29 models · 52 providers · snapshot October 6, 2026
LLM API pricing, model by model
Historical comparison data and current dated vendor facts are separate. The task bars keep their original prices and token mixes.
Claude Sonnet 5.5 and GPT-6.1 Sol both list $2.00 input and $10.00 output. GPT-6.1 Sol lists a $0.10 cache read. Claude Sonnet 5.5 lists $0.20. Prices are USD per million tokens.
15 of 22 models with two or more providers list one price at every provider. The widest gap is DeepSeek V4 Flash 0423: 12.6x between its most expensive provider (Cloudflare) and its cheapest (StreamLake). That is a calculation on blended prices.
OpenRouter’s public API reported the provider prices on October 6, 2026. They are third-party-reported. Crewshal did not measure them.
Each dot is one provider. The diamond is the maker’s own list price. Blended price is (3 × input + output) ÷ 4, a calculation.
Price per million tokens, every model
Pick a sort. Select a provider to pin it on every row. The Table view lists every number.
Sorted by blended price, lowest first. Blended price is (3 × input + output) ÷ 4, a calculation. It is not a quality rank.
29 models · Select a provider to pin it on every row.
- DeepSeek V4 Flash 0423Spread 12.6x: highest ÷ lowest Blended 3:1 price (calculation)
- gpt-oss-120bSpread 6.9x: highest ÷ lowest Blended 3:1 price (calculation)
- Llama 3.3 70B InstructSpread 6.7x: highest ÷ lowest Blended 3:1 price (calculation)
- GPT-6 Lunasame
- Qwen3.8 Flash1 listing
- DeepSeek V4 Pro 0423Spread 11.2x: highest ÷ lowest Blended 3:1 price (calculation)
- Llama 4 MaverickSpread 1.7x: highest ÷ lowest Blended 3:1 price (calculation)
- GLM 5.3Spread 4.7x: highest ÷ lowest Blended 3:1 price (calculation)
- Mistral Large 3 25121 listing
- Gemini 3.5 Flash Litesame
- Gemini 3.8 Flashsame
- Gemini 3.x Flash1 listing
- Claude Haiku 4.5same
- Grok 4.71 listing
- Mistral Medium 3.51 listing
- Qwen3.8 Max (0902)1 listing
- Gemini 3.5 Flashsame
- Kimi K3Spread 1.8x: highest ÷ lowest Blended 3:1 price (calculation)
- Claude Sonnet 5same
- Claude Sonnet 5.5same
- GPT-6 Solsame
- GPT-6.1 Sol1 listing
- Gemini 3.1 Pro Previewsame
- Claude Opus 5.5same
- Claude Opus 4.8same
- Claude Opus 5same
- GPT-5.5same
- Claude Fable 5.1same
- GPT-6 Astrasame
- one provider (blended: a calculation, hollow)
- first-party list price
- Spread: highest ÷ lowest price of the row (calculation)
| Maker | Price from | Cache read $/M | Providers | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| TypeSafe | ◇ TypeSafe list price, 2026-09-23 | $0.042 | $0.00 | $0.0042 | $0.0315 | 0 | No comparison | Not reported | Routing decisions only | Routing decisions only | |
| DeepSeek | Reported: StreamLake | $0.042 | $0.084 | $0.0084 | $0.0525 | 15 | 12.6x | 1.05M | $0.056 | $0.00005 | |
| OpenAI | Reported: CoreWeave | $0.03 | $0.17 | $0.03 | $0.065 | 20 | 6.9x | 131k | $0.16 | $0.000081 | |
| Meta | Reported: DeepInfra | $0.10 | $0.32 | — | $0.155 | 10 | 6.7x | 131k | $0.51 | $0.00024 | |
| OpenAI | ◇ OpenAI list price, 2026-10-03 | $0.10 | $0.50 | $0.01 | $0.20 | 2 | 1x | 1.05M | $0.1 | $0.00014 | |
| Alibaba | Reported: Alibaba | $0.15 | $0.47 | $0.016 | $0.23 | 1 | No comparison | 1M | $0.14 | $0.00018 | |
| DeepSeek | Reported: StreamLake | $0.2088 | $0.4176 | $0.0174 | $0.261 | 15 | 11.2x | 1.05M | $0.16 | $0.00021 | |
| Meta | Reported: DigitalOcean | $0.1875 | $0.6525 | — | $0.3037 | 3 | 1.7x | 1.05M | $0.96 | $0.00046 | |
| Z.ai | Reported: Novita | $0.42 | $1.32 | $0.078 | $0.645 | 32 | 4.7x | 1.05M | $0.56 | $0.00054 | |
| Mistral AI | Reported: Mistral | $0.50 | $1.50 | $0.05 | $0.75 | 1 | No comparison | 262k | $0.46 | $0.00057 | |
| Reported: Google AI Studio | $0.30 | $2.50 | $0.03 | $0.85 | 2 | 1x | 1.05M | $0.36 | $0.00052 | ||
| ◇ Google list price, 2026-09-21 | $0.75 | $3.75 | $0.075 | $1.50 | 2 | 1x | 1.05M | $0.77 | $0.001 | ||
| ◇ Google list price, 2026-09-21 | $0.75 | $3.75 | $0.075 | $1.50 | 0 | No comparison | Not reported | $0.77 | $0.001 | ||
| Anthropic | ◇ Anthropic list price, 2026-09-21 | $1.00 | $5.00 | $0.10 | $2.00 | 4 | 1x | 200k | $1.32 | $0.0014 | |
| xAI | Reported: xAI | $2.00 | $6.00 | $0.50 | $3.00 | 1 | No comparison | 500k | $3.23 | $0.0027 | |
| Mistral AI | Reported: Mistral | $1.50 | $7.50 | — | $3.00 | 1 | No comparison | 262k | $7.80 | $0.0039 | |
| Alibaba | Reported: Alibaba | $2.00 | $6.00 | $0.25 | $3.00 | 1 | No comparison | 1M | $2.07 | $0.0024 | |
| ◇ Google list price, 2026-09-21 | $1.50 | $9.00 | $0.15 | $3.38 | 2 | 1x | 1.05M | $1.62 | $0.0022 | ||
| Moonshot AI | Reported: Relace | $0.83 | $13.00 | $0.45 | $3.87 | 19 | 1.8x | 1.05M | $3.03 | $0.0026 | |
| Anthropic | ◇ Anthropic list price, 2026-09-21 | $2.00 | $10.00 | $0.20 | $4.00 | 5 | 1x | 1M | $2.64 | $0.0027 | |
| Anthropic | ◇ Anthropic list price, 2026-09-21 | $2.00 | $10.00 | $0.20 | $4.00 | 5 | 1x | 1M | $2.64 | $0.0027 | |
| OpenAI | Reported: Azure | $2.00 | $10.00 | $0.20 | $4.00 | 2 | 1x | 1.05M | $2.05 | $0.0027 | |
| OpenAI | ◇ OpenAI list price, 2026-10-03 | $2.00 | $10.00 | $0.10 | $4.00 | 0 | No comparison | Not reported | $1.59 | $0.0026 | |
| Reported: Google AI Studio | $2.00 | $12.00 | $0.20 | $4.50 | 2 | 1x | 1.05M | $2.16 | $0.0029 | ||
| Anthropic | ◇ Anthropic list price, 2026-09-21 | $4.00 | $20.00 | Historical record under re-check | $8.00 | 5 | 1x | 1M | $4.36 | $0.0052 | |
| Anthropic | ◇ Anthropic list price, 2026-09-21 | $5.00 | $25.00 | $0.50 | $10.00 | 5 | 1x | 1M | $6.61 | $0.0068 | |
| Anthropic | ◇ Anthropic list price, 2026-09-21 | $5.00 | $25.00 | $0.50 | $10.00 | 5 | 1x | 1M | $6.61 | $0.0068 | |
| OpenAI | Reported: Azure | $5.00 | $30.00 | $0.50 | $11.25 | 2 | 1x | 1.05M | $5.40 | $0.0073 | |
| Anthropic | ◇ Anthropic list price, 2026-09-21 | $10.00 | $50.00 | $0.25 | $20.00 | 4 | 1x | 1M | $9.74 | $0.013 | |
| OpenAI | Reported: Azure | $10.00 | $50.00 | $1.00 | $20.00 | 2 | 1x | 1.05M | $10.26 | $0.014 |
List-price calculation, not a run. 170 models, 4 series: Blended 3:1, Input, Output, Cache read. Blended 3:1: highest Claude Fable 5.1 · Amazon Bedrock $20.00 (n 4). Lowest DeepSeek V4 Flash 0423 · StreamLake (fp8) $0.053 (n 15). Input: highest Claude Fable 5.1 · Amazon Bedrock $10.00 (n 4). Lowest DeepSeek V4 Flash 0423 · Relace (fp4) $0.012 (n 15).
Notesn 1–32 per row
29 models, USD per million tokens, standard tier, snapshot 2026-10-06. Blended 3:1 is a calculation: (3 × input + output) ÷ 4.
Provider prices are reported by OpenRouter’s public API (third-party-reported, not measured by Crewshal). One dot per provider: its cheapest standard-tier endpoint. The diamond is the maker’s own list price, as recorded on the date of its price list. Spread: highest ÷ lowest price of the row, a calculation. A parenthesis names the quantization the provider reported. n = providers with a standard-tier endpoint.
Sources: OpenRouter public API: models and provider endpoints (snapshot), OpenAI list prices, Google Gemini list prices, Anthropic list prices (Claude models)
Claude, GPT and Gemini list prices
The makers’ own list prices, with the day each was recorded. A model with no maker price shows the provider with the lowest blended price (reported 2026-10-09).
Current dated vendor list prices
Official vendor facts have their own observation dates. They do not reprice the historical task bars below.
- Claude Haiku 4.5: $1.00 input, $5.00 output, $0.10 cache read per million tokens. Official pricing, recorded 2026-10-08.
- Claude Sonnet 5.5: $2.00 input, $10.00 output, $0.10 cache read per million tokens. Official pricing, recorded 2026-10-08.
- Claude Opus 5.5: $4.00 input, $20.00 output, $0.20 cache read per million tokens. Official pricing, recorded 2026-10-08.
- Claude Opus 5: $5.00 input, $25.00 output, $0.50 cache read per million tokens. Official pricing, recorded 2026-10-08.
- Claude Fable 5.1: $10.00 input, $50.00 output, $0.25 cache read per million tokens. Official pricing, recorded 2026-10-08.
Historical comparison list prices
The preserved historical Opus cache-read caveat belongs to the old record. The current vendor list above records its separate official value.
Claude (Anthropic)
Anthropic list prices.
- Claude Haiku 4.5: $1.00 input, $5.00 output, $0.10 cache read per million tokens (Anthropic list price, 2026-09-21). Provider directory Benchmarks for Claude Haiku 4.5
- Claude Sonnet 5: $2.00 input, $10.00 output, $0.20 cache read per million tokens (Anthropic list price, 2026-09-21; cache read as reported by Anthropic). Provider directory
- Claude Sonnet 5.5: $2.00 input, $10.00 output, $0.20 cache read per million tokens (Anthropic list price, 2026-09-21). Provider directory Benchmarks for Claude Sonnet 5.5
- Claude Opus 5.5: $4.00 input, $20.00 output per million tokens, cache read under re-check (Anthropic list price, 2026-09-21). Provider directory Benchmarks for Claude Opus 5.5
- Claude Opus 4.8: $5.00 input, $25.00 output, $0.50 cache read per million tokens (Anthropic list price, 2026-09-21; cache read as reported by Anthropic). Provider directory
- Claude Opus 5: $5.00 input, $25.00 output, $0.50 cache read per million tokens (Anthropic list price, 2026-09-21). Provider directory Benchmarks for Claude Opus 5
- Claude Fable 5.1: $10.00 input, $50.00 output, $0.25 cache read per million tokens (Anthropic list price, 2026-09-21). Provider directory Benchmarks for Claude Fable 5.1
GPT (OpenAI)
OpenAI list prices and provider prices. GPT-6 Sol is the snapshot name. GPT-6.1 Sol is the price table name. Crewshal does not merge the two.
- gpt-oss-120b: $0.03 input, $0.17 output, $0.03 cache read per million tokens, at CoreWeave, the lowest blended price of 20 providers (reported 2026-10-06). Provider directory
- GPT-6 Luna: $0.10 input, $0.50 output, $0.01 cache read per million tokens (OpenAI list price, 2026-10-03; cache read as reported by OpenAI). Provider directory
- GPT-6 Sol: $2.00 input, $10.00 output, $0.20 cache read per million tokens, where 2 providers tie for the lowest blended price (reported 2026-10-06). Provider directory
- GPT-6.1 Sol: $2.00 input, $10.00 output, $0.10 cache read per million tokens (OpenAI list price, 2026-10-03). Provider directory
- GPT-5.5: $5.00 input, $30.00 output, $0.50 cache read per million tokens, where 2 providers tie for the lowest blended price (reported 2026-10-06). Provider directory
- GPT-6 Astra: $10.00 input, $50.00 output, $1.00 cache read per million tokens, where 2 providers tie for the lowest blended price (reported 2026-10-06). Provider directory
Gemini (Google)
Google list prices and provider prices. Gemini 3.8 Flash is the snapshot name. Gemini 3.x Flash is the price table name. Crewshal does not merge the two.
- Gemini 3.5 Flash Lite: $0.30 input, $2.50 output, $0.03 cache read per million tokens, where 2 providers tie for the lowest blended price (reported 2026-10-06). Provider directory
- Gemini 3.8 Flash: $0.75 input, $3.75 output, $0.075 cache read per million tokens (Google list price, 2026-09-21; cache read as reported by Google AI Studio). Provider directory
- Gemini 3.x Flash: $0.75 input, $3.75 output, $0.075 cache read per million tokens (Google list price, 2026-09-21). Provider directory
- Gemini 3.5 Flash: $1.50 input, $9.00 output, $0.15 cache read per million tokens (Google list price, 2026-09-21; cache read as reported by Google AI Studio). Provider directory
- Gemini 3.1 Pro Preview: $2.00 input, $12.00 output, $0.20 cache read per million tokens, where 2 providers tie for the lowest blended price (reported 2026-10-06). Provider directory
Other makers
Providers report these prices. Jev 1.13 is a router: it prices routing decisions, not coding tasks.
- Jev 1.13 (router): $0.042 input, output free per million tokens (TypeSafe list price, 2026-09-23). A router prices decisions, not coding tasks. Provider directory
- DeepSeek V4 Flash 0423: $0.042 input, $0.084 output, $0.0084 cache read per million tokens, at StreamLake, the lowest blended price of 15 providers (reported 2026-10-06). Provider directory
- Llama 3.3 70B Instruct: $0.10 input, $0.32 output per million tokens, cache read not reported, at DeepInfra, the lowest blended price of 10 providers (reported 2026-10-06). Provider directory
- Qwen3.8 Flash: $0.15 input, $0.47 output, $0.016 cache read per million tokens, at Alibaba, its only standard-tier provider (reported 2026-10-06). Provider directory
- DeepSeek V4 Pro 0423: $0.2088 input, $0.4176 output, $0.0174 cache read per million tokens, at StreamLake, the lowest blended price of 15 providers (reported 2026-10-06). Provider directory
- Llama 4 Maverick: $0.1875 input, $0.6525 output per million tokens, cache read not reported, at DigitalOcean, the lowest blended price of 3 providers (reported 2026-10-06). Provider directory
- GLM 5.3: $0.42 input, $1.32 output, $0.078 cache read per million tokens, at Novita, the lowest blended price of 32 providers (reported 2026-10-06). Provider directory
- Mistral Large 3 2512: $0.50 input, $1.50 output, $0.05 cache read per million tokens, at Mistral, its only standard-tier provider (reported 2026-10-06). Provider directory
- Grok 4.7: $2.00 input, $6.00 output, $0.50 cache read per million tokens, at xAI, its only standard-tier provider (reported 2026-10-06). Provider directory
- Mistral Medium 3.5: $1.50 input, $7.50 output per million tokens, cache read not reported, at Mistral, its only standard-tier provider (reported 2026-10-06). Provider directory
- Qwen3.8 Max (0902): $2.00 input, $6.00 output, $0.25 cache read per million tokens, at Alibaba, its only standard-tier provider (reported 2026-10-06). Provider directory
- Kimi K3: $0.83 input, $13.00 output, $0.45 cache read per million tokens, at Relace, the lowest blended price of 19 providers (reported 2026-10-06). Provider directory
One model, every provider
Pick a model. Each strip is one price type. Point at a provider to link its prices on the strips.
All 5 standard-tier providers list the same price: $2.00 input and $10.00 output.
Historical Anthropic list price, 2026-09-21: $2.00 input, $10.00 output. Cache read: $0.20. Price source
Input
Output
Cache read
- one provider (reported price, USD per million tokens, log scale per strip)
- first-party list price
| Provider | Input | Output | Cache read |
|---|---|---|---|
| Amazon Bedrock | $2.00 | $10.00 | $0.2 |
| Anthropic | $2.00 | $10.00 | $0.2 |
| Azure | $2.00 | $10.00 | $0.2 |
| Claude Platform on AWS | $2.00 | $10.00 | $0.2 |
| Google Vertex | $2.00 | $10.00 | $0.2 |
Third-party reported values. 5 providers, 3 series: Input, Output, Cache read. Input: all at $2.00. Output: all at $10.00.
Notes
USD per million tokens, standard tier, 5 providers. Reported by OpenRouter’s public API; the diamond is the maker’s list price.
Prices reported by OpenRouter’s public API (third-party-reported, not measured by Crewshal). Check the endpoint tier, region and quantization before you rely on a price.
Sources: OpenRouter public API: models and provider endpoints (snapshot), Anthropic list prices (Claude models)
- CLI task: $0.0027. Calculation using historical prices and recorded tokens.
- SWE-bench attempt: $2.64. Calculation using historical prices and recorded tokens.
- Amazon Bedrock: lowest blended price in this recorded standard-tier market (calculation).
- Anthropic: lowest blended price in this recorded standard-tier market (calculation).
- Azure: lowest blended price in this recorded standard-tier market (calculation).
- Claude Platform on AWS: lowest blended price in this recorded standard-tier market (calculation).
- Google Vertex: lowest blended price in this recorded standard-tier market (calculation).
What one task costs
The same recorded tokens, priced at each model’s price. Pick a task. This is a calculation, not a bill.
SWE-bench-style coding task. 33 recorded attempts.
Hover or focus a bar for its ratio to Claude Sonnet 5.5 (the highlighted row): a ratio of list-price calculations, not a measurement.
| Model | USD per task (calculation) | n |
|---|---|---|
| DeepSeek V4 Flash 0423 · via StreamLake | $0.056 | 33 |
| GPT-6 Luna · same price at 2 providers | $0.1 | 33 |
| Qwen3.8 Flash · via Alibaba | $0.14 | 33 |
| gpt-oss-120b · via CoreWeave | $0.16 | 33 |
| DeepSeek V4 Pro 0423 · via StreamLake | $0.16 | 33 |
| Gemini 3.5 Flash Lite · same price at 2 providers | $0.36 | 33 |
| Mistral Large 3 2512 · via Mistral | $0.46 | 33 |
| Llama 3.3 70B Instruct · via DeepInfra | $0.51 | 33 |
| GLM 5.3 · via Novita | $0.56 | 33 |
| Gemini 3.8 Flash · same price at 2 providers | $0.77 | 33 |
| Gemini 3.x Flash · list price | $0.77 | 33 |
| Llama 4 Maverick · via DigitalOcean | $0.96 | 33 |
| Claude Haiku 4.5 · list price | $1.32 | 33 |
| GPT-6.1 Sol · list price | $1.59 | 33 |
| Gemini 3.5 Flash · same price at 2 providers | $1.62 | 33 |
| GPT-6 Sol · same price at 2 providers | $2.05 | 33 |
| Qwen3.8 Max (0902) · via Alibaba | $2.07 | 33 |
| Gemini 3.1 Pro Preview · same price at 2 providers | $2.16 | 33 |
| Claude Sonnet 5 · same price at 5 providers | $2.64 | 33 |
| Claude Sonnet 5.5 · list price | $2.64 | 33 |
| Kimi K3 · via Relace | $3.03 | 33 |
| Grok 4.7 · via xAI | $3.23 | 33 |
| Claude Opus 5.5 · list price, cache read under re-check | $4.36 | 33 |
| GPT-5.5 · same price at 2 providers | $5.40 | 33 |
| Claude Opus 4.8 · same price at 5 providers | $6.61 | 33 |
| Claude Opus 5 · list price | $6.61 | 33 |
| Mistral Medium 3.5 · via Mistral | $7.80 | 33 |
| Claude Fable 5.1 · list price | $9.74 | 33 |
| GPT-6 Astra · same price at 2 providers | $10.26 | 33 |
List-price calculation, not a run. 29 models. Highest GPT-6 Astra · same price at 2 providers $10.26 (n 33). Lowest DeepSeek V4 Flash 0423 · via StreamLake $0.056 (n 33).
Notesn = 33 per row
One task is 239 uncached input, 295,379 cache write, 4,639,572 cache read, 53,353 output (tokens, recorded). Calculation: recorded tokens × each model’s price per million tokens.
A calculation, not a run and not a bill. Recorded totals over 33 SWE-bench Verified attempts, divided by 33. The AI cost calculator shows how the token mix is derived. Models use different token counts for the same job, so this compares prices, not the cost of getting the job done. "List price" bars use the maker’s list price (2026-09-21, 2026-10-03). Other bars use the provider with the lowest blended price in OpenRouter’s snapshot (reported, 2026-10-06). Cache writes cost 2x the input price for Claude models and the input price for others; a missing cache-read price is priced as input. The cache-read price of Claude Opus 5.5 is under re-check; the bar uses the number in our price table.
Sources: Repricing calculation, Crewshal on SWE-bench Verified, campaign 1 (25 instances), Crewshal on SWE-bench Verified, campaign 2 (8 compiled-extension instances), OpenRouter public API: models and provider endpoints (snapshot), OpenAI list prices, Google Gemini list prices, Anthropic list prices (Claude models)
- Models use different numbers of tokens for the same job. The bars compare prices, not the cost of getting the job done.
- A bar that says “list price” uses the maker’s list price. A bar that says “via” a provider uses the lowest blended price in the snapshot, as reported.
- Cache writes cost twice the input price for Claude models and the input price for other models. A missing cache-read price is priced as input.
What the snapshot shows
29 models, 52 providers, 265 endpoints. Reported 2026-10-09.
265
Provider endpoints in the snapshot
endpoints, 52 providers, 27 models · n = 27
12.6x
Largest standard-tier price spread (DeepSeek V4 Flash 0423)
(most expensive: Cloudflare; cheapest: StreamLake (fp8)) · n = 15
10 of 10
Models where OpenRouter’s per-token price equals the first-party list price
n = 10
5.5%
Credit-purchase fee on OpenRouter’s Standard plan (third-party-reported)
($0.80 minimum by card)
Why prices differ
- The snapshot lists 265 endpoints. 175 are standard tier. The rest are 54 regional, 13 fast, 12 flex, 10 priority and 1 ultrafast. This page prices the standard tier only.
- Some endpoints run a model at lower precision (fp4 or fp8) or with a shorter context. Check the tag before you compare two prices.
- A cache read costs less than a plain input token. The cache-read price differs by model: see the Cache read strip.
- OpenRouter charges the maker’s per-token price on 10 of 10 models with a maker list price. It adds a 5.5% fee when you buy credits.
Price changes
2 dated snapshots: 2026-10-06 to 2026-10-09. Third-party-reported observations, not vendor change timestamps.
53 known changes across 255 compared endpoint keys. 11 ambiguous or changed-lineage keys omitted. Partial known-change history; not complete.
Raw observations: 2026-10-06 · 2026-10-09 · Known-change extract · Latest observation pointer
Known changes and source coverage
- 2026-10-09: Claude Sonnet 5.5 at Amazon Bedrock: cache read $0.20 to $0.10 Calculation: -50.0% rate change
- 2026-10-09: Claude Sonnet 5.5 at Anthropic: cache read $0.20 to $0.10 Calculation: -50.0% rate change
- 2026-10-09: Claude Sonnet 5.5 at Azure: cache read $0.20 to $0.10 Calculation: -50.0% rate change
- 2026-10-09: Claude Sonnet 5.5 at Azure (regional): cache read $0.22 to $0.11 Calculation: -50.0% rate change
- 2026-10-09: Claude Sonnet 5.5 at Claude Platform on AWS: cache read $0.20 to $0.10 Calculation: -50.0% rate change
- 2026-10-09: Claude Sonnet 5.5 at Google Vertex (regional): cache read $0.22 to $0.11 Calculation: -50.0% rate change
- 2026-10-09: Claude Sonnet 5.5 at Google Vertex: cache read $0.20 to $0.10 Calculation: -50.0% rate change
- 2026-10-09: Claude Sonnet 5.5 at Google Vertex (regional): cache read $0.22 to $0.11 Calculation: -50.0% rate change
- 2026-10-09: DeepSeek V4 Flash 0423 at OpenInference: input $0.0132 to $0.0046 Calculation: -65.5% rate change
- 2026-10-09: DeepSeek V4 Flash 0423 at OpenInference: output $1.41 to $0.2803 Calculation: -80.1% rate change
- 2026-10-09: DeepSeek V4 Flash 0423 at OpenInference: cache read $0.0132 to $0.0046 Calculation: -65.5% rate change
- 2026-10-09: DeepSeek V4 Flash 0423 at Relace: input $0.012 to $0.03 Calculation: 150.0% rate change
- 2026-10-09: DeepSeek V4 Flash 0423 at Relace: cache read $0.012 to $0.03 Calculation: 150.0% rate change
- 2026-10-09: DeepSeek V4 Flash 0423 at Wafer: added
- 2026-10-09: DeepSeek V4 Pro 0423 at NextBit: removed
- 2026-10-09: DeepSeek V4 Pro 0423 at Relace: input $0.2067 to $0.30 Calculation: 45.1% rate change
- 2026-10-09: DeepSeek V4 Pro 0423 at Relace: cache read $0.21 to $0.23 Calculation: 9.5% rate change
- 2026-10-09: Llama 4 Maverick at DigitalOcean: cache read — to $0.05
- 2026-10-09: Kimi K3 at AkashML: input $1.30 to $1.20 Calculation: -7.7% rate change
- 2026-10-09: Kimi K3 at AkashML: cache read $1.30 to $1.20 Calculation: -7.7% rate change
- 2026-10-09: Kimi K3 at InferenceNet (fast): input $3.50 to $2.99 Calculation: -14.6% rate change
- 2026-10-09: Kimi K3 at InferenceNet: input $0.95 to $0.58 Calculation: -38.9% rate change
- 2026-10-09: Kimi K3 at InferenceNet: output $14.00 to $12.00 Calculation: -14.3% rate change
- 2026-10-09: Kimi K3 at InferenceNet: cache read $0.31 to $0.30 Calculation: -3.2% rate change
- 2026-10-09: Kimi K3 at Makora: removed
- 2026-10-09: Kimi K3 at Makora: added
- 2026-10-09: Kimi K3 at Morph: input $1.27 to $2.50 Calculation: 96.2% rate change
- 2026-10-09: Kimi K3 at Morph: output $13.30 to $14.00 Calculation: 5.3% rate change
- 2026-10-09: Kimi K3 at Morph: cache read $0.278 to $0.29 Calculation: 4.3% rate change
- 2026-10-09: Kimi K3 at Parasail (fast): added
- 2026-10-09: Kimi K3 at Parasail: input $3.00 to $2.60 Calculation: -13.3% rate change
- 2026-10-09: Kimi K3 at Parasail: output $15.00 to $13.00 Calculation: -13.3% rate change
- 2026-10-09: Kimi K3 at Parasail: cache read $0.30 to $0.26 Calculation: -13.3% rate change
- 2026-10-09: Kimi K3 at Relace: input $0.83 to $0.57 Calculation: -31.3% rate change
- 2026-10-09: Kimi K3 at Relace: cache read $0.45 to $0.30 Calculation: -33.3% rate change
- 2026-10-09: Kimi K3 at Wafer: input $0.95 to $0.79 Calculation: -16.8% rate change
- 2026-10-09: Kimi K3 at Wafer: output $14.00 to $15.00 Calculation: 7.1% rate change
- 2026-10-09: Kimi K3 at Wafer: cache read $0.40 to $0.45 Calculation: 12.5% rate change
- 2026-10-09: gpt-oss-120b at Mancer 2: input $0.045 to $0.05 Calculation: 11.1% rate change
- 2026-10-09: GLM 5.3 at InferenceNet: removed
- 2026-10-09: GLM 5.3 at InferenceNet: added
- 2026-10-09: GLM 5.3 at Io Net: added
- 2026-10-09: GLM 5.3 at Morph: input $0.179 to $0.145 Calculation: -19.0% rate change
- 2026-10-09: GLM 5.3 at Morph: output $3.55 to $3.74 Calculation: 5.3% rate change
- 2026-10-09: GLM 5.3 at Morph: cache read $0.137 to $0.108 Calculation: -21.2% rate change
- 2026-10-09: GLM 5.3 at Relace: input $0.03 to $0.0312 Calculation: 4.0% rate change
- 2026-10-09: GLM 5.3 at Relace: cache read $0.03 to $0.029 Calculation: -3.3% rate change
- 2026-10-09: GLM 5.3 at Wafer: input $0.15 to $0.17 Calculation: 13.3% rate change
- 2026-10-09: GLM 5.3 at Wafer: output $7.00 to $3.39 Calculation: -51.6% rate change
- 2026-10-09: GLM 5.3 at Wafer: cache read $0.14 to $0.16 Calculation: 14.3% rate change
- 2026-10-09: GLM 5.3 at Wafer (regional): input $0.07 to $0.039 Calculation: -44.3% rate change
- 2026-10-09: GLM 5.3 at Wafer (regional): output $7.00 to $3.39 Calculation: -51.6% rate change
- 2026-10-09: GLM 5.3 at Wafer (regional): cache read $0.065 to $0.038 Calculation: -41.5% rate change
2026-10-06 to 2026-10-09: 255 of 266 keys compared
- deepseek/deepseek-v4-flash · streamlake/fp8: changed identity lineage. Source indices before: 19:2; after: 19:3.
- deepseek/deepseek-v4-pro · reka: changed identity lineage. Source indices before: 18:3; after: 18:5.
- deepseek/deepseek-v4-pro · streamlake/fp8: changed identity lineage. Source indices before: 18:1; after: 18:2.
- moonshotai/kimi-k3 · decart/mxfp4: changed identity lineage. Source indices before: 25:8; after: 25:6.
- moonshotai/kimi-k3 · phala: changed identity lineage. Source indices before: 25:7; after: 25:8.
- z-ai/glm-5.3 · decart/fp4: changed identity lineage. Source indices before: 26:18; after: 26:14.
- z-ai/glm-5.3 · makora/fp4: changed identity lineage. Source indices before: 26:6; after: 26:4.
- z-ai/glm-5.3 · mistral/nvfp4: ambiguous identity. Source indices before: 26:23, 26:36; after: 26:24, 26:37.
- z-ai/glm-5.3 · novita/fp8: changed identity lineage. Source indices before: 26:10; after: 26:12.
- z-ai/glm-5.3 · reka: changed identity lineage. Source indices before: 26:4; after: 26:3.
- z-ai/glm-5.3 · siliconflow/fp8: changed identity lineage. Source indices before: 26:13; after: 26:17.
Questions about LLM API pricing
These answers describe the historical comparison. Current official values have separate dates above.
How much does the Claude API cost per million tokens?
Claude Haiku 4.5 lists $1.00 input and $5.00 output. Claude Sonnet 5 and Claude Sonnet 5.5 list $2.00 input and $10.00 output. Claude Opus 5.5 lists $4.00 input and $20.00 output. Claude Opus 4.8 and Claude Opus 5 list $5.00 input and $25.00 output. Claude Fable 5.1 lists $10.00 input and $50.00 output. Prices are USD per million tokens. They are Anthropic list prices, recorded 2026-09-21.
Is the Claude API cheaper than the GPT API?
Not by list price alone. Claude Sonnet 5.5 and GPT-6.1 Sol both list $2.00 input and $10.00 output. GPT-6.1 Sol lists a $0.10 cache read. Claude Sonnet 5.5 lists $0.20. A price per token is not a price per task. Models use different numbers of tokens and fail at different rates. See the cost per task section and the benchmarks.
Which LLM API is the cheapest?
In this snapshot the lowest blended price is DeepSeek V4 Flash 0423 at StreamLake (fp8): $0.0525 per million tokens (calculation). A low price can come with lower precision, a shorter context or lower quality. Price says nothing about quality.
Why does one model cost different amounts at different providers?
Often it does not. 15 of 22 models list one price at every standard-tier provider. The other 7 differ, from 1.7x (Llama 4 Maverick) to 12.6x (DeepSeek V4 Flash 0423), a calculation. A provider can run a model at lower precision, such as fp8, or with a shorter context. The endpoint tags show both.
Does OpenRouter cost more than going direct?
On 10 of 10 models with a maker list price, OpenRouter’s per-token price equals the maker’s price. OpenRouter adds a 5.5% fee when you buy credits (third-party-reported, Standard plan, card).
How often do these prices change?
2 dated snapshots, 2026-10-06 to 2026-10-09. 53 known changes across 255 compared endpoint keys. 11 ambiguous or changed-lineage keys omitted. Partial known-change history; not complete. These are observation dates, not vendor change times. Check the maker’s price page before you buy.
What does one task cost, not one token?
The recorded tokens of one CLI task cost $0.0027 at the list price of Claude Sonnet 5.5 (15 calls). One recorded SWE-bench attempt costs $2.64 (33 attempts). Both are calculations. Other models use other token counts, so the section compares prices, not task cost.
How these numbers were made
- One snapshot of OpenRouter’s public, keyless API on October 6, 2026: 265 endpoints, 52 providers, 27 models. Each price is copied as the API reported it.
- A provider dot is its cheapest standard-tier endpoint. The model table lists headline comparisons; the linked historical study holds every endpoint.
- The diamond is the maker’s list price from the price list named in the table, with its own date.
- Blended price = (3 × input + output) ÷ 4. Spread = the highest provider blended price ÷ the lowest. Cost per task = recorded tokens × price. All three are calculations.
- Sorted lists name their key. A tie is not a rank. This page computes no overall score.
Historical record caveats
- The comparison ladder keeps its original snapshot. The separate change history names its own dated observations.
- A provider’s own site can show a different price from the one OpenRouter reports.
- Maker list prices were recorded on 2026-09-21, 2026-09-23, 2026-10-03, before the snapshot. A maker change in between shows as a gap.
- The data holds no latency, throughput or quality for these prices. A lower price is not shown to be better or worse.
- The old record marked Opus cache read under re-check. Its separate official value was recorded on 2026-10-08.
- Google Gemini list prices. The vendor announced a doubling from 2027-01-01.
- Anthropic list prices (Claude models), 2026-09-21
- Google Gemini list prices, 2026-09-21
- OpenAI list prices, 2026-10-03
- Jev 1.13 list price, 2026-09-23
- OpenRouter public API: models and provider endpoints (snapshot), 2026-10-06