Self-hosting an LLM vs paying for an API: break-even calculator

OpenAI: gpt-5.4-mini

From the 'All models' Standard table (rows embedded in page data, collapsed by default); row has 3 values mapped as input / cached input / output (no cache-write price). Most recent model actually named '-mini' on the page. Batch/Flex $0.375 / $0.0375 / $2.25; Fast $1.50 / $0.15 / $9.00. No long-context tier shown. Standard tier. Batch and Flex are 50% of Standard. Output price includes reasoning tokens. Regional processing (data residency) and FedRAMP endpoints +10% for models released on/after 2026-03-05. Requested URL platform.openai.com/docs/pricing redirected to developers.openai.com/api/docs/pricing; openai.com/api/pricing/ returned 403.

OpenAI: gpt-5.4-nano

From the 'All models' Standard table (collapsed by default); 3 values mapped as input / cached input / output. Most recent model actually named '-nano' on the page. Batch/Flex $0.10 / $0.01 / $0.625. No long-context tier shown. Standard tier. Batch and Flex are 50% of Standard. Output price includes reasoning tokens. Regional processing (data residency) and FedRAMP endpoints +10% for models released on/after 2026-03-05. Requested URL platform.openai.com/docs/pricing redirected to developers.openai.com/api/docs/pricing; openai.com/api/pricing/ returned 403.

OpenAI: gpt-6-astra

Listed first under 'Flagship models'. Short context (<=272K input tokens) shown; long context (>272K input): $20 in / $2 cached / $75 out. Cache writes $12.50 (short) / $25 (long); per page, input tokens are billed as input, cached input, or cache write (not additive). Batch/Flex $5 / $0.50 / $25. Fast $20 / $2 / $100; Ultrafast $60 / $6 / $300. Standard tier. Batch and Flex are 50% of Standard. Output price includes reasoning tokens. Regional processing (data residency) and FedRAMP endpoints +10% for models released on/after 2026-03-05. Requested URL platform.openai.com/docs/pricing redirected to developers.openai.com/api/docs/pricing; openai.com/api/pricing/ returned 403.

OpenAI: gpt-6-luna

Flagship-table small tier (current generation uses astra/sol/luna naming rather than mini/nano). Short context (<=272K) shown; long context (>272K): $0.20 in / $0.02 cached / $0.75 out. Cache writes $0.125 (short) / $0.25 (long). Batch/Flex $0.05 / $0.005 / $0.25. Fast $0.20 / $0.02 / $1.00. Standard tier. Batch and Flex are 50% of Standard. Output price includes reasoning tokens. Regional processing (data residency) and FedRAMP endpoints +10% for models released on/after 2026-03-05. Requested URL platform.openai.com/docs/pricing redirected to developers.openai.com/api/docs/pricing; openai.com/api/pricing/ returned 403.

OpenAI: gpt-6.1-sol

Flagship-table mid tier. Short context (<=272K) shown; long context (>272K): $4 in / $0.20 cached / $15 out. Cache writes $2.50 (short) / $5 (long). Batch/Flex $1 / $0.05 / $5. Fast $4 / $0.20 / $20. Also listed: gpt-6-sol at $2 / $0.20 cached / $10. Standard tier. Batch and Flex are 50% of Standard. Output price includes reasoning tokens. Regional processing (data residency) and FedRAMP endpoints +10% for models released on/after 2026-03-05. Requested URL platform.openai.com/docs/pricing redirected to developers.openai.com/api/docs/pricing; openai.com/api/pricing/ returned 403.

Anthropic: Claude Fable 5.1

Top row of model table ('For demanding reasoning and long-horizon agentic work'). Cache hit is 0.025x input ($0.25). 5m write $12.50, 1h write $20. Batch $5 / $25. Full 1M context at standard pricing (Claude 4.6+). Page says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text, so compare per-task rather than per-token cost. Batch API 50% off input and output. Prompt caching: 5-min write 1.25x input, 1-hour write 2x input. inference_geo 'us' (data residency) 1.1x on all token categories for Claude 4.6+ models. Requested docs.claude.com URL redirected to platform.claude.com.

Anthropic: Claude Haiku 4.5

Current Haiku. Cache hit $0.10 (0.1x). 5m write $1.25, 1h write $2. Batch $0.50 / $2.50. Uses the previous tokenizer (Claude 4.6 and earlier). Batch API 50% off input and output. Prompt caching: 5-min write 1.25x input, 1-hour write 2x input. inference_geo 'us' (data residency) 1.1x on all token categories for Claude 4.6+ models. Requested docs.claude.com URL redirected to platform.claude.com.

Anthropic: Claude Opus 5.5

Current Opus. Cache hit is 0.05x input ($0.20). 5m write $5, 1h write $8. Batch $2 / $10. Fast mode (research preview) $8 in / $40 out. Full 1M context at standard pricing (Claude 4.6+). Page says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text, so compare per-task rather than per-token cost. Batch API 50% off input and output. Prompt caching: 5-min write 1.25x input, 1-hour write 2x input. inference_geo 'us' (data residency) 1.1x on all token categories for Claude 4.6+ models. Requested docs.claude.com URL redirected to platform.claude.com.

Anthropic: Claude Sonnet 5.5

Current Sonnet. Cache hit $0.20 (0.1x). 5m write $2.50, 1h write $4. Batch $1 / $5. Full 1M context at standard pricing (Claude 4.6+). Page says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text, so compare per-task rather than per-token cost. Batch API 50% off input and output. Prompt caching: 5-min write 1.25x input, 1-hour write 2x input. inference_geo 'us' (data residency) 1.1x on all token categories for Claude 4.6+ models. Requested docs.claude.com URL redirected to platform.claude.com.

Google: Gemini 2.5 Pro (gemini-2.5-pro)

Older generally available Pro model, kept for reference. Prompts <=200k shown; >200k: $2.50 in / $15.00 out / $0.25 cached. Batch $0.625 / $5.00 (<=200k). Paid tier, Standard. Output price includes thinking tokens. Batch and Flex are 50% of Standard. First plain curl was redirected to a Google sign-in check; the page loaded fine with cookies kept (public, no login).

Google: Gemini 3.1 Pro Preview (gemini-3.1-pro-preview)

Only Gemini 3.x Pro text model with its own price table (still 'Preview'). Prompts <=200k tokens shown; prompts >200k: $4.00 in / $18.00 out / $0.40 cached. Context-cache storage $4.50 per 1M tokens per hour. Batch/Flex $1.00 / $6.00 (<=200k), $2.00 / $9.00 (>200k). Priority $3.60 / $21.60 (<=200k). Paid tier, Standard. Output price includes thinking tokens. Batch and Flex are 50% of Standard. First plain curl was redirected to a Google sign-in check; the page loaded fine with cookies kept (public, no login).

Google: Gemini 3.5 Flash (gemini-3.5-flash)

Page calls it 'Our earlier Flash model'; it is not on the promotional pricing the 3.6 to 3.8 Flash models have. Cache storage $1.00 per 1M tokens per hour. Batch/Flex $0.75 / $4.50. Paid tier, Standard. Output price includes thinking tokens. Batch and Flex are 50% of Standard. First plain curl was redirected to a Google sign-in check; the page loaded fine with cookies kept (public, no login).

Google: Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite)

Input price applies to text/image/video/audio. Cache storage $1.00 per 1M tokens per hour. Batch $0.15 / $1.25. Older Gemini 3.1 Flash-Lite is cheaper: $0.25 in (text/image/video; $0.50 audio) / $1.50 out / $0.025 cached. Paid tier, Standard. Output price includes thinking tokens. Batch and Flex are 50% of Standard. First plain curl was redirected to a Google sign-in check; the page loaded fine with cookies kept (public, no login).

Google: Gemini 3.8 Flash (gemini-3.8-flash)

Page calls it 'Our most intelligent Flash model'. PROMOTIONAL: $0.75 in / $3.75 out / $0.075 cached through 2026-12-31; from 2027-01-01 $1.50 / $7.50 / $0.15. Cache storage $0.50 per 1M tokens per hour (rises to $1.00). No >200k tier shown. Batch/Flex $0.375 / $1.875. Priority $1.35 / $6.75. Gemini 3.7 Flash and 3.6 Flash carry the same promotional prices. Paid tier, Standard. Output price includes thinking tokens. Batch and Flex are 50% of Standard. First plain curl was redirected to a Google sign-in check; the page loaded fine with cookies kept (public, no login).

DeepSeek: deepseek-flash (DeepSeek-V4.1-Flash)

Cache-miss input $0.30, cache-hit input $0.006, output $1.20 (peak). Off-peak: $0.15 / $0.003 / $0.60. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by V4.1-Flash at this price. Weights on HF (MIT) per HF API check. PEAK rates recorded (the higher, undiscounted rate). Off-peak is half price: peak hours are 01:00-04:00 and 06:00-10:00 UTC, Mon-Fri, excluding Chinese public holidays; all other hours (most of the week) are off-peak. Context 1M, max output 384K. Thinking mode is on by default.

DeepSeek: deepseek-v4-pro (DeepSeek-V4-Pro-0813)

Cache-miss input $1.32, cache-hit input $0.044, output $3.96 (peak). Off-peak: $0.66 / $0.022 / $1.98. No vision. Weights on HF (MIT) per HF API check. PEAK rates recorded (the higher, undiscounted rate). Off-peak is half price: peak hours are 01:00-04:00 and 06:00-10:00 UTC, Mon-Fri, excluding Chinese public holidays; all other hours (most of the week) are off-peak. Context 1M, max output 384K. Thinking mode is on by default.

Mistral: Mistral Large 3

675B MoE with open weights (Apache-2.0 on HF). The mistral.ai/pricing FAQ example also quotes 'Mistral Large' at $0.5 in / $1.5 out. Standard tier, default (non-regional) view of docs.mistral.ai/inference/pricing. mistral.ai/pricing (the requested URL) shows only plans, no per-model API table. Batch is 50% off per Mistral's pages.

Mistral: Mistral Large 4

SALE PRICE recorded (what is charged today). The page shows the original price struck through: $1.36 in / $0.14 cached / $4.18 out. No sale end date is given. No mistralai Large 4 repo found on Hugging Face, so treated as closed weights. Standard tier, default (non-regional) view of docs.mistral.ai/inference/pricing. mistral.ai/pricing (the requested URL) shows only plans, no per-model API table. Batch is 50% off per Mistral's pages.

Mistral: Mistral Medium 3.5

The pricing page recommends it for most tasks and coding. HF has mistralai/Mistral-Medium-3.5-128B under a non-Apache license ('other', probably Mistral Research License), so check commercial self-hosting terms. Standard tier, default (non-regional) view of docs.mistral.ai/inference/pricing. mistral.ai/pricing (the requested URL) shows only plans, no per-model API table. Batch is 50% off per Mistral's pages.

Mistral: Mistral Small 4

Open weights (Apache-2.0 on HF, 119B). Same page lists Ministral 3 14B $0.2/$0.2, 8B $0.15/$0.15, 3B $0.1/$0.1, and third-party Z.ai GLM 5.3 at $1.4 / $0.14 cached / $4.4. Standard tier, default (non-regional) view of docs.mistral.ai/inference/pricing. mistral.ai/pricing (the requested URL) shows only plans, no per-model API table. Batch is 50% off per Mistral's pages.

Together AI: DeepSeek V4 Flash 0731

Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: DeepSeek V4 Pro 0813

Input and output match DeepSeek's own peak prices; cached is $0.13 here vs $0.044 at DeepSeek. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: DeepSeek V4.1 Flash

Matches DeepSeek's own peak prices. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Gemma 4 31B

Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: GLM-5.3

GLM-5.2 is also listed at the same $1.40 / $0.26 / $4.40. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: GLM-5.3-Flash

Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: gpt-oss-120B

No cached price shown. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Kimi K3

Marked 'PROMO' in the Vision tab of the same table, so this price may be temporary. Kimi K2.x is not listed on Together's page. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Llama 3 8B Instruct Lite

This is Llama 3 (not 3.1) 8B 'Lite'. Together lists no Llama 3.1 8B chat model. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Llama 3.3 70B

Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: MiniMax M2.7

Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: MiniMax M3

Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Qwen3 235B A22B Instruct 2507 FP8 Throughput

FP8, 'Throughput' variant (as named). Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Qwen3.5-397B-A17B

Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Qwen3.5 9B

Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Qwen3.6-Plus

No Qwen/Qwen3.6-Plus repo on HF (only Qwen3.6-27B and 35B-A3B), so treated as Alibaba's closed 'Plus' API model resold by Together. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Qwen3.8-2.4T-A95B

Page URL slug is 'qwen3-8-max', so this is probably what Fireworks and DeepInfra call 'Qwen3.8 Max'. HF license is 'other'. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Together AI: Qwen3.8 Flash

Open-weights status unclear: HF has Qwen/Qwen3.8-Flash-Next but no 'Qwen3.8-Flash' repo. Serverless Chat table. The page has a 'Batch API price' toggle, but batch prices are not in the static HTML (filled in by JavaScript). Quantization not stated unless named.

Fireworks AI: Other base models: 4B to 16B parameters

Would cover Llama 3.1 8B, Qwen3 8B, Qwen3 14B and Gemma 3 12B if served. SIZE-BUCKET RULE: 'For any text or vision model not listed individually, pricing is set by parameter count and architecture.' The same price applies to input and output, with no separate cached-input price. This page does NOT confirm which specific models are on serverless; check Fireworks' model library before mapping a model to this bucket. Batch is 50%.

Fireworks AI: Other base models: more than 16B parameters

Would cover dense Llama 3.3 70B, Llama 3.1 405B, Qwen3 32B, Gemma 3 27B and Mistral Small 24B if served. SIZE-BUCKET RULE: 'For any text or vision model not listed individually, pricing is set by parameter count and architecture.' The same price applies to input and output, with no separate cached-input price. This page does NOT confirm which specific models are on serverless; check Fireworks' model library before mapping a model to this bucket. Batch is 50%.

Fireworks AI: Other base models: less than 4B parameters

Applies to dense models under 4B parameters. SIZE-BUCKET RULE: 'For any text or vision model not listed individually, pricing is set by parameter count and architecture.' The same price applies to input and output, with no separate cached-input price. This page does NOT confirm which specific models are on serverless; check Fireworks' model library before mapping a model to this bucket. Batch is 50%.

Fireworks AI: Other base models: MoE 56.1B to 176B parameters (e.g. DBRX, Mixtral 8x22B)

No bucket is listed for MoE above 176B (e.g. Qwen3 235B-A22B), so those models have no rule-based price. SIZE-BUCKET RULE: 'For any text or vision model not listed individually, pricing is set by parameter count and architecture.' The same price applies to input and output, with no separate cached-input price. This page does NOT confirm which specific models are on serverless; check Fireworks' model library before mapping a model to this bucket. Batch is 50%.

Fireworks AI: Other base models: MoE up to 56B parameters (e.g. Mixtral 8x7B)

Would cover Qwen3 30B-A3B and gpt-oss-20b (about 21B MoE) if served. SIZE-BUCKET RULE: 'For any text or vision model not listed individually, pricing is set by parameter count and architecture.' The same price applies to input and output, with no separate cached-input price. This page does NOT confirm which specific models are on serverless; check Fireworks' model library before mapping a model to this bucket. Batch is 50%.

Fireworks AI: DeepSeek V4.1 Flash

US region $0.45 / $0.009 / $1.80. Standard serverless tier; cells are input / cached input / output. Batch inference is 50% of serverless on input and output. Priority costs about 1.25x (1.5x for some models); '(US)' region variants cost 1.5x. fireworks.ai/pricing has no per-token table and links to this docs page.

Fireworks AI: GLM 5.3

GLM 5.3 Fast and US region both $2.10 / $0.39 / $6.60. Standard serverless tier; cells are input / cached input / output. Batch inference is 50% of serverless on input and output. Priority costs about 1.25x (1.5x for some models); '(US)' region variants cost 1.5x. fireworks.ai/pricing has no per-token table and links to this docs page.

Fireworks AI: GLM 5.3 Flash

Standard serverless tier; cells are input / cached input / output. Batch inference is 50% of serverless on input and output. Priority costs about 1.25x (1.5x for some models); '(US)' region variants cost 1.5x. fireworks.ai/pricing has no per-token table and links to this docs page.

Fireworks AI: OpenAI GPT OSS 120B

Priority $0.18 / $0.018 / $0.72. Standard serverless tier; cells are input / cached input / output. Batch inference is 50% of serverless on input and output. Priority costs about 1.25x (1.5x for some models); '(US)' region variants cost 1.5x. fireworks.ai/pricing has no per-token table and links to this docs page.

Fireworks AI: Kimi K3

Kimi K3 Fast $4.50 / $0.45 / $22.50; US region $4.50 / $0.45 / $22.50. Standard serverless tier; cells are input / cached input / output. Batch inference is 50% of serverless on input and output. Priority costs about 1.25x (1.5x for some models); '(US)' region variants cost 1.5x. fireworks.ai/pricing has no per-token table and links to this docs page.

Fireworks AI: MiniMax M3

Priority $0.45 / $0.09 / $1.80. Standard serverless tier; cells are input / cached input / output. Batch inference is 50% of serverless on input and output. Priority costs about 1.25x (1.5x for some models); '(US)' region variants cost 1.5x. fireworks.ai/pricing has no per-token table and links to this docs page.

Fireworks AI: Qwen 3.8 Max

Same price as Together's 'Qwen3.8-2.4T-A95B' (slug qwen3-8-max), so probably the open Qwen/Qwen3.8-2.4T-A95B, but this page does not say so. Priority $3.00 / $0.375 / $9.00. Standard serverless tier; cells are input / cached input / output. Batch inference is 50% of serverless on input and output. Priority costs about 1.25x (1.5x for some models); '(US)' region variants cost 1.5x. fireworks.ai/pricing has no per-token table and links to this docs page.

DeepInfra: DeepSeek-V3.2

Context 160k. DeepSeek-V3.1 is also listed: $0.25 / $0.13 cached / $0.95. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: DeepSeek-V4-Flash

Context 1024k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: DeepSeek-V4-Flash-0731

Context 1024k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: DeepSeek-V4-Pro

Context 1024k. This is the original V4-Pro checkpoint; DeepSeek's own API now serves V4-Pro-0813. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: gemma-3-27b-it

Context 128k. gemma-3-12b-it is also listed at $0.05 / $0.15. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: gemma-4-26B-A4B-it

Context 256k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: gemma-4-31B-it

Context 256k. Variants also listed: gemma-4-31B-it-turbo $0.09 / $0.05 cached / $0.34, and gemma-4-31B-it-Ultra $0.27 / $0.76 (128k). Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Kimi-K2.6

Context 256k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Kimi-K3

Context 1024k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Meta-Llama-3.1-8B-Instruct-Turbo

Turbo variant; context 128k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Llama-3.3-70B-Instruct-Turbo

Turbo variant; context 128k. Meta-Llama-3.1-70B-Instruct-Turbo is also listed at $0.40 / $0.40. No 405B listed. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Mistral-Small-3.2-24B-Instruct-2506

Context 125k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3-14B

Context 40k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3-235B-A22B-Instruct-2507

Context 256k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3-30B-A3B

Context 40k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3-32B

Context 40k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3.5-27B

Context 256k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3.5-35B-A3B

Context 256k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3.5-397B-A17B

Context 256k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3.5-9B

Context 256k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3.6-27B

Context 256k. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3.6-35B-A3B

Context 256k. The page shows cached input equal to input ($0.10 / $0.10 cached). Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

DeepInfra: Qwen3.8-Max

Context 250k. Probably the same model as Together's Qwen3.8-2.4T-A95B, but there is no Qwen/Qwen3.8-Max repo on HF and DeepInfra also resells closed models (e.g. Qwen3-Max, Gemini, Claude), so open-weights status is unconfirmed. Standard service tier (1x). Priority 1.5x, Flex 0.8x. Hint is taken from DeepInfra's model URL path. Quantization is not stated on the pricing page; DeepInfra 'Turbo' variants are often quantized, so check the model page.

Groq: GPT OSS 120B (openai/gpt-oss-120b)

Production model, about 500 tokens/s. groq.com/pricing now redirects (308) to the homepage, so prices come from GroqDocs 'Supported Models' (console.groq.com/docs/models), the provider's own public page. No cached-input price shown on that page.

Groq: GPT OSS 20B (openai/gpt-oss-20b)

Production model, about 1000 tokens/s. groq.com/pricing now redirects (308) to the homepage, so prices come from GroqDocs 'Supported Models' (console.groq.com/docs/models), the provider's own public page. No cached-input price shown on that page.

Groq: Llama 3.1 8B (llama-3.1-8b-instant)

Price shown as 'Contact Sales' (tagged Enterprise); no public per-token price. groq.com/pricing now redirects (308) to the homepage, so prices come from GroqDocs 'Supported Models' (console.groq.com/docs/models), the provider's own public page. No cached-input price shown on that page.

Groq: Llama 3.3 70B (llama-3.3-70b-versatile)

Price shown as 'Contact Sales' (tagged Enterprise); no public per-token price. groq.com/pricing now redirects (308) to the homepage, so prices come from GroqDocs 'Supported Models' (console.groq.com/docs/models), the provider's own public page. No cached-input price shown on that page.

Groq: MiniMax M2.7 (minimaxai/minimax-m2.7)

Preview, tagged Enterprise; price shown as 'Contact Sales'. groq.com/pricing now redirects (308) to the homepage, so prices come from GroqDocs 'Supported Models' (console.groq.com/docs/models), the provider's own public page. No cached-input price shown on that page.

Groq: Qwen/Qwen3.8-27B (qwen/qwen3.8-27b)

PREVIEW model (evaluation only, may be discontinued at short notice). groq.com/pricing now redirects (308) to the homepage, so prices come from GroqDocs 'Supported Models' (console.groq.com/docs/models), the provider's own public page. No cached-input price shown on that page.

Hyperstack: OpenAI gpt-oss-120b

From the 'Token-based Pricing' table on Hyperstack's GPU pricing page.

Hyperstack: Llama 3.1 8B

From the 'Token-based Pricing' table on Hyperstack's GPU pricing page.

Hyperstack: Llama 3.3 70B

From the 'Token-based Pricing' table on Hyperstack's GPU pricing page.

Crusoe: Gemma 4 31B-it

From the 'Serverless Inference pricing' table on Crusoe's cloud pricing page.

Crusoe: GLM 5.3

From the 'Serverless Inference pricing' table on Crusoe's cloud pricing page.

Crusoe: GPT-OSS 120B

From the 'Serverless Inference pricing' table on Crusoe's cloud pricing page.

Pages we could not use as published, and what we used instead:

• https://openai.com/api/pricing/: HTTP 403 (bot protection). Used https://developers.openai.com/api/docs/pricing instead (platform.openai.com/docs/pricing redirects there).
• https://groq.com/pricing: HTTP 308 redirect to the groq.com homepage, which has no prices. Used https://console.groq.com/docs/models (GroqDocs Supported Models, which lists per-model prices).
• https://mistral.ai/pricing: Lists plans only, with no per-model API price table in the raw HTML or via WebFetch; the only API figure is an FAQ example ('Mistral Large costs $0.5/M in, $1.5/M out'). Used https://docs.mistral.ai/inference/pricing instead.
• https://fireworks.ai/pricing: No per-token serverless price table (only embeddings, training and GPU prices); it links to https://docs.fireworks.ai/serverless/pricing, which was used.