- MiniMax released M3.1 Flash on October 19, 2026 at $0.15 input and $0.60 output per million tokens, the same as DeepSeek V4.1 Flash 1017's off-peak price. MiniMax says weights 'will be released later' and hasn't published a size or license.
- Claude Haiku 5.5 and GPT-6 Luna list the lowest prices here, $0.10/$0.50, with surcharges above 100K-token (Haiku) and 272K-token (Luna) prompts. Gemini 3.8 Flash costs $0.75/$3.75 through December 31, 2026, then $1.50/$7.50.
- On Artificial Analysis's Intelligence Index v4.3.2 (checked October 20), DeepSeek 1017 (Max) scores 45, Haiku 5.5 (Max) and M3.1 Flash 43, Gemini 3.8 Flash (High) 41 and GPT-6 Luna (Max) 38. Vals AI's index (version 2.2, October 19) puts DeepSeek 1017 first of these at 61.26%.
Contents
Five budget models now compete for the same jobs: classification, extraction, routing, subagents and high-volume chat. MiniMax released MiniMax M3.1 Flash on October 19, 2026, two days after DeepSeek's V4.1 Flash 1017 update. They join Anthropic's Claude Haiku 5.5 (October 7), OpenAI's GPT-6 Luna (September 22) and Google's Gemini 3.8 Flash (September 2) in the small-model tier. This comparison of cheap, fast AI models covers official prices, context and output limits, open weights, and what independent evaluators have published, with each evaluator kept separate.
This comparison is based on each company's pricing pages, model documentation and launch posts, plus results published by Artificial Analysis, Vals AI and LMArena, all checked October 20, 2026. For the DeepSeek release itself, see our DeepSeek V4.1 Flash 1017 story. For Anthropic's model, see our Claude Haiku 5.5 story.
The short version
- Cheapest list price: Claude Haiku 5.5 and GPT-6 Luna, at $0.10 input / $0.50 output per million tokens. Both cost more on long prompts: Haiku 5.5 above 100K tokens, Luna above 272K.
- Cheapest open-weight model: DeepSeek V4.1 Flash 1017, at $0.15/$0.60 off-peak (double at peak), with MIT-licensed weights on Hugging Face.
- MiniMax M3.1 Flash: $0.15/$0.60 flat. Weights are promised "later," and MiniMax hasn't published a parameter count or license.
- Gemini 3.8 Flash costs the most, at $0.75/$3.75 through December 31, 2026, then $1.50/$7.50. Google's Flash models are the only ones here with audio and PDF input and a free API tier.
- Independent scores are close. Artificial Analysis puts the five within 7 points of each other, and Vals AI within about 9 points. Neither says one model is best at everything.
How much do the cheap AI models cost?
US dollars per million tokens, standard tier.
| Model (API ID) | Released | Input | Cached input | Output | Long-prompt rate |
|---|---|---|---|---|---|
Claude Haiku 5.5 (claude-haiku-5-5) | Oct 7, 2026 | $0.10 | $0.01 read; $0.125 write (5 min) | $0.50 | Prompts over 100K tokens: $0.50 / $2.50 |
GPT-6 Luna (gpt-6-luna) | Sep 22, 2026 | $0.10 | $0.01 read; $0.125 write | $0.50 | Prompts over 272K tokens: $0.20 / $0.75 |
DeepSeek V4.1 Flash 1017 (deepseek-flash) | Oct 17, 2026 | $0.15 off-peak; $0.30 peak | $0.003 off-peak; $0.006 peak | $0.60 off-peak; $1.20 peak | None listed |
MiniMax M3.1 Flash (MiniMax-M3.1-Flash) | Oct 19, 2026 | $0.15 | $0.015 | $0.60 | None listed ("flat pricing") |
Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite) | Updated July 2026 | $0.30 | $0.03 | $2.50 | None listed |
Gemini 3.8 Flash (gemini-3.8-flash), through Dec 31, 2026 | Sep 2, 2026 | $0.75 | $0.075 | $3.75 | None listed |
| Gemini 3.8 Flash, from Jan 1, 2027 | $1.50 | $0.15 | $7.50 | None listed |
Table 1Haiku 5.5 and Luna share the lowest list price, but only on prompts under their long-prompt thresholds; DeepSeek's price doubles at peak hours. Source: each company's API pricing page, checked October 20, 2026.
Last verified: October 20, 2026 (Claude, OpenAI, DeepSeek, MiniMax and Gemini API pricing pages; full rows with source links in O16-cheap-models-compared.csv)
Details that change the bill:
- Haiku 5.5's threshold covers the whole request. Anthropic's pricing page says a prompt over 100,000 tokens "pays the higher prices even when part of its prompt is a cache hit." Long-document work costs five times the headline rate.
- Luna's long-prompt rate applies to the full request over 272K input tokens: 2x input and cache, 1.5x output, per OpenAI's model page.
- DeepSeek's peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. Peak prices are double. Artificial Analysis lists DeepSeek at the $0.30/$1.20 peak rate.
- Batch discounts: Anthropic and Google list batch rates at half the standard price, and OpenAI's pricing page shows half-price batch and flex tiers. DeepSeek and MiniMax list no batch tier on the pages we checked.
- Reasoning tokens are billed as output. A model that thinks at length can cost more per task than its token price suggests (see the cost-per-task figures below).
Illustrative calculation, not a measurement. A job of 10 million input tokens and 2 million output tokens, every prompt under each model's long-prompt threshold, no caching, at list prices:
If the same job's prompts run over 100K tokens each, Haiku 5.5's cost rises to $10.00. For why list prices only tell part of the story, see frontier APIs vs open models.
Context, output limits and inputs
| Model | Context window | Max output | Inputs | Thinking control |
|---|---|---|---|---|
| Claude Haiku 5.5 | 1M tokens | 128,000 | Text, image | Adaptive, effort low to max (default medium) |
| GPT-6 Luna | 1,050,000 | 128,000 | Text, image | Effort none to max (default medium) |
| DeepSeek V4.1 Flash 1017 | 1M | 384K | Text, image | Not stated on the pricing page |
| MiniMax M3.1 Flash | 1,000,000 | Not stated | Text, images, video | Effort low to max (default max); can't be turned off |
| Gemini 3.8 Flash | 1,048,576 | 65,536 | Text, image, video, audio, PDF | Low, medium, high (minimal not supported) |
| Gemini 3.5 Flash-Lite | 1,048,576 | 65,536 | Text, image, video, audio, PDF | Supported |
Table 2All six take about 1M tokens of input, but maximum output runs from 65,536 tokens on Google's Flash models to 384K on DeepSeek. Source: each company's model documentation, checked October 20, 2026.
Last verified: October 20, 2026 (Anthropic models overview, OpenAI's gpt-6-luna page, DeepSeek pricing page, MiniMax "Model Invocation" docs, Gemini model pages)
Every model here now takes about 1M tokens of input. Output limits differ more: DeepSeek allows 384K, Haiku and Luna 128K, and Google's Flash models 65,536. MiniMax's docs don't state a maximum. Knowledge cutoffs: Anthropic lists June 2026 for Haiku 5.5, and OpenAI lists May 18, 2026 for Luna. Google, DeepSeek and MiniMax don't state one on the pages we checked.
Which ones have open weights?
| Model | Weights | License | Size |
|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | On Hugging Face | MIT | 763B parameters in the Hugging Face metadata; Artificial Analysis lists 552B total and 16B active |
| MiniMax M3.1 Flash | Not yet: "Model weights will be released later" | Not stated | Not disclosed |
| Claude Haiku 5.5, GPT-6 Luna, Gemini 3.8 Flash and 3.5 Flash-Lite | No (API only) | Proprietary | Not disclosed |
Table 3DeepSeek V4.1 Flash 1017 is the only model here with downloadable weights today. Sources: Hugging Face, MiniMax's October 19 post and Artificial Analysis, checked October 20, 2026.
Last verified: October 20, 2026
The DeepSeek size figures conflict because they count different things. DeepSeek's V4.1-Flash card describes 552B "backbone" parameters plus a separate memory component, while the weights' metadata counts 763B. Our DeepSeek 1017 story explains the split and the 510 GB download.
What we know about MiniMax M3.1 Flash's size and license: nothing official yet. MiniMax's October 19 post doesn't give a parameter count, and Artificial Analysis lists it as proprietary, with weights not publicly available as of October 20. MiniMax's previous model, MiniMax-M3, is on Hugging Face at about 428B total and 23B active parameters under a "minimax-community" license. That says nothing certain about M3.1 Flash. Our open-weight license comparison explains how such custom licenses differ from MIT. If you need a model you can run on a laptop, none of these qualifies; see how to choose a small AI model for your laptop.
Is MiniMax M3.1 Flash free? Not on MiniMax's API. Its pricing page lists $0.15/$0.60, and the launch post mentions no free tier. MiniMax also sells it through its M Plan subscription and MiniMax Code. Of the models here, only Google's Flash models list a free API tier.
What independent evaluators report
Three evaluators had published results for some or all of these models by October 20, 2026. Each measures something different, so we keep them in separate columns and don't combine them.
| Model | Artificial Analysis Intelligence Index v4.3.2 (variant) | AA cost per index task | AA output speed | Vals Index v2.2 (Oct 19) | Vals cost per test |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | 45 (Max) | $0.26 | 210.6 tokens/s | 61.26% | $0.66 |
| Claude Haiku 5.5 | 43 (Max); 38 (High) | $0.21 (Max) | 234.1 tokens/s (Max) | 58.71% | $3.17 |
| MiniMax M3.1 Flash | 43 (reasoning) | $0.22 | 104.7 tokens/s | 53.78% | $0.66 |
| Gemini 3.8 Flash | 41 (High) | $1.24 (High) | 149 tokens/s (High) | 55.95% | $6.06 |
| GPT-6 Luna | 38 (Max) | $0.07 | 124.2 tokens/s | 52.32% | $0.45 |
Table 4DeepSeek 1017 tops both indexes in this group, while GPT-6 Luna scores lowest but costs least per Artificial Analysis task. Results as published by Artificial Analysis and Vals AI, kept separate.
Last verified: October 20, 2026 (Artificial Analysis model and comparison pages, which carry no data date; Vals AI leaderboard updated October 19, 2026)
How to read these:
- Artificial Analysis Intelligence Index v4.3.2 is a composite of the firm's chosen evaluations, run at the effort setting shown. The spread from 38 to 45 is small. AA also reports verbosity: Haiku 5.5 (Max) used 440 million output tokens to complete the index, the most among the 68 models in that ranking, against 210 million for DeepSeek 1017, 180 million for M3.1 Flash and 140 million for Luna. That is why Luna's cost per task is the lowest. Scores from different index versions aren't comparable: AA's September 2 launch writeup gave Gemini 3.8 Flash (high) 59 on an earlier version, while today's v4.3.2 comparison page shows 41.
- Vals Index v2.2 (released October 19, replacing v2.1) is a weighted average of agentic tasks in finance, coding, legal and tax. DeepSeek 1017 leads this group at 61.26%, about 2.6 points ahead of Haiku 5.5, at $0.66 per test. M3.1 Flash, newly listed, scores 53.78% at the same cost per test. Vals doesn't show each model's reasoning setting.
- LMArena Text (snapshot dated October 15, 2026) ranks Gemini 3.8 Flash (High) ninth overall at 1496 ±5 with style control applied, from crowd votes on chat answers.
Vendor-reported results are a separate category, because each company picked the tests and settings. MiniMax reports 73.7% on DeepSWE v1.1 and 61.7 on OSWorld V2 for M3.1 Flash, without naming the comparison models. DeepSeek's model card reports 77.0 on DeepSWE v1.1 (Mini-SWE-V2 harness) and 42.3 on Terminal-Bench 4.0 (Claude Code harness) for 1017. Anthropic reports Haiku 5.5 at 39.2% on Terminal-Bench 4.0 against 16.4% for GPT-6 Luna, without saying who produced the Luna figure. Google reports 54.9% on HLE-Verified for Gemini 3.8 Flash. These can't be lined up into a ranking.
Which cheap model should you pick?
| If you need… | Start with | Why |
|---|---|---|
| The lowest token price for short prompts | Claude Haiku 5.5 or GPT-6 Luna | Both $0.10/$0.50 |
| The lowest cost per task on Artificial Analysis's index | GPT-6 Luna | $0.07 per task, the lowest here, though it scores lowest (38) |
| Long prompts over 100K tokens at low cost | GPT-6 Luna, DeepSeek 1017 or MiniMax M3.1 Flash | Haiku 5.5 rises to $0.50/$2.50 above 100K; Luna's surcharge starts at 272K; the other two list none |
| Open weights you can host yourself | DeepSeek V4.1 Flash 1017 | MIT license; MiniMax's weights aren't out yet |
| Agentic work in finance, legal or coding on a budget | DeepSeek 1017 | Highest Vals Index score in this group (61.26%) at $0.66 per test |
| Audio, video or PDF input | Gemini 3.8 Flash or 3.5 Flash-Lite | Only Google lists audio and PDF; MiniMax lists video |
| Very long outputs | DeepSeek 1017 | 384K max output |
| A free API tier to prototype | Gemini Flash models | The only free tier on these pricing pages |
| The fastest output on AA's measurements | Claude Haiku 5.5 or DeepSeek 1017 | About 234 and 211 tokens per second |
Table 5Our starting points by job, based on the prices, limits and published results above; test on your own workload before switching.
Prices and rankings in this tier change monthly. Test on your own workload before switching. For which consumer plans include these models, see Claude vs ChatGPT vs Gemini, and for the frontier models above them, Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6.
What's confirmed, reported and uncertain
- Confirmed (official pages, October 20, 2026): all prices, the Gemini 3.8 Flash price change on January 1, 2027, the long-prompt thresholds, context and output limits, input types, DeepSeek's MIT license, and MiniMax's October 19 release and "later" weights statement.
- Vendor-reported: every benchmark in the vendor paragraph, MiniMax's speed claims (200+ tokens per second) and its "70% lower" agentic cost claim.
- Third-party, with limits: the Artificial Analysis, Vals AI and LMArena results, each tied to its date and configuration.
- Uncertain:
- MiniMax M3.1 Flash's parameter count, license and weights release date
- its maximum output length
- the gap between DeepSeek's 763B and 552B parameter figures, which reflects different counting
About this storyBased on the sources linked below. Editorial standards
