One Bug Fix, Three Cheap Models: $0.0006 to $0.0045

Three cheap-tier models fixed the same two-bug Python function and passed all 215 hidden tests — at costs ranging from $0.0006 to $0.0045 per fix.

The run comes from AI gateway AIHubMix, which published it on 19 October 2026: "One two-bug Python fix, 215 hidden tests, three runs per model through one AIHubMix key, on 19 October 2026. All 12 runs passed every test." The models were Claude Haiku 5.5, DeepSeek V4.1 Flash 1017 and MiniMax M3.1 Flash. AIHubMix notes these are "AIHubMix's own measurements, not yet reproduced by anyone else" — it sells access to all three, so treat the figures as a vendor benchmark, not an independent one.

Same correctness, sevenfold cost spread

At default settings the post reports Haiku 5.5 finishing "in about 7 s for $0.0007 per fix", DeepSeek V4.1 Flash 1017 "in about 15 s for $0.0014", and MiniMax M3.1 Flash "in 56 s for $0.0045". Sending reasoning_effort: "low" dropped M3.1 Flash to 9.5 s and $0.0006 — "the cheapest configuration tested".

Per 1,000 such fixes, AIHubMix puts the bill at "about $0.70" for Haiku, "about $1.35" for DeepSeek, and "about $4.45 at default or $0.64 at low" for M3.1 Flash. Measured throughput was roughly 183 tok/s for Haiku, 148 for DeepSeek and 112 for Flash.

The price card is the smaller half

Model	Input / output per 1M	Max output	Thinking can be off
Claude Haiku 5.5	$0.10 / $0.50	128K	Yes
DeepSeek V4.1 Flash 1017	$0.155 / $0.62	384K	Yes
MiniMax M3.1 Flash	$0.165 / $0.66	512K	No

Input prices span $0.10 to $0.165 — a 65% gap — yet the measured cost spread was sevenfold. The gap is thinking: at its default, M3.1 Flash "spent 6,364 thinking tokens against Haiku's 755", and thinking is billed as output on all three. The default effort level shaped the bill more than the per-token price did.

Three catches before you route traffic

• Omitting the effort field on M3.1 Flash means maximum thinking. MiniMax documents five levels — low, medium, high, xhigh, max — and "the model uses max when the field is omitted". Code ported from M2.7 never sends the field, so every request lands on max. The same day, MiniMax closed M2.5/M2.7 to new users and routed plan requests to M3.1 Flash.
• Haiku has a long-prompt cliff. "Above 100K prompt tokens, Haiku 5.5 bills $0.50 / $2.50" — five times its base input rate. Large-context calls are where the cheapest model stops being cheapest.
• One listed price is promotional. A 30% AIHubMix discount on DeepSeek V4.1 Flash 1017 "runs from 19 to 25 October" and is excluded from the figures above. Budget on list price.

What the test does not settle

Twelve of twelve runs passed, so the task separates the models on cost and latency only — not on quality. AIHubMix says as much: "Three runs per setting is a small sample, and thinking length varies a lot between runs." For M3.1 Flash there is no external score to fall back on either: "MiniMax hasn't published one and Artificial Analysis doesn't list it." DeepSeek is the only one of the three shipping open weights (MIT); Haiku and M3.1 Flash do not.

If you run a cheap tier in production, the actionable step is narrower than a model switch: log reasoning tokens per call, and set the effort field explicitly on any model whose default is maximum. Then re-run your own two or three hardest tasks — this one was too easy to rank anybody.