Compare / LLM APIs
Hosted LLM API pricing, normalized
Uncached list rates, checked 18 Oct 2026. The cheapest published cheap/fast model is Groq Llama 3.1 8B Instant at $0.05 / $0.08 per 1M input/output. Among flagships, Mistral Large 4 is $0.68 / $2.09 on the listed sale (regular $1.36 / $4.18), DeepSeek V4 Pro peak cache-miss is $1.32 / $3.96, and Groq Llama 3.3 70B is $0.59 / $0.79. Frontier APIs cost more: grok-4.7 $2 / $6, Claude Sonnet 5.5 and Gemini 4 Argon $2 / $10, GPT-5.2 $1.75 / $14. OpenRouter does not have its own model rates — it passes through providers (plus 5.5% on credits), so those cells are em dashes.
Prices from each provider's official pricing page, checked 18 Oct 2026. USD, excluding tax. Headline cells are uncached list rates. Cached-input, batch, off-peak, and sale-regular figures are in the cell tooltips, not the headline. Blended = (3 × input + 1 × output) ÷ 4, a 3:1 input:output mix. "Not published" means the page did not list a first-party rate. Estimates, not guarantees.
Price per 1,000,000 tokens, uncached list rates
OpenAI Details
| $1.75 GPT-5.2 uncached. Cached input $0.175. Batch 50% off. | $14.00 GPT-5.2. Batch 50% off ($7). | $4.81 (3 × $1.75 + $14.00) ÷ 4 = $4.8125. Cached/batch in the input and output tooltips. | GPT-5-mini $0.25 / $2.00 per 1M input/output. Cached input $0.025. Batch 50% off. |
|---|---|---|---|---|
Anthropic Details
| $2.00 Sonnet 5.5 uncached, prompts ≤200k. Cache-write 1.25×; cache-hit 0.1×. Batch 50% off. | $10.00 Sonnet 5.5. Batch 50% off ($5). | $4.00 (3 × $2.00 + $10.00) ÷ 4. Cache/batch in the input and output tooltips. | Haiku 5.5 $0.10 / $0.50 per 1M, prompts ≤100k. Above 100k: $0.20 / $2.50. Cache-hit and Batch 50% off in the Anthropic table. |
Google Gemini Details
| $2.00 Gemini 4 Argon through 30 Nov 2026. Then $4.00. Batch 50% off. Cached input is cheaper. | $10.00 Gemini 4 Argon through 30 Nov 2026. Then $20.00. Batch 50% off. | $4.00 (3 × $2.00 + $10.00) ÷ 4 through 30 Nov 2026. After that (3 × $4 + $20) ÷ 4 = $8.00. | 3.5 Flash-Lite $0.30 / $2.50 per 1M input/output. Batch 50% off. |
xAI Details
| $2.00 grok-4.7, prompts <200k. Above 200k: $4.00. Cached input is cheaper. | $6.00 grok-4.7, prompts <200k. Above 200k: $12.00. | $3.00 (3 × $2.00 + $6.00) ÷ 4 below 200k. Above 200k: (3 × $4 + $12) ÷ 4 = $6.00. | grok-4.20 $1.25 / $2.50 per 1M input/output. Cached input cheaper. |
Mistral AI Details
| $0.68 Large 4 listed sale. Regular $1.36. Cached input 0.1× ($0.136 regular / $0.068 sale). | $2.09 Large 4 listed sale. Regular $4.18. | $1.03 Sale (3 × $0.68 + $2.09) ÷ 4 = $1.0325. Regular blended $2.07. | Small 4 $0.15 / $0.60 per 1M input/output. Cached input 0.1×. |
DeepSeek Details
| $1.32 V4 Pro peak, cache-miss. Cache-hit is much cheaper. Off-peak is half. | $3.96 V4 Pro peak. Off-peak is half ($1.98). | $1.98 (3 × $1.32 + $3.96) ÷ 4 peak cache-miss. Off-peak half. Cache-hit omitted from the headline. | Flash $0.30 / $1.20 per 1M peak cache-miss. Off-peak half. Cache-hit input cheaper. |
Groq Details
| $0.59 Llama 3.3 70B on-demand. Not a frontier model — a fast hosted open weight. | $0.79 Llama 3.3 70B on-demand. | $0.64 (3 × $0.59 + $0.79) ÷ 4 = $0.64. | Llama 3.1 8B Instant $0.05 / $0.08 per 1M input/output. GPT OSS 120B is $0.15 / $0.60. |
Together AI Details
| $1.40 GLM-5.3 serverless. Cached input cheaper. Batch up to 50% off. | $4.40 GLM-5.3 serverless. Batch up to 50% off. | $2.15 (3 × $1.40 + $4.40) ÷ 4. Cache/batch in the input and output tooltips. | GLM-5.3 Flash $0.15 / $0.50 per 1M input/output. Cached input cheaper. |
Fireworks AI Details
| $1.40 GLM 5.3 serverless. Cached input cheaper. Batch 50% off. | $4.40 GLM 5.3 serverless. Batch 50% off. | $2.15 (3 × $1.40 + $4.40) ÷ 4. Cache/batch in the input and output tooltips. | GLM 5.3 Flash $0.15 / $0.50 per 1M input/output. Cached input cheaper. Batch 50% off. |
OpenRouter Details
| — Pass-through of each provider's price, not a first-party model table. Credits add 5.5%. | — Pass-through of each provider's price, not a first-party model table. Credits add 5.5%. | — Pass-through; no first-party blended rate to publish. | Pass-through Same model can route to several hosts. Pay-as-you-go credits add 5.5% on top of the provider price. |
What you actually get
| OpenAI | GPT-5.2 | GPT-5-mini | Standard list; longer-context SKUs are separate rows on the pricing page | Cached input 0.1×; Batch 50% off (tooltip only) | OpenAI API |
|---|---|---|---|---|---|
| Anthropic | Claude Sonnet 5.5 | Claude Haiku 5.5 | Sonnet list is ≤200k prompt tokens; Haiku cheap rate is ≤100k | Prompt-cache hits and Batch 50% off (tooltip only) | Anthropic API / Amazon Bedrock / Google Vertex (first-party rates here) |
| Google Gemini | Gemini 4 Argon | Gemini 3.5 Flash-Lite | Argon promo through 30 Nov 2026, then $4 / $20 | Batch 50% off; cached input cheaper (tooltip only) | Gemini Developer API (ai.google.dev) |
| xAI | grok-4.7 | grok-4.20 | Flagship list is <200k prompt tokens; 200k+ is 2× | Cached input cheaper (tooltip only) | xAI API |
| Mistral AI | Mistral Large 4 | Mistral Small 4 | Sale is the listed current rate; regular is in the tooltip | Cached input 0.1× (tooltip only) | Mistral API (La Plateforme) |
| DeepSeek | DeepSeek V4 Pro | DeepSeek Flash | Peak cache-miss is the conservative headline; off-peak is half | Cache-hit and off-peak in the tooltip, not the cell | DeepSeek API |
| Groq | Llama 3.3 70B | Llama 3.1 8B Instant | On-demand list; reserved/compound SKUs are separate | Not used as a headline discount | GroqCloud |
| Together AI | GLM-5.3 | GLM-5.3 Flash | Serverless list; dedicated endpoints are a different product | Cached input and Batch up to 50% off (tooltip only) | Together serverless |
| Fireworks AI | GLM 5.3 | GLM 5.3 Flash | Serverless list; on-demand deployments are a different product | Cached input cheaper; Batch 50% off (tooltip only) | Fireworks serverless |
| OpenRouter | Whatever model you pick | Whatever model you pick | No first-party list rate — omitted rather than copied | Some listed models are $0; that is the host, not OpenRouter | Router in front of many providers |
Best for
- OpenAI
The default frontier API when you want GPT-5.2 and will pay the output rate.
- Anthropic
Claude quality at a published $2 / $10 Sonnet rate, or Haiku at $0.10 / $0.50 for cheap classification.
- Google Gemini
Gemini 4 Argon at the current $2 / $10 promo, or Flash-Lite when cheap/fast Google is enough.
- xAI
Grok when you want a $2 / $6 flagship (cheaper output than Sonnet or GPT-5.2).
- Mistral AI
The cheapest published frontier-class flagship on the current Large 4 sale.
- DeepSeek
A cheap capable flagship when you will miss cache at peak and still beat Sonnet on price.
- Groq
The cheapest published tokens when an 8B or 70B Llama on an LPU is enough.
- Together AI
Hosted GLM-5.3 when you want an open model without running a GPU yourself.
- Fireworks AI
The same GLM 5.3 serverless rates as Together, on Fireworks' API.
- OpenRouter
One key in front of many hosts, knowing the cell cannot show a single official rate.
A $0.05/1M cheap model vs a $2/$10 frontier API vs a pass-through router
These pages compare hosted inference APIs, not self-hosted vLLM. One row is one provider. The headline number is the flagship's uncached list rate. A representative cheap or fast sibling sits in the last column as text so it does not mix into the flagship color scale. Cached-input hits, Batch (often 50% off), DeepSeek off-peak (half), Mistral's listed sale vs regular, and Gemini 4 Argon's post-30 Nov 2026 rate live in tooltips.
Blended 3:1 is (3 × input + 1 × output) ÷ 4. That is a common chat mix — more prompt than completion — stated on every page that uses it. A 1:1 mix would move the blended number toward the output rate. Worked examples use the same 3:1 ratio at three monthly volumes.
OpenRouter is a router, not a model host. It charges the provider's price and adds 5.5% when you pay with account credits. Those cells are em dashes rather than a copied vendor rate, because the official page does not publish a first-party model table of its own.
What would a real project cost?
Estimates, not guarantees. Built from the published rates above (checked 18 Oct 2026); check each provider's current pricing before you commit. Tax is not included. Numbers marked unverified or not published were left out rather than guessed.
- Small: hobby chatbot. 2,000,000 input tokens and 667,000 output tokens a month on the cheap/fast model (3:1). A personal bot, light RAG, or weekend prototype. Tax not included.
- Medium: production SaaS feature. 20,000,000 input tokens and 6,670,000 output tokens a month on the flagship (3:1). One product feature: support replies, an agent, or structured extraction. Uncached list rates. Tax not included.
- Large: high-volume pipeline. 200,000,000 input tokens and 66,700,000 output tokens a month on the flagship (3:1). A batch or streaming pipeline. Headline still uses uncached list rates — real bills drop if you hit cache or Batch. Tax not included.
| OpenAI | $1.83 GPT-5-mini | $128.38 GPT-5.2 uncached | $1,283.80 GPT-5.2 uncached |
|---|---|---|---|
| Anthropic | $0.53 Haiku 5.5 ≤100k | $106.70 Sonnet 5.5 | $1,067.00 Sonnet 5.5 |
| Google Gemini | $2.27 3.5 Flash-Lite | $106.70 Gemini 4 Argon promo | $1,067.00 Gemini 4 Argon promo |
| xAI | $4.17 grok-4.20 | $80.02 grok-4.7 <200k | $800.20 grok-4.7 <200k |
| Mistral AI | $0.70 Small 4 | $27.54 Large 4 sale; regular ~$55.08 | $275.40 Large 4 sale; regular ~$550.80 |
| DeepSeek | $1.40 Flash peak cache-miss | $52.81 V4 Pro peak cache-miss | $528.13 V4 Pro peak cache-miss |
| Groq | $0.15 Llama 3.1 8B Instant | $17.07 Llama 3.3 70B | $170.69 Llama 3.3 70B |
| Together AI | $0.63 GLM-5.3 Flash | $57.35 GLM-5.3 | $573.48 GLM-5.3 |
| Fireworks AI | $0.63 GLM 5.3 Flash | $57.35 GLM 5.3 | $573.48 GLM 5.3 |
| OpenRouter | — Pass-through; no first-party rate | — Pass-through; no first-party rate | — Pass-through; no first-party rate |
Small: hobby chatbot
Try first: Groq Llama 3.1 8B Instant at about $0.15 — cheapest published cheap-model bill at this volume.
Runner-up: Anthropic Haiku 5.5 at about $0.53, or Mistral Small 4 at about $0.70, when you want a provider model instead of an 8B Llama.
Avoid: OpenRouter as a priced row (pass-through, omitted) and frontier flagships for a hobby bot — GPT-5.2 at this volume on the cheap SKU is already about $1.83, and the flagship would be about $12.84.
Medium: production SaaS feature
Try first: Mistral Large 4 at about $27.54 on the listed sale — cheapest published frontier-class flagship at this volume (regular would be about $55.08).
Runner-up: DeepSeek V4 Pro peak cache-miss at about $52.81, or Groq Llama 3.3 70B at about $17.07 if a fast 70B open weight is enough and you do not need a frontier model.
Avoid: Paying GPT-5.2 ($128.38) or Sonnet/Argon ($106.70) for this volume when Mistral Large 4 or DeepSeek Pro still fit, and treating OpenRouter as a priced platform.
Large: high-volume pipeline
Try first: Mistral Large 4 at about $275.40 on the listed sale (regular about $550.80).
Runner-up: DeepSeek V4 Pro peak cache-miss at about $528.13, or Groq Llama 3.3 70B at about $170.69 if the 70B quality is enough. Batch/cache (tooltips) can cut these further.
Avoid: GPT-5.2 at about $1,283.80 uncached, and OpenRouter as a first-party price. Hyperscaler model gardens are not extra rows — use the first-party pages above.
Show the math
- Project — Small: hobby chatbot
- 2,000,000 input tokens and 667,000 output tokens a month on the cheap/fast model (3:1). A personal bot, light RAG, or weekend prototype. Tax not included. Cheap 3:1: OpenAI mini 2×$0.25 + 0.667×$2.00 = $1.83. Haiku 2×$0.10 + 0.667×$0.50 = $0.53. Flash-Lite 2×$0.30 + 0.667×$2.50 = $2.27. grok-4.20 2×$1.25 + 0.667×$2.50 = $4.17. Mistral Small 2×$0.15 + 0.667×$0.60 = $0.70. DeepSeek Flash 2×$0.30 + 0.667×$1.20 = $1.40. Groq 8B 2×$0.05 + 0.667×$0.08 = $0.15. GLM Flash 2×$0.15 + 0.667×$0.50 = $0.63. OpenRouter omitted.
- Project — Medium: production SaaS feature
- 20,000,000 input tokens and 6,670,000 output tokens a month on the flagship (3:1). One product feature: support replies, an agent, or structured extraction. Uncached list rates. Tax not included. Flagship 3:1: GPT-5.2 20×$1.75 + 6.67×$14.00 = $128.38. Sonnet/Argon 20×$2.00 + 6.67×$10.00 = $106.70. grok-4.7 20×$2.00 + 6.67×$6.00 = $80.02. Mistral Large sale 20×$0.68 + 6.67×$2.09 = $27.54. DeepSeek Pro 20×$1.32 + 6.67×$3.96 = $52.81. Groq 70B 20×$0.59 + 6.67×$0.79 = $17.07. GLM 20×$1.40 + 6.67×$4.40 = $57.35. OpenRouter omitted.
- Project — Large: high-volume pipeline
- 200,000,000 input tokens and 66,700,000 output tokens a month on the flagship (3:1). A batch or streaming pipeline. Headline still uses uncached list rates — real bills drop if you hit cache or Batch. Tax not included. 10× the medium flagship bills: GPT-5.2 $1,283.80. Sonnet/Argon $1,067.00. grok-4.7 $800.20. Mistral Large sale $275.40. DeepSeek Pro $528.13. Groq 70B $170.69. GLM $573.48. OpenRouter omitted.
- OpenAI
- Official: GPT-5.2 $1.75 / $14 per 1M; cached input $0.175; Batch 50% off. GPT-5-mini $0.25 / $2.00; cached $0.025. Blended 3:1 = (3×1.75 + 14) ÷ 4 = $4.81. Source: https://developers.openai.com/api/docs/pricing (checked 18 Oct 2026)
- Anthropic
- Official: Sonnet 5.5 $2 / $10 per 1M (≤200k). Haiku 5.5 $0.10 / $0.50 (≤100k); $0.20 / $2.50 above. Cache-write 1.25×, cache-hit 0.1×, Batch 50% off. Blended 3:1 = $4.00. Source: https://platform.claude.com/docs/en/about-claude/pricing (checked 18 Oct 2026)
- Google Gemini
- Official: Gemini 4 Argon $2 / $10 per 1M through 30 Nov 2026, then $4 / $20. Gemini 3.5 Flash-Lite $0.30 / $2.50. Batch 50% off. Blended 3:1 on the promo = $4.00. Source: https://ai.google.dev/gemini-api/docs/pricing (checked 18 Oct 2026)
- xAI
- Official: grok-4.7 $2 / $6 per 1M below 200k; $4 / $12 above. grok-4.20 $1.25 / $2.50. Cached input cheaper. Blended 3:1 below 200k = $3.00. Source: https://docs.x.ai/developers/pricing (checked 18 Oct 2026)
- Mistral AI
- Official: Large 4 $0.68 / $2.09 sale (regular $1.36 / $4.18). Small 4 $0.15 / $0.60. Cached input 0.1×. Blended 3:1 on sale = $1.03. Source: https://docs.mistral.ai/inference/pricing (checked 18 Oct 2026)
- DeepSeek
- Official: V4 Pro peak cache-miss $1.32 / $3.96 per 1M; Flash $0.30 / $1.20. Off-peak is half. Cache-hit input is much cheaper — tooltip only. Blended 3:1 peak miss = $1.98. Source: https://api-docs.deepseek.com/quick_start/pricing/ (checked 18 Oct 2026)
- Groq
- Official on-demand: Llama 3.3 70B $0.59 / $0.79 per 1M. Llama 3.1 8B Instant $0.05 / $0.08. GPT OSS 120B $0.15 / $0.60. Blended 3:1 on 70B = $0.64. Source: https://home.cloud.groq.io/pricing (checked 18 Oct 2026)
- Together AI
- Official serverless: GLM-5.3 $1.40 / $4.40 per 1M. GLM-5.3 Flash $0.15 / $0.50. Cached input cheaper; Batch up to 50% off. Blended 3:1 = $2.15. Source: https://www.together.ai/pricing (checked 18 Oct 2026)
- Fireworks AI
- Official serverless: GLM 5.3 $1.40 / $4.40 per 1M. GLM 5.3 Flash $0.15 / $0.50. Cached input cheaper; Batch 50% off. Blended 3:1 = $2.15. Source: https://docs.fireworks.ai/serverless/pricing (checked 18 Oct 2026)
- OpenRouter
- Official: OpenRouter passes through the selected provider's price. Account credits add 5.5%. No first-party per-model table is treated as a c17 headline — cells are em dashes. Source: https://openrouter.ai/pricing (checked 18 Oct 2026)
FAQ
- What is the cheapest hosted LLM API?
- On published cheap/fast models, Groq Llama 3.1 8B Instant is $0.05 / $0.08 per 1M. Haiku 5.5 is $0.10 / $0.50. Among flagships, Groq Llama 3.3 70B is $0.59 / $0.79 and Mistral Large 4 is $0.68 / $2.09 on the listed sale. Frontier APIs start around $2 / $6–$14.
- What does blended 3:1 mean?
- Three input tokens for every output token. The column is (3 × input + 1 × output) ÷ 4, in USD per 1 million tokens. The worked examples use the same ratio at 2M / 20M / 200M input tokens.
- Why are cached and batch prices in the tooltip?
- Those rates only apply when you hit cache or submit Batch jobs. The headline is the uncached list rate you pay on a normal request. Hover or tap the cell for cache, Batch, off-peak, and sale-regular figures.
- Why is OpenRouter an em dash?
- OpenRouter passes through each provider's price and adds 5.5% on credits. It is not a first-party model host, so there is no official OpenRouter $/1M to put in the cell.
- Is Groq comparable to GPT-5.2 or Sonnet?
- No. Groq hosts fast open weights (Llama 8B/70B, GPT OSS). It wins on published $/1M and latency. It is not a frontier-model substitute. The try-first picks use Groq for the hobby bot and Mistral Large 4 for production flagship volume.
- What would a real project cost?
- Estimates from published uncached rates: a hobby bot at 2M+667k cheap tokens is about $0.15 on Groq 8B, $0.53 on Haiku, $0.70 on Mistral Small. A production feature at 20M+6.67M flagship tokens is about $28 on Mistral Large 4 (sale), $53 on DeepSeek Pro, $107 on Sonnet or Gemini 4 Argon, $128 on GPT-5.2. Ten times that volume is ten times those bills.
Related pages
Stack of the Week
Get one stack a week that replaces a SaaS bill.