiers—enable PromptCache on existing endpoints and watch inference bills shrink immediately.

out 2808 tok (+2656 thinking) latency 6.0 s

Constraint obedience (word budget, banned-word list, the single question), style fingerprint, and length control.

Switch between them with one line

Both ids are in every tab below; the highlighted pair of lines is the only edit. Same endpoint, same key, same request shape.

from openai import OpenAI

client = OpenAI(
    base_url="https://synthorai.io/v1",
    api_key="sk-syn-...",
)

resp = client.chat.completions.create(
    model="kimi-k3",
    # model="minimax-m3",  # uncomment this line, comment the one above
    messages=[{"role": "user", "content": "Summarize this diff"}],
    reasoning_effort="medium",
)
print(resp.choices[0].message.content)

Get your API key →

FAQ

Which is cheaper, Kimi K3 or MiniMax M3?

MiniMax M3 is cheaper on the "Input / 1M tokens" row ($0.3 vs $3, 10× apart). Other rows may point the other way; the table above carries the full rate card, and real cost depends on your mix.

Can I A/B test Kimi K3 against MiniMax M3 without two integrations?

Yes. Both are served through the same OpenAI-compatible endpoint with one API key. Switching is a one-line change to the model id, so you can route a fraction of traffic to each and compare bills directly.

Do Kimi K3 and MiniMax M3 support prompt caching?

Yes. Both bill cache reads below their input rate, so warm-prefix workloads cost less than the list rates suggest. The exact cache-read rows are in the pricing table above.

Related comparisons

From our measured studies