The Vibe Father
Roundup
Cheap is easy to rank. Cheap and worth using is harder, because the cheapest token in the world is expensive if the model burns three attempts on a task a better model finishes in one. So this roundup does both: we rank the genuinely cheap models by price, then apply a benchmark-per-dollar lens using the public scores on our live board at /benchmarks — and we tell you honestly where each one stops being worth it.
One framing note before the list: these are per-token API prices. If you are paying a middleman markup on top of them, fix that first — we wrote up why in stop paying token markup.
The budget board
To make the prices concrete, we cost out a realistic heavy month for someone running agents daily: 50M input tokens and 10M output tokens. Where a benchmark score has not been published, we say so — we never fill in numbers.
| Model | In / Out per M | Heavy month | SWE-bench | Terminal-Bench | LiveCodeBench |
|---|---|---|---|---|---|
 MiniMax M3 | $0.30 / $1.20 | $27.00 | not published | not published | 82.2 |
 DeepSeek V4 Pro | $0.435 / $0.87 | $30.45 | 77.6 | not published | 87.5 |
 Kimi K2.7 Code | $0.95 / $4 | $87.50 | not published | not published | 82.1 |
 Claude Haiku 4.5 | $1 / $5 | $100.00 | 66.6 | 35.5 | 41.2 |
 Grok 4.5 | $2 / $6 | $160.00 | not published | not published | 87.4 |
 Gemini 3.5 Flash | $1.50 / $9 | $165.00 | 79.3 | 76.2 | 87.6 |
 GPT-5.3 Codex | $1.75 / $14 | $227.50 | 74.8 | 78.4 | 87.3 |
MiniMax M3 — the price floor
At $0.30 in and $1.20 out, MiniMax M3 is the cheapest serious coding model we track: a heavy month costs $27, roughly a lunch. It streams at 95 tokens per second, carries a 1M context window, and posts a respectable 82.2 on LiveCodeBench. The honest caveat is what we do not know — its SWE-bench Verified and Terminal-Bench scores are not yet published, so we cannot tell you how it handles real repository surgery or a live shell. On our index it is renormalized around the one score that exists. It is a fantastic scout and drafting model at a price where experimenting costs nothing; we go deeper in our MiniMax M3 review.
DeepSeek V4 Pro — the best evidence per dollar
DeepSeek V4 Pro costs $0.435/$0.87 — a $30.45 heavy month — and, unlike M3, it comes with receipts: 77.6 on SWE-bench Verified and 87.5 on LiveCodeBench. That SWE score is within striking distance of models charging four to seven times as much. Add open weights and a 1M context window and this is, by the published evidence, the strongest benchmark-per-dollar deal on the board. Its Terminal-Bench score is unpublished, and its 64 tokens per second is on the slower side, but for a builder seat on a budget it is our first call. Full writeup in our DeepSeek V4 Pro review.
Kimi K2.7 Code — the coding-plan sleeper
Kimi K2.7 Code runs $0.95/$4 ($87.50 heavy month) with a 262k context and an 82.1 on LiveCodeBench. Its SWE and Terminal-Bench numbers are not yet published, and its 46 tokens per second is the slowest on this list, so treat it as a value builder you audition rather than one you assume. Worth noting: its predecessor K2.6 did publish a full slate — 76.7 SWE, 66.7 Terminal-Bench, 86.8 LiveCodeBench — which gives you a reasonable prior for the family.
Claude Haiku 4.5 — cheap, with a hard ceiling
Here is where honesty matters. Haiku 4.5 is $1/$5, fast at 96 tokens per second, and its 66.6 SWE-bench is fine for a small model. But its 35.5 on Terminal-Bench is the lowest score anywhere on our board, and it tells you exactly what Haiku is not: an agent that can drive a shell through a multi-step task. Use Haiku as a scout — fast reads, summaries, triage, cheap first drafts — and never as an autonomous builder. Cheap tokens that fail agentic tasks are not cheap.
Grok 4.5, Gemini Flash, and Codex — the value flagships
Above the $100-a-month line sit three models that blur the budget category. Grok 4.5, released July 9, is $2/$6 ($160 heavy month) with an 87.4 LiveCodeBench and no SWE or Terminal-Bench scores published yet — promising, unproven. Gemini 3.5 Flash at $1.50/$9 ($165) is the most complete cheap model in existence: 79.3 SWE, 76.2 Terminal-Bench, 87.6 LiveCodeBench, 167 tokens per second, 1M context. And GPT-5.3 Codex at $227.50 a month is the most expensive model on this list but earns it with a 78.4 Terminal-Bench — the best agentic score in the budget tier.
The trap: retries are the real price
The per-token sticker is only half the equation; the other half is how many attempts a task takes. A model that costs a fifth as much per token but needs three runs to land a change is not saving you 80% — it is saving you 40% while tripling your wall-clock time and your review burden. This is why we insist on reading price against published capability rather than in isolation, and why the "worth using" filter in this roundup leans so hard on SWE-bench and Terminal-Bench numbers.
The practical version: match the model to the difficulty band. Routine, well-specified tasks — CRUD endpoints, test scaffolding, mechanical refactors — are where the cheap tier genuinely delivers, because first-attempt success rates stay high and every token saved is real. Ambiguous, multi-file, or architectural work is where cheap models start burning retries, and where escalating to a Sonnet 5 or an Opus 4.8 for one clean pass is usually the cheaper total outcome. The teams with the lowest real bills are not the ones on the cheapest model; they are the ones routing each task to the cheapest model that clears it first try — the role-based pattern we detail in the best model for each agent role.
How to choose
Our rule of thumb: if you need proven repository competence at the lowest price, take DeepSeek. If you want the most complete cheap all-rounder, take Flash. If you are filling a scout seat where speed and price matter more than depth, take M3 or Haiku — with eyes open about Haiku's terminal cliff. And if the task is genuinely hard, stop shopping in this aisle; a failed cheap run costs more than a successful expensive one. The full board, with every score cited to its source, lives at /benchmarks.