The cheapest LLM APIs in 2026, ranked by real cost.
Per million tokens and per month at three real workloads, for every model we track. The cheapest overall, the cheapest in each tier, and the catches that make a cheap model expensive.
The cheapest model in each tier.
The cheapest current model is GPT OSS 20B (Groq) from Groq, at $0.13 per million tokens blended. For a growing product with 1,000 users a day, it costs about $39 a month. The most expensive current model costs about 152 times as much.
GPT OSS 20B (Groq)
- Input / output
- $0.075 / $0.30
- At 1,000 users a day
- $39 a month
Claude Haiku 5.5
- Input / output
- $0.10 / $0.50
- At 1,000 users a day
- $60 a month
Prompts over 100,000 tokens cost $0.50 input and $2.50 output. Counts about 30 percent more tokens than Haiku 4.5 for the same text
Mistral Large
- Input / output
- $0.50 / $1.50
- At 1,000 users a day
- $225 a month
57 models, one ranking.
The 15 cheapest current models
Blended price per million tokens, three input tokens to one output.
- GPT OSS 20B (Groq)$0.13
- Claude Haiku 5.5$0.20
- GPT-6 Luna$0.20
- DeepSeek V4.1 Flash$0.26
- GPT OSS 120B (Groq)$0.26
- GPT-5.6 Luna$0.45
- Mistral Large$0.75
- Gemini 3.5 Flash Lite$0.85
- DeepSeek V4 Pro$0.99
- Sonar (web aware)$1.00
- Llama 3.3 70B (Together)$1.04
- Gemini 3.7 Flash$1.50
- Gemini 3.8 Flash$1.50
- Grok 4.3$1.56
- GPT-5.4 mini$1.69
What it costs at your size.
Prices per million tokens are hard to feel. These are monthly bills for a chat style feature: five requests per active user per day, 1,500 tokens in and 500 out each, over 30 days. The cheapest picks sit next to the flagships most teams start with.
| Model | 100 users a day | 1,000 users a day | 10,000 users a day |
|---|---|---|---|
| GPT OSS 20B (Groq)Groq · Basic | $3.94 | $39 | $394 |
| Claude Haiku 5.5Anthropic · Standard | $6.00 | $60 | $600 |
| Mistral LargeMistral · Premium | $23 | $225 | $2,250 |
| Claude Sonnet 5.5Anthropic · Premium | $120 | $1,200 | $12,000 |
| GPT-6 SolOpenAI · Premium | $120 | $1,200 | $12,000 |
| Gemini 3.1 ProGoogle · Premium | $135 | $1,350 | $13,500 |
| Claude Opus 5.5Anthropic · Premium | $240 | $2,400 | $24,000 |
| GPT-6 AstraOpenAI · Premium | $600 | $6,000 | $60,000 |
Every model, cheapest first.
All 57 text models in our roster, ranked by blended price. Legacy and retiring models stay on the list because teams still run them, but they are marked and left out of the picks above.
| # | Model | Tier | Input / 1M | Output / 1M | Blended / 1M | 1,000 users a day |
|---|---|---|---|---|---|---|
| 1 | GPT OSS 20B (Groq)Groq | Basic | $0.075 | $0.30 | $0.13 | $39 |
| 2 | GPT-5 nanoOpenAIRetired or retiringSnapshot shuts down 11 Dec 2026. OpenAI names gpt-5.6-luna as the replacement | Basic | $0.050 | $0.40 | $0.14 | $41 |
| 3 | Gemini 2.5 Flash LiteGoogleLegacy | Basic | $0.10 | $0.40 | $0.18 | $53 |
| 4 | Claude Haiku 5.5AnthropicPrompts over 100,000 tokens cost $0.50 input and $2.50 output. Counts about 30 percent more tokens than Haiku 4.5 for the same text | Standard | $0.10 | $0.50 | $0.20 | $60 |
| 5 | GPT-6 LunaOpenAIRate shown is for prompts up to 272K input tokens. Above that, $0.20 input and $0.75 output | Standard | $0.10 | $0.50 | $0.20 | $60 |
| 6 | DeepSeek V4.1 FlashDeepSeekOff peak rate. Peak (01:00-04:00 and 06:00-10:00 UTC, Mon to Fri) is 2x. API name deepseek-flash; the legacy name deepseek-v4-flash is still accepted and billed at this price | Standard | $0.15 | $0.60 | $0.26 | $79 |
| 7 | GPT OSS 120B (Groq)Groq | Standard | $0.15 | $0.60 | $0.26 | $79 |
| 8 | GPT-4o miniOpenAIRetired or retiringShuts down 14 Jan 2027. OpenAI names gpt-6-luna as the replacement | Standard | $0.15 | $0.60 | $0.26 | $79 |
| 9 | Mistral Small 3MistralLegacySuperseded by Small 4. Price last verified 23 Jul 2026 | Standard | $0.20 | $0.60 | $0.30 | $90 |
| 10 | GPT-5.6 LunaOpenAI | Standard | $0.20 | $1.20 | $0.45 | $135 |
| 11 | GPT-5.4 nanoOpenAIRetired or retiringShuts down 1 Apr 2027. OpenAI names gpt-6-luna as the replacement | Basic | $0.20 | $1.25 | $0.46 | $139 |
| 12 | DeepSeek V3DeepSeekRetired or retiringRetired, superseded by V4. Price last verified 23 Jul 2026 | Standard | $0.27 | $1.10 | $0.48 | $143 |
| 13 | Grok 4 minixAIRetired or retiringNo longer listed by xAI. Price last verified 23 Jul 2026 | Standard | $0.30 | $1.20 | $0.52 | $158 |
| 14 | GPT-5 miniOpenAIRetired or retiringSnapshot shuts down 11 Dec 2026. OpenAI names gpt-5.6-terra as the replacement | Standard | $0.25 | $2.00 | $0.69 | $206 |
| 15 | GPT-4.1 miniOpenAIRetired or retiringShuts down 14 Jan 2027. OpenAI names gpt-6-luna as the replacement | Standard | $0.40 | $1.60 | $0.70 | $210 |
| 16 | Mistral LargeMistral | Premium | $0.50 | $1.50 | $0.75 | $225 |
| 17 | Gemini 2.5 FlashGoogleLegacy | Standard | $0.30 | $2.50 | $0.85 | $255 |
| 18 | Gemini 3.5 Flash LiteGoogle | Basic | $0.30 | $2.50 | $0.85 | $255 |
| 19 | DeepSeek ReasonerDeepSeekRetired or retiringRetired, superseded by V4 Pro. Price last verified 23 Jul 2026 | Premium | $0.55 | $2.19 | $0.96 | $288 |
| 20 | DeepSeek V4 ProDeepSeekOff peak rate. Peak (01:00-04:00 and 06:00-10:00 UTC, Mon to Fri) is 2x | Premium | $0.66 | $1.98 | $0.99 | $297 |
| 21 | Sonar (web aware)PerplexityPlus a per request search fee of $5 to $14 per 1,000 requests by context size | Standard | $1.00 | $1.00 | $1.00 | $300 |
| 22 | Llama 3.3 70B (Together)TogetherGroq no longer publishes a per token price for this model (contact sales), so this is the Together rate | Standard | $1.04 | $1.04 | $1.04 | $312 |
| 23 | Gemini 3.7 FlashGoogleScheduled increase to $1.50 input and $7.50 output on 1 Jan 2027 | Standard | $0.75 | $3.75 | $1.50 | $450 |
| 24 | Gemini 3.8 FlashGoogleScheduled increase to $1.50 input and $7.50 output on 1 Jan 2027 | Standard | $0.75 | $3.75 | $1.50 | $450 |
| 25 | Grok 4.3xAIRate doubles on prompts of 200k tokens or more | Standard | $1.25 | $2.50 | $1.56 | $469 |
| 26 | GPT-5.4 miniOpenAI | Standard | $0.75 | $4.50 | $1.69 | $506 |
| 27 | o3-mini (reasoning)OpenAIRetired or retiringShut down 23 Oct 2026. OpenAI names gpt-5.6-sol as the replacement | Premium | $1.10 | $4.40 | $1.93 | $578 |
| 28 | o4-mini (reasoning)OpenAIRetired or retiringShut down 23 Oct 2026. OpenAI names gpt-5.6-terra as the replacement | Premium | $1.10 | $4.40 | $1.93 | $578 |
| 29 | Claude Haiku 4.5Anthropic | Standard | $1.00 | $5.00 | $2.00 | $600 |
| 30 | Grok 4.6xAIRate doubles on prompts of 200k tokens or more | Premium | $2.00 | $6.00 | $3.00 | $900 |
| 31 | Grok 4.7xAIRate doubles on prompts of 200k tokens or more | Premium | $2.00 | $6.00 | $3.00 | $900 |
| 32 | Mistral Large 2MistralLegacySuperseded. Price last verified 23 Jul 2026 | Premium | $2.00 | $6.00 | $3.00 | $900 |
| 33 | Gemini 3.5 FlashGoogle | Standard | $1.50 | $9.00 | $3.38 | $1,013 |
| 34 | Gemini 2.5 ProGoogleLegacy | Premium | $1.25 | $10 | $3.44 | $1,031 |
| 35 | GPT-5OpenAIRetired or retiringSnapshot shuts down 11 Dec 2026. OpenAI names gpt-5.6-sol as the replacement | Premium | $1.25 | $10 | $3.44 | $1,031 |
| 36 | Llama 3.1 405BTogetherRetired or retiringNo longer listed by Together. Price last verified 23 Jul 2026 | Premium | $3.50 | $3.50 | $3.50 | $1,050 |
| 37 | o3 (reasoning)OpenAIRetired or retiringSnapshots shut down 11 Dec 2026. OpenAI points to the GPT-5.6 family | Premium | $2.00 | $8.00 | $3.50 | $1,050 |
| 38 | Sonar Reasoning ProPerplexityPlus a per request search fee of $5 to $14 per 1,000 requests by context size | Premium | $2.00 | $8.00 | $3.50 | $1,050 |
| 39 | Claude Sonnet 5Anthropic | Premium | $2.00 | $10 | $4.00 | $1,200 |
| 40 | Claude Sonnet 5.5AnthropicCache reads $0.10 (cut from $0.20 on 7 October 2026). Uses Sonnet 5 tokenizer: about 30% more tokens than Sonnet 4.6, 4.5 or Haiku 4.5 for the same text, so compare on token counts, not rates | Premium | $2.00 | $10 | $4.00 | $1,200 |
| 41 | GPT-6 SolOpenAIRate shown is for prompts up to 272K input tokens. Above that, $4 input and $15 output | Premium | $2.00 | $10 | $4.00 | $1,200 |
| 42 | GPT-6.1 SolOpenAIRate shown is for prompts up to 272K input tokens. Above that, $4 input and $15 output. Cached input $0.10, half the rate on GPT-6 Sol | Premium | $2.00 | $10 | $4.00 | $1,200 |
| 43 | Command R+CohereLegacyCohere lists this under legacy models, available to existing customers | Premium | $2.50 | $10 | $4.38 | $1,313 |
| 44 | GPT-4.1OpenAIRetired or retiringShuts down 14 Jan 2027. OpenAI names gpt-6-luna as the replacement | Premium | $2.50 | $10 | $4.38 | $1,313 |
| 45 | GPT-4oOpenAIRetired or retiringShuts down 14 Jan 2027. OpenAI names gpt-6-luna as the replacement | Premium | $2.50 | $10 | $4.38 | $1,313 |
| 46 | Gemini 3.1 ProGoogle | Premium | $2.00 | $12 | $4.50 | $1,350 |
| 47 | GPT-5.6 TerraOpenAI | Premium | $2.00 | $12 | $4.50 | $1,350 |
| 48 | GPT-5.4OpenAI | Premium | $2.50 | $15 | $5.63 | $1,688 |
| 49 | Grok 4xAIRetired or retiringNo longer listed by xAI. Price last verified 23 Jul 2026 | Premium | $3.00 | $15 | $6.00 | $1,800 |
| 50 | Sonar Pro (web aware)PerplexityPlus a per request search fee of $5 to $14 per 1,000 requests by context size | Premium | $3.00 | $15 | $6.00 | $1,800 |
| 51 | Claude Opus 5.5AnthropicCache hits are 0.05x base input ($0.20), not the usual 0.1x. Thinking cannot be disabled and tool_choice any/tool return 400 | Premium | $4.00 | $20 | $8.00 | $2,400 |
| 52 | GPT-5.6 SolOpenAIOpenAI describes this as promotional pricing, available at least through 21 Nov 2026 | Premium | $4.00 | $20 | $8.00 | $2,400 |
| 53 | Claude Opus 5Anthropic | Premium | $5.00 | $25 | $10 | $3,000 |
| 54 | GPT-5.5OpenAI | Premium | $5.00 | $30 | $11 | $3,375 |
| 55 | Claude Fable 5.1Anthropic | Premium | $10 | $50 | $20 | $6,000 |
| 56 | GPT-6 AstraOpenAIRate shown is for prompts up to 272K input tokens. Above that, $20 input and $75 output | Premium | $10 | $50 | $20 | $6,000 |
| 57 | Claude Opus 4.8AnthropicLegacy | Premium | $15 | $75 | $30 | $9,000 |
Six ways a cheap model gets expensive.
Long context surcharges
Several models double the rate past a prompt length, 200,000 or 272,000 tokens. Retrieval apps that stuff context can cross it without noticing.
Peak hour pricing
DeepSeek charges twice the off peak rate during set UTC windows on weekdays. A bill modelled at the off peak rate can double.
Per request fees
Web aware models such as Perplexity Sonar add a search fee per thousand requests on top of token prices.
Promotions and scheduled rises
Some prices carry an end date or a published increase. Budget at the later price, not the launch one.
Output heavy work
Output costs several times input on most models. Long answers, code and reasoning tokens move the bill more than the blend shows.
Retries and evals
A cheap model that needs two attempts costs double. Measure cost per finished task, not per request.
How this is worked out.
Every price comes from the roster the Cost Comparator runs on, reconciled against each provider’s own pricing page and dated 23 Oct 2026. Nothing is taken from aggregators. Blended price assumes three input tokens for every output token. Monthly figures assume 1,500 input and 500 output tokens per request, five requests per active user per day, 30 days, with no caching or batch discounts.
Tiers are our grouping by price band and the job each provider positions the model for. They are not a benchmark score. Legacy means the provider has moved the model to a legacy list or replaced it; retiring means it is retired, delisted, or has a published shutdown date.
Common questions.
What is the cheapest LLM API in 2026?
On the prices we checked on 23 Oct 2026, the cheapest current model is GPT OSS 20B (Groq) from Groq, at $0.075 per million input tokens and $0.30 per million output tokens, or $0.13 per million on a three to one blend. For a product with 1,000 daily users making five requests each, that is about $39 a month.
What is the cheapest model in each tier?
Basic: GPT OSS 20B (Groq) at $0.13 per million tokens blended. Standard: Claude Haiku 5.5 at $0.20 per million tokens blended. Premium: Mistral Large at $0.75 per million tokens blended. Tiers are our own grouping of each model by the price band and the job its provider positions it for.
How much cheaper is the cheapest model than the most expensive?
About 152 times. Claude Fable 5.1 costs $20 per million tokens blended against $0.13 for GPT OSS 20B (Groq). That gap is why routing simple requests to a cheap model and only hard ones to a premium model cuts most bills sharply.
What does blended price mean?
It is the price of a million tokens at three input tokens for every output token, a common shape for chat and retrieval requests. Output tokens cost more on almost every model, so a workload that writes long answers costs more than the blend suggests. The tables also show input and output prices separately.
Is the cheapest model good enough for my product?
Sometimes. Cheap models handle classification, extraction, routing and short answers well, and they struggle more with long reasoning and tool use. The way to know is to run your own prompts through two or three candidates and compare. The Cost Comparator shows what each one would cost you at your real volume.
What catches can make a cheap model cost more?
Long context surcharges (several models double the rate past a prompt length), peak hour pricing, per request search fees on web aware models, promotional prices with an end date, and scheduled price rises. Every row with one of these carries a note in the full table.
How often are these prices updated?
The roster behind this page is reconciled against each provider's own pricing page; the current version is dated 23 Oct 2026. Every change is dated and sourced in the price change log.
Model your real bill.
Set your users, requests and tokens and see all 57 text models priced for your volume, with a cheap, mid and premium pick.
Price my own usage