As of October 2026, GPT-6.1 Terra (80.5, Estimated) leads the models priced at or below $1.50 per million input tokens. The strongest open-weight option in that price band is DeepSeek V4.1 Flash 1017 (65.2, Supported), and below $0.30 input the leader is Claude Haiku 5.5 (66.4, Supported).
This guide ranks every major LLM under $1.50 per million input tokens by benchmark performance and price. The tables rebuild from our leaderboard and pricing page on every site update, so the leaders named above are current.
The budget tier, ranked
| Rank | Model | Type | License | Score | Evidence | Input / 1M |
|---|---|---|---|---|---|---|
| 1 | GPT-6.1 Terra | Reasoning | Proprietary | 80.5 | Estimated | $1 |
| 2 | Gemini 3.7 Flash | Reasoning | Proprietary | 66.5 | Supported | $0.75 |
| 3 | Claude Haiku 5.5 | Reasoning | Proprietary | 66.4 | Supported | $0.1 |
| 4 | Gemini 3.8 Flash | Reasoning | Proprietary | 66.2 | Supported | $0.75 |
| 5 | Muse Spark 1.2 | Reasoning | Proprietary | 65.4 | Supported | $1.25 |
| 6 | DeepSeek V4.1 Flash 1017 | Reasoning | Open Weight | 65.2 | Supported | $0.3 |
| 7 | GPT-5.6 Luna | Reasoning | Proprietary | 63.8 | Supported | $0.2 |
| 8 | DeepSeek V4 Pro 0813 | Reasoning | Open Weight | 62.5 | Estimated | $1.32 |
| 9 | Gemini 3.6 Flash | Reasoning | Proprietary | 62.4 | Supported | $0.75 |
| 10 | GPT-6 Luna | Reasoning | Proprietary | 62.1 | Supported | $0.1 |
| 11 | Gemini 3.5 Flash | Reasoning | Proprietary | 61.4 | Supported | $1.5 |
| 12 | GLM-5.2 | Reasoning | Open Weight | 60 | Supported | $1.4 |
| 13 | MiniMax M3.1 Flash | Reasoning | Proprietary | 59.7 | Estimated | $0.15 |
| 14 | GPT-5.1 | Reasoning | Proprietary | 56.8 | Supported | $1.25 |
| 15 | MiMo-V2.6-Pro | Reasoning | Open Weight | 56.4 | Estimated | $0.435 |
The score is not a value judgment by itself. It is the capability estimate; the adjacent evidence label shows whether the row has enough independent support to carry normally or needs the wider uncertainty attached to Estimated evidence.
The ultra-budget tier: $0.30 input or less
| Rank | Model | Type | License | Score | Evidence | Input / 1M |
|---|---|---|---|---|---|---|
| 1 | Claude Haiku 5.5 | Reasoning | Proprietary | 66.4 | Supported | $0.1 |
| 2 | DeepSeek V4.1 Flash 1017 | Reasoning | Open Weight | 65.2 | Supported | $0.3 |
| 3 | GPT-5.6 Luna | Reasoning | Proprietary | 63.8 | Supported | $0.2 |
| 4 | GPT-6 Luna | Reasoning | Proprietary | 62.1 | Supported | $0.1 |
| 5 | MiniMax M3.1 Flash | Reasoning | Proprietary | 59.7 | Estimated | $0.15 |
| 6 | MiMo-V2.6-Flash | Reasoning | Open Weight | 54.8 | Estimated | $0.14 |
| 7 | DeepSeek V3.2 | Non-Reasoning | Open Weight | 49.7 | Supported | $0.28 |
| 8 | MiniMax M2.5 | Non-Reasoning | Proprietary | 49.2 | Supported | $0.3 |
At these prices a 100M-token monthly classification pipeline costs tens of dollars. The step up to the full budget table above is worth it when the task needs multi-step reasoning, coding or agent work.
How to pick from the budget tables
- High-volume classification, tagging and extraction: start from the ultra-budget table. Cost dominates, and the top rows are more than capable for well-defined tasks.
- Budget coding assistant: take the top three rows of the main table and run them on your own repository tasks; the overall score is a shortlist, not a coding verdict. The coding leaderboard adds the category view.
- Budget agent: agent work punishes small error rates over many steps, so favour Supported rows near the top of the main table and test on the full workflow, not single calls.
- Long-context workloads: check advertised context on each model page, then test retrieval quality; the largest window is not automatically the best long-document system.
- Self-hosting: the open-weight rows let you trade API cost for serving cost. The self-host calculator estimates the hardware.
The March 2026 launch snapshot
The sections below are the original March 18, 2026 analysis of that week's budget launches, with the scores published then. They are kept for context; the live tables above are current, and where they disagree, the tables win.
GPT-5.4 mini and nano: OpenAI on a budget
GPT-5.4 mini is OpenAI's reasoning model at budget pricing: $0.75/M input, 3.3x cheaper than GPT-5.4. Its current public score appears in the live table above; the model has a 400K context window.
Where mini stands out:
- Agentic tasks: OSWorld-Verified 72.2 (vs full GPT-5.4's 85), Terminal-Bench 2.0 at 60, Tau2Bench 93.4. These are strong agentic scores for a budget model: OSWorld 72.2 beats Claude Haiku 4.5 (57) and Gemini 3 Flash (53) comfortably.
- Knowledge: GPQA 88 is solid, above Gemini 3 Flash (69) and Claude Haiku 4.5 (67), though below Gemini 3.1 Pro (97). HLE 41.5 is a standout on a benchmark where most budget models score in single digits.
- Multimodal: MMMU-Pro 76.6 is competitive, trailing only Gemini 3.1 Pro (95) and Claude Haiku 4.5 (82) in the budget tier.
Where mini falls short:
Coding is the big one: SWE-bench Pro 54.4 is a long way down from GPT-5.4's 85, so for a coding-focused workload the quality gap is real. MRCR v2 at 40.7 shows the long-context reasoning ceiling (full GPT-5.4 scores 97 here), and with no AIME or HMMT scores published yet, math capability is hard to evaluate at all.
The pitch: GPT-5.4 mini makes sense only after it wins a trial on the work you actually send. The broad data no longer supports calling it the default budget model.
One tier further down, GPT-5.4 nano costs $0.20/M input, 12.5x cheaper than full GPT-5.4. It lands at 66.72 overall with Supported evidence and materially outperforms the older GPT-5 nano budget row.
Key scores:
| Benchmark | GPT-5.4 nano | GPT-5 nano | GPT-5.4 mini |
|---|---|---|---|
| GPQA | 82.8 | 71.2 | 88 |
| HLE | 37.7 | — | 41.5 |
| SWE-bench Pro | 52.4 | 22 | 54.4 |
| Terminal-Bench 2.0 | 46.3 | 38 | 60 |
| OSWorld-Verified | 39 | 30 | 72.1 |
| MMMU-Pro | 66.1 | 58 | 76.6 |
GPT-5.4 nano beats GPT-5 nano on every available benchmark, especially coding (SWE-bench Pro 52.4 vs 22) and knowledge (GPQA 82.8 vs 71.2). The gap is large enough that GPT-5.4 nano effectively replaces GPT-5 nano for anything beyond the cheapest possible classification tasks.
The cost math: At $0.20/M input, nano processes 5 million input tokens per dollar. For a classification pipeline handling 100M tokens/month, GPT-5.4 nano costs $20/month. GPT-5.4 mini would cost $75/month for the same volume. That 3.75x multiplier matters at scale.
Where nano makes sense: High-volume tasks where cost dominates: classification, tagging, simple extraction, content filtering. For anything requiring strong reasoning or coding, the step up to mini ($0.75) is worth the extra cost.
MiniMax M2.7: the coding wildcard
MiniMax M2.7 is the surprise of this batch. At $0.30/M input (cheaper than both GPT-5.4 mini and nano for quality coding) it posts the highest SWE-bench Pro score in the budget tier: 56.22.
| Benchmark | MiniMax M2.7 | GPT-5.4 mini | GPT-5.4 nano | Claude Haiku 4.5 |
|---|---|---|---|---|
| SWE-bench Pro | 56.22 | 54.4 | 52.4 | 46 |
| Terminal-Bench 2.0 | 57 | 60 | 46.3 | 53 |
| SWE-Multilingual | 76.5 | — | — | — |
| MLE-Bench-Lite | 66.6 | — | — | — |
| Toolathlon | 46.3 | 42.9 | 35.5 | — |
MiniMax M2.7 beats GPT-5.4 mini on SWE-bench Pro by nearly 2 points while costing 2.5x less on input tokens. On SWE-Multilingual (76.5) and MLE-Bench-Lite (66.6), it shows strong coding breadth that the OpenAI budget models haven't been tested on yet.
The caveat: MiniMax M2.7 has since been superseded by MiniMax M3; check the live tables above for where the MiniMax line sits today.
200K context is another differentiator. At $0.30/M input, feeding large codebases into M2.7 is dramatically cheaper than any alternative with comparable SWE-bench scores.
Head-to-head in March 2026
Coding
| Model | SWE-bench Pro | LiveCodeBench | Price (in/out) |
|---|---|---|---|
| MiniMax M2.7 | 56.22 | — | $0.30/$1.20 |
| GPT-5.4 mini | 54.4 | — | $0.75/$4.50 |
| GPT-5.4 nano | 52.4 | — | $0.20/$1.25 |
| Claude Haiku 4.5 | 46 | 36 | $1.00/$5.00 |
| Gemini 3 Flash | 44 | 36 | $0.50/$3.00 |
| DeepSeek V3 | — | 37.6 | $0.27/$1.10 |
MiniMax M2.7 leads. For budget coding workloads (code review, bug fixing, refactors) it's the best value option in the tier. GPT-5.4 mini is close behind with the added benefit of being a reasoning model.
Agentic tasks
| Model | Terminal-Bench 2.0 | OSWorld-Verified | Price (in/out) |
|---|---|---|---|
| GPT-5.4 mini | 60 | 72.2 | $0.75/$4.50 |
| MiniMax M2.7 | 57 | — | $0.30/$1.20 |
| Gemini 3 Flash | 56 | 53 | $0.50/$3.00 |
| Claude Haiku 4.5 | 41 | 57 | $1.00/$5.00 |
| GPT-5.4 nano | 46.3 | 39 | $0.20/$1.25 |
| GPT-5 nano | 38 | 30 | $0.05/$0.40 |
GPT-5.4 mini dominates agentic benchmarks in this tier. OSWorld-Verified 72.2 is a standout: closer to full GPT-5.4 (85) than any other budget model gets to its flagship sibling. If you're building an agent on a budget, mini is the pick.
Knowledge
| Model | GPQA | HLE | Price (in/out) |
|---|---|---|---|
| GPT-5.4 mini | 88 | 41.5 | $0.75/$4.50 |
| GPT-5.4 nano | 82.8 | 37.7 | $0.20/$1.25 |
| GPT-5 nano | 71.2 | — | $0.05/$0.40 |
| Gemini 3 Flash | 69 | 6 | $0.50/$3.00 |
| Claude Haiku 4.5 | 67 | 11 | $1.00/$5.00 |
| DeepSeek V3 | 59.1 | — | $0.27/$1.10 |
GPT-5.4 mini and nano dominate knowledge benchmarks in the budget tier. HLE scores of 41.5 and 37.7 are particularly impressive; Claude Haiku 4.5 scores 11 and Gemini 3 Flash scores 6 on the same benchmark.
Multimodal
| Model | MMMU-Pro | Price (in/out) |
|---|---|---|
| Claude Haiku 4.5 | 82 | $1.00/$5.00 |
| Gemini 3 Flash | 80 | $0.50/$3.00 |
| GPT-5.4 mini | 76.6 | $0.75/$4.50 |
| GPT-5.4 nano | 66.1 | $0.20/$1.25 |
| GPT-5 nano | 58 | $0.05/$0.40 |
Claude Haiku 4.5 and Gemini 3 Flash lead the budget tier on multimodal. MiniMax M2.7 has no MMMU-Pro score, another gap in its benchmark coverage.
When to use what (March 2026)
High-volume classification and tagging: GPT-5 nano ($0.05/$0.40) or GPT-5.4 nano ($0.20/$1.25). If you're processing millions of tokens daily on simple tasks, nano-tier pricing is hard to argue with. GPT-5.4 nano is substantially better on quality if the 4x price increase fits your budget.
Budget coding assistant: MiniMax M2.7 ($0.30/$1.20). Highest SWE-bench Pro in the tier (56.22) at the second-lowest price. The 200K context window handles large codebases well. The caveat: limited benchmark coverage outside coding, so evaluate on your specific tasks.
Budget AI agent: GPT-5.4 mini ($0.75/$4.50). OSWorld-Verified 72.2 and Terminal-Bench 60 are the best agentic scores in the budget tier by a wide margin. The reasoning capability helps with multi-step agent workflows.
Long-context workloads: Gemini 3 Flash ($0.50/$3.00) with 1M context, or GPT-5.4 mini ($0.75/$4.50) with 400K. Context length still needs a retrieval-quality test; the largest advertised window is not automatically the best long-document system.
Best budget all-rounder: see the live table at the top of this post.
Cheapest reasoning model: GPT-5.4 nano ($0.20/$1.25). The only reasoning model under $0.50/M input with broad benchmark coverage. GPQA 82.8 and HLE 37.7 show real reasoning capability at an ultra-budget price.
The ChatGPT alternatives table offers another API-model comparison. For a fixed monthly consumer subscription, check the app allowances separately.
The data gap problem
MiniMax M2.7's overall score still does not tell the full story. BenchAlign v5.8 estimates the model from the evidence it has, preserves uncertainty where it does not, and separates Supported from Estimated rows instead of assigning a zero for an unreported benchmark.
On the direct benchmarks that exist, M2.7 remains competitive with GPT-5.4 mini. SWE-bench Pro 56.22 and Terminal-Bench 57 are useful signals. They should inform a coding trial, not stand in for a general-purpose verdict.
This is a recurring problem in AI benchmarking. As we covered in the benchmark reliability explainer, benchmark coverage and provenance matter as much as the scores themselves. A model with 10 strong scores and 20 unknowns is a riskier choice than a model with 25 moderate scores.
The practical takeaway: published coding rows justify a coding trial; they do not stand in for a general-purpose verdict. Measure candidates on your own acceptance tests.
What the March launches meant for the market
Three takeaways from this week's releases:
1. The reasoning gap is closing at the bottom. GPT-5.4 mini and nano bring reasoning-class capability to the budget tier. A year ago, reasoning models started at $2.50/M input. Now you can get HLE 37.7 for $0.20/M input.
2. Chinese models keep punching above on coding. MiniMax M2.7 posting the highest SWE-bench Pro score in the budget tier (above both GPT-5.4 mini and nano) continues the trend of Chinese labs producing strong coding models at aggressive price points.
3. Budget doesn't mean weak anymore. GPT-5.4 mini's OSWorld-Verified 72.2 would have been a frontier-class score 12 months ago. The models that cost $0.30–$0.75/M input today are materially better than the $15/M models of early 2025.
Check the live leaderboard for the latest scores as more benchmarks roll in for these models. Prices and capabilities shift fast; what's budget today is obsolete tomorrow.
Frequently asked questions
01What is the best budget LLM in 2026?
The live table at the top of this post ranks every model priced at or below $1.50 per million input tokens by BenchAlign score, and its first row is the best budget model on our board today. In October 2026 that was GPT-6.1 Terra at $1 input, on Estimated evidence; the strongest Supported rows were Google's Gemini Flash models and Claude Haiku 5.5.
02What is the best open-weight budget LLM?
This post names the open-weight leader under $1.50 input in its opening line, from the same live data. In October 2026 it was DeepSeek V4.1 Flash at $0.30 input with Supported evidence. Open weights also let you self-host, which changes the cost math at high volume.
03What is the cheapest good LLM API in 2026?
Use the second live table in this post, which ranks models at or below $0.30 per million input tokens. In October 2026 Claude Haiku 5.5 led it at $0.10 input. Ultra-budget models suit classification, extraction and summarization; test them before giving them hard coding or agent work.
04Is a budget LLM good enough for production?
Often, for well-defined tasks. The top budget rows now score where flagship models did a year earlier. Check the Evidence column: a Supported row has enough independent results behind it, while an Estimated row carries wider uncertainty. Then run your own acceptance tests, because a cheap model that fails one in ten tasks can cost more than a pricier one.