LLM Price Index
Token and prompt cache prices, one table
Input, output, cached input, cache write and batch pricing for 36 models from 10 providers and hosts. Search, sort, then price your own cache hit rate.
| Line | Tokens | Rate / 1M | Cost |
|---|---|---|---|
| Input, uncached | 15,000 | $3.00 | $0.0450 |
| Input, cache read | 135,000 | $0.30 | $0.0405 |
| Output | 2,000 | $15.00 | $0.0300 |
- Models tracked
- 36
- Providers and hosts
- 10
- With cache-read pricing
- 26
- Lowest blended 3:1
- $0.058
- Last full review
- June 2026
The index
Every model, every token type
Click a column to sort. Search syncs to the URL, so a filtered view is a link you can share. Model names link to full cost guides.
36 of 36 models · USD per 1M tokens · sorted by blended
| Tier | Batch | ||||||||
|---|---|---|---|---|---|---|---|---|---|
 Llama 3.1 8B Instant | Meta (Groq) | Small | $0.050 | $0.080 | none | none | Yes | 128K | $0.058 |
 Amazon Nova Micro | Amazon | Small | $0.035 | $0.14 | $0.009 | none | Yes | 128K | $0.061 |
| Amazon Nova Lite | Amazon | Small | $0.060 | $0.24 | $0.015 | none | Yes | 300K | $0.10 |
 GPT-5 nano | OpenAI | Small | $0.050 | $0.40 | $0.005 | none | Yes | 400K | $0.14 |
 Mistral Small 3.1 | Mistral | Small | $0.10 | $0.30 | none | none | Yes | 128K | $0.15 |
| GPT-4.1 nano | OpenAI | Small | $0.10 | $0.40 | $0.025 | none | Yes | 1M | $0.18 |
 Gemini 2.5 Flash-Lite | Small | $0.10 | $0.40 | $0.025 | none | Yes | 1M | $0.18 | |
| Gemini 2.0 Flash | Small | $0.10 | $0.40 | $0.025 | none | Yes | 1M | $0.18 | |
| GPT-4o mini | OpenAI | Small | $0.15 | $0.60 | $0.075 | none | Yes | 128K | $0.26 |
 Command R | Cohere | Small | $0.15 | $0.60 | none | none | No | 128K | $0.26 |
| Llama 4 Maverick | Meta (Groq) | Mid | $0.20 | $0.60 | none | none | Yes | 128K | $0.30 |
 DeepSeek V3.2 (chat) | DeepSeek | Mid | $0.28 | $0.42 | $0.028 | none | No | 128K | $0.32 |
| DeepSeek V3.2 (reasoner) | DeepSeek | Reasoning | $0.28 | $0.42 | $0.028 | none | No | 128K | $0.32 |
| Codestral | Mistral | Mid | $0.30 | $0.90 | none | none | Yes | 256K | $0.45 |
| Llama 3.3 70B | Meta (Groq) | Mid | $0.59 | $0.79 | none | none | Yes | 128K | $0.64 |
| GPT-5 mini | OpenAI | Small | $0.25 | $2.00 | $0.025 | none | Yes | 400K | $0.69 |
| GPT-4.1 mini | OpenAI | Small | $0.40 | $1.60 | $0.10 | none | Yes | 1M | $0.70 |
| Mistral Medium 3 | Mistral | Mid | $0.40 | $2.00 | none | none | Yes | 128K | $0.80 |
| Gemini 2.5 Flash | Mid | $0.30 | $2.50 | $0.075 | none | Yes | 1M | $0.85 | |
 Llama 3.3 70B Turbo | Meta (Together) | Mid | $0.88 | $0.88 | none | none | Yes | 128K | $0.88 |
| Amazon Nova Pro | Amazon | Mid | $0.80 | $3.20 | $0.20 | none | Yes | 300K | $1.40 |
 Claude 3.5 Haiku | Anthropic | Small | $0.80 | $4.00 | $0.080 | $1.00 | Yes | 200K | $1.60 |
| o4-mini | OpenAI | Reasoning | $1.10 | $4.40 | $0.28 | none | Yes | 200K | $1.93 |
| Claude Haiku 4.5 | Anthropic | Small | $1.00 | $5.00 | $0.10 | $1.25 | Yes | 200K | $2.00 |
| Mistral Large 2 | Mistral | Frontier | $2.00 | $6.00 | none | none | Yes | 128K | $3.00 |
| GPT-5 | OpenAI | Frontier | $1.25 | $10.00 | $0.13 | none | Yes | 400K | $3.44 |
| Gemini 2.5 Pro | Frontier | $1.25 | $10.00 | $0.31 | none | Yes | 1M | $3.44 | |
| GPT-4.1 | OpenAI | Frontier | $2.00 | $8.00 | $0.50 | none | Yes | 1M | $3.50 |
| o3 | OpenAI | Reasoning | $2.00 | $8.00 | $0.50 | none | Yes | 200K | $3.50 |
| GPT-4o | OpenAI | Frontier | $2.50 | $10.00 | $1.25 | none | Yes | 128K | $4.38 |
| Command R+ | Cohere | Frontier | $2.50 | $10.00 | none | none | No | 128K | $4.38 |
| Claude Sonnet 4.5 | Anthropic | Frontier | $3.00 | $15.00 | $0.30 | $3.75 | Yes | 200K | $6.00 |
| Claude Sonnet 4 | Anthropic | Frontier | $3.00 | $15.00 | $0.30 | $3.75 | Yes | 200K | $6.00 |
 Grok 4 | xAI | Frontier | $3.00 | $15.00 | $0.75 | none | No | 256K | $6.00 |
| Claude Opus 4.5 | Anthropic | Frontier | $5.00 | $25.00 | $0.50 | $6.25 | Yes | 200K | $10.00 |
| Claude Opus 4.1 | Anthropic | Frontier | $15.00 | $75.00 | $1.50 | $18.75 | Yes | 200K | $30.00 |
Standard on-demand tier, shortest context band. Cache write is the Anthropic five-minute TTL write at 1.25x input. Gemini 2.5 Pro and Claude Sonnet long-context requests above 200K input tokens are billed at a higher band. Reasoning and thinking tokens are billed at the output rate.
Cache savings calculator
What your hit rate is worth
Caching discounts the stable prefix of a prompt: system instructions, tool schemas, documents. Set how much of your input is cacheable and how often it hits.
- OpenAI, Google, DeepSeek, xAI: reads are discounted, misses cost the normal input rate.
- Anthropic: reads cost a tenth of input, writes cost 1.25x. Low hit rates lose money.
- Amazon Bedrock: Nova cache reads are discounted on the cacheable prefix.
- Without caching
- $938
- With caching
- $501
Monthly saving
$43747%
Effective input rate $1.25 per 1M vs $3.00 list
- Uncached input$150
- Cache reads$51
- Cache writes$113
- Output$188
Anthropic bills cache writes at $3.75 per 1M, so caching pays off only above a 22% hit rate on the cacheable prefix.
- Llama 3.1 8B Instant$0.058
- Amazon Nova Micro$0.061
- Amazon Nova Lite$0.10
- GPT-5 nano$0.14
- Mistral Small 3.1$0.15
- GPT-4.1 nano$0.18
- Gemini 2.5 Flash-Lite$0.18
- Gemini 2.0 Flash$0.18
- GPT-4o mini$0.26
- Command R$0.26
- Llama 4 Maverick$0.30
- DeepSeek V3.2 (chat)$0.32
- DeepSeek V3.2 (reasoner)$0.32
- Codestral$0.45
- Llama 3.3 70B$0.64
- GPT-5 mini$0.69
- GPT-4.1 mini$0.70
- Mistral Medium 3$0.80
- Gemini 2.5 Flash$0.85
- Llama 3.3 70B Turbo$0.88
- Amazon Nova Pro$1.40
- Claude 3.5 Haiku$1.60
- o4-mini$1.93
- Claude Haiku 4.5$2.00
- Mistral Large 2$3.00
- GPT-5$3.44
- Gemini 2.5 Pro$3.44
- GPT-4.1$3.50
- o3$3.50
- GPT-4o$4.38
- Command R+$4.38
- Claude Sonnet 4.5$6.00
- Claude Sonnet 4$6.00
- Grok 4$6.00
- Claude Opus 4.5$10.00
- Claude Opus 4.1$30.00
Blended cost
A 522x spread at the same traffic shape
At 3:1, Llama 3.1 8B Instant costs $0.058 per 1M blended tokens and Claude Opus 4.1 costs $30.00. Most production traffic does not need the top of this chart.
The log scale keeps the cheapest models readable; switch to linear to see how far the frontier tier sits from the rest. Change the ratio to match RAG (input heavy) or generation (output heavy) traffic.
Route over-budget tenants to cheaper modelsMethodology
How the index is built
Source
Each row is copied from the provider's public pricing page for the standard on-demand tier, in USD per 1M tokens.
Context band
Where a provider charges more above a prompt length threshold, the index lists the shortest band and notes the surcharge in the model guide.
Cache columns
Cached input is the published cache-read or cache-hit price. Cache write is listed only where it is billed separately.
Review
Price changes are applied as they are announced, with a full line-by-line review. The last full review was June 2026.
Latency
We do not publish TTFT or throughput figures. PALM SOUQ measures both per request on the gateway and reports them per tenant, model and region from your own traffic.
Cite this dataset
Free for developers, researchers and journalists under Creative Commons Attribution 4.0. Please credit PALM SOUQ and link back to this page.
PALM SOUQ. (2026). PALM SOUQ LLM Price Index: token and prompt cache prices, last reviewed June 2026. https://palmsouqltd.com/resources/llm-pricing-index. Licensed CC BY 4.0.
Found a price that changed? Email support@palmsouqltd.com with the provider link and we will correct the row.
Price index questions
How often is the LLM Price Index updated?
Provider price changes are applied as they are announced, and the whole table gets a full line-by-line review against provider pricing pages on a regular schedule. The last full review was in June 2026. Each row is the standard on-demand tier at the shortest context band.
What is a blended price?
A blended price weights input and output tokens by a typical traffic shape. At 3:1, the default, it is (3 x input price + output price) / 4 per 1M tokens. Chat and agent traffic usually sends far more input than it receives as output, so input price matters more than a simple average suggests.
Why do some models show no cached input price?
Some providers do not publish a discounted rate for cached prompt tokens on that model, or serve open-weight models without a cache discount. For those rows every input token is billed at the full input rate, and the cache calculator excludes them.
Do you publish latency or throughput benchmarks?
No. Time to first token and output tokens per second depend on region, prompt size, time of day and your account's capacity, so a single published number misleads more than it helps. PALM SOUQ records both on every request that passes through the gateway and reports them per tenant, per model and per region from your own traffic.
Are reasoning tokens included in the output price?
Yes. OpenAI o-series and GPT-5, Gemini 2.5 thinking models, Claude extended thinking, Grok 4 and DeepSeek's reasoner all bill hidden reasoning tokens at the output rate. That is why a reasoning model with a modest list price can cost more per answer than a non-reasoning model with a higher one.
Can I cite or reuse this data?
Yes. The index is published under the Creative Commons Attribution 4.0 license. Cite it as the PALM SOUQ LLM Price Index with a link to https://palmsouqltd.com/resources/llm-pricing-index. Use the Copy as CSV control above the table to export the rows you have filtered.
List prices are the start. Per-tenant cost is the answer.
PALM SOUQ applies these prices to every request your customers make, including cache reads, cache writes and reasoning tokens, so you see gross margin per customer instead of one provider invoice.
14-day free production trial · 100k requests included · no card required