MiniMax M3.1 Flash vs DeepSeek and Claude Haiku: A Same-Bug Fix Benchmark

A Controlled Showdown of Cheap Models

When three vendors release lightweight models within weeks of each other, the selection question becomes concrete: for the same coding task, who is faster, more reliable, and cheaper? According to testing by AIHubMix, a Python function repair task containing two bugs, evaluated against 215 hidden tests, passed in all 12 runs across MiniMax M3.1 Flash, DeepSeek^Ad V4.1 Flash 1017, and Claude Haiku 5.5. The real difference isn't correctness — it's the cost and latency gap created by default thinking budgets.

The Basics of the Three Models

The three models come from the budget product lines of MiniMax, DeepSeek, and Anthropic respectively. All have a 1M context window, but their pricing and capability orientations differ.

Metric	MiniMax M3.1 Flash	DeepSeek V4.1 Flash 1017	Claude Haiku 5.5
Input/Output (per million tokens)	$0.165 / $0.66	$0.155 / $0.62	$0.10 / $0.50
Cache read (per million tokens)	$0.0165	$0.0031	$0.01
Max output	512K	384K	128K
Input types	Text, image, video	Text, image	Text, image
Default thinking level	max	on	medium
Open weights	Not yet	Yes (MIT)	No

DeepSeek V4.1 Flash was released on September 10, with the 1017 update arriving on October 17; Haiku 5.5 launched on October 7; M3.1 Flash ran anonymously as Space Bunny Alpha before being named on September 27, and opened for pay-per-token calls on AIHubMix on October 19. During its free trial week it topped the OpenRouter daily chart, ahead of DeepSeek V4.1 Flash.

Context window refers to the maximum token sequence length (including input and output) a model can hold in a single call, and is a key metric for handling long documents, multi-turn conversations, or large codebases. A 1M context means, in theory, reading roughly 750,000 English words or hundreds of thousands of lines of code in one go. But note that a large window does not equal economy and stability at that window: most vendors set higher unit prices for prompts exceeding conventional lengths (the "price cliff" for Haiku discussed later), and effective retrieval within long contexts is usually weaker than in short ones. Also, "open weights" in the table means the model parameters can be downloaded and deployed locally, which differs fundamentally from API-only closed models in data compliance and cost structure.

Test Method: Two-Bug Fix with 215 Hidden Cases

The function under test calculates the active days of a subscription within a half-open interval, and has two defects: when a pause interval extends beyond the end of the period, an extra day is subtracted; overlapping pauses are deducted twice. The original function passes only 138 of the 215 tests.

The evaluator extracts the last code block from the model's response and runs it in an isolated subprocess: 15 hand-written edge cases (adjacent, nested, duplicate pauses, covering the full period, empty period, pause starting on the end date) plus 200 fixed-seed random cases, the latter compared against a brute-force reference implementation that counts day by day. The prompt describes only the symptoms and asks for a complete corrected function.

Results: Correctness Tied, Cost Determined by Thinking Level

Three runs, Chat Completions API, max_tokens at 32768, measured by AIHubMix on October 19:

Model & Setting	Avg. thinking tokens	Avg. output tokens	Avg. duration	Avg. cost per run	Tests passed
Claude Haiku 5.5 (default)	755	1,323	7.2s	$0.0007	215/215 ×3
DeepSeek V4.1 Flash 1017 (default)	1,823	2,148	14.6s	$0.0014	215/215 ×3
MiniMax M3.1 Flash (default=max)	6,364	6,673	56.3s	$0.0045	215/215 ×3
MiniMax M3.1 Flash (low)	591	889	9.5s	$0.0006	215/215 ×3

Key readings:

• No difference in correctness. All 12 fixes were perfect; this task cannot distinguish model quality.
• Default thinking volume differs by about 8×. Flash consumes 6,364 thinking tokens by default versus only 755 for Haiku, which directly explains why the former's default cost is about six times the latter's.
• Flash is cheapest at the low setting. $0.64 per thousand fixes versus $0.70 for Haiku, but Haiku is still faster.
• Throughput. Output tokens divided by actual duration: Haiku about 183 tok/s, DeepSeek about 148 tok/s, Flash about 112 tok/s.

Per thousand fixes: Haiku about $0.70, DeepSeek about $1.35, Flash about $4.45 at default and about $0.64 at low.

A second check classified five support tickets into four labels, with each model run at both default and minimum thinking settings. All 30 answers were correct, with cost per call under $0.0001; thinking tokens per call were 0–6 for Flash, 0 for Haiku, and about 25 for DeepSeek.

Reasoning tokens are the internal reasoning content generated by reasoning models before producing a final answer. They are usually not counted as visible output, but they are billed as output tokens and significantly lengthen latency. reasoning_effort is the request parameter controlling this budget, with common levels none/low/medium/high/max. Vendors do not agree on level naming or defaults: some default to off, others default to maximum. This is precisely the root of the cost differences in this test — the bill depends not only on the per-token price but also on "how long the model is willing to think by default." When tuning, treat effort as an explicit parameter as important as temperature, and avoid relying on vendor defaults.

Failure Modes and How to Avoid Them

Flash's default setting looks like a hang. When the effort field is omitted, M3.1 Flash thinks at the max level, and a small fix takes an average of 56 seconds, which will trigger timeouts in clients tuned for fast models. Explicitly pass reasoning_effort on every request, and start from low.

DeepSeek's none doesn't zero out thinking. Under the three-class classification prompt, DeepSeek consumes about 25 reasoning tokens under both none and default settings; the cost difference is negligible, but don't assume none returns a non-thinking response on this route.

Haiku's long-prompt price cliff. Beyond 100K prompt tokens, Haiku 5.5 bills at $0.50/$2.50. Estimating a 300K-token uncached prompt, counting input only and assuming the higher price applies to the entire prompt: Haiku costs 300,000 × $0.50 ÷ 1M = $0.15, versus about $0.05 for M3.1 Flash, comparable to DeepSeek — a roughly 3× difference per call. Bypass Haiku for large-context calls or use caching aggressively.

Promotional prices drift. The discounted price for DeepSeek V4.1 Flash 1017 on AIHubMix is shown through October 25; budget at list price.

Prompt caching lets vendors reuse previously processed identical prefix tokens, with cache-read prices typically only 1% to 10% of normal input prices. The three cache-read prices in the table are $0.0165, $0.0031, and $0.01 respectively — a clear gap. For calls that repeatedly carry the same long system prompt or codebase, placing stable content at the front of the prompt and dynamic content at the end can greatly improve hit rates, thereby easing the price pressure of long prompts. But caches usually have short lifetimes (minutes to hours) and may not hit across sessions, so don't be overly optimistic when budgeting.

The Boundaries of This Test

All runs passed; distinguishing quality would require harder multi-file tasks. Each setting had only three runs, a small sample, and thinking length varied noticeably between runs. M3.1 Flash has no public benchmark yet — MiniMax hasn't released one and Artificial Analysis hasn't listed it; DeepSeek's benchmark scores are self-reported from its V4.1 Flash technical report; Haiku's Artificial Analysis score of 43 is cited by OrcaRouter. Video input via AIHubMix was not verified in this test, and the exact billing rules for Haiku's long prompts are estimates.

Selection Recommendations

• Human waiting, short prompts: Claude Haiku 5.5.
• Benchmark focus, self-hosting, or a true non-thinking mode: DeepSeek V4.1 Flash 1017.
• Large context, long output, or video input, with effort explicitly set: MiniMax M3.1 Flash, starting from low or medium.

The most memorable conclusion from this test: on short coding tasks like this, the default thinking level affects the bill more than the per-token price itself. When selecting a model, ask about the default level first, then compare price lists.