OpenRouter pricing guide 2026: GPT, Claude, Gemini, and DeepSeek API costs and a monthly budget

OpenRouter calls GPT, Claude, Gemini, and DeepSeek with one API key. On the public list for 30 September 2026, input, cache-read, and output prices line up with the vendors. There is no extra per-token markup. The extra line is the platform fee: 5.5% on standard pay-as-you-go, 8% on Business.

The figures below come from that day’s model list and pricing page. Units are US dollars per million tokens, standard endpoints, batch excluded. Swapping models on the same key changes the unit price, not the shape of the bill. If a price moves, the model page wins. The workload matches the Agents API cost note, so the two bills can sit side by side.

Platform fee vs model price

The free plan is about 50 requests a day. It has no platform fee, and it also has no spend controls or prompt caching. Use it to see whether a model answers. Do not use it as a monthly budget. Standard pay-as-you-go buys credits by card or crypto at 5.5%. Business is 8% and adds EU or US in-region routing. Enterprise fees are negotiated, and invoicing is available.

BYOK means you bring the vendor key. Standard and Business both say the fee is waived on the first $25,000 of list-price inference each month, then 5%. Enterprise waives the first $200,000 a month, then the same 5%. Your own key does not change the model price. It removes OpenRouter’s platform fee.

Unit prices across four families

The table uses one workload, cheapest first: 22 working days, 40 requests a day, 880 calls. Each call is 20,000 input tokens and 4,000 output tokens, with no cache. That is 17.6M input and 3.52M output. Model cost = 17.6 × input price + 3.52 × output price.

DeepSeek V4 Flash 0731 is $0.018 in and $0.32 out. On this workload the output is about $1.13 and the input about $0.32. The cheap side is input, not the whole call. GPT-6 Sol and Claude Sonnet 5.5 are both $2 / $10, so both model lines are $70.40. GPT-6 Astra and Claude Fable 5.1 are both $10 / $50, so both are $352. Gemini 3.1 Pro Preview is also $2 in, but $12 out, so the line is $77.44. The gap versus Sol is output.

Three monthly budgets

None of the three bands includes the platform fee. If you buy credits, multiply by 1.055. To use your own volume, change three numbers: call count, input tokens per call, and output tokens per call.

Personal trial: 200 calls a month

Each call is 4,000 tokens in and 800 out. That is 0.8M in and 0.16M out. Flash 0731 is about $0.07, Luna about $0.16, Gemini 3.8 Flash about $1.20, Sol or Sonnet 5.5 about $3.20, Opus 5.5 about $6.40, and Astra or Fable about $16. Every line is under $20. The gap comes from the default model, not from the number of calls.

Daily coding: 880 calls

The comparison table is this band. Ninety percent Sol and ten percent Astra is a model line of $98.56 (0.9 × 70.40 + 0.1 × 352). Ninety percent DeepSeek V4.1 Flash and ten percent Opus 5.5 is $22.63 (0.9 × 9.50 + 0.1 × 140.80). Keep the daily loop on Sol, Sonnet 5.5, or V4.1. Save the flagship for cases you have already measured.

model = 17.6 * input_per_million + 3.52 * output_per_million
with_card_fee = model * 1.055

Heavy agent: multiply by 10

Multiply the 880 calls by 10. The model lines scale about tenfold: Flash 0731 about $14.43, Luna about $35.20, Gemini 3.8 Flash about $264, Sol or Sonnet about $704, Opus about $1,408, Astra or Fable about $3,520. Defaulting this band to Astra is about $3,714 after the card fee, before any tool calls. Lower the default model before you add a machine.

Cache, batch, and routing

A cache read is cheaper than a fresh input. Still on 880 Sol calls: if 80% of input hits the cache, the bill is 14.08M × $0.20 + 3.52M × $2 + 3.52M × $10 = $45.06, not $70.40. Count the discount only when the cache actually hits. Do not budget the first request, or a system prompt you keep rewriting, at that rate.

Batch endpoints on these models are about half the standard price. Sol batch is $1 / $5, so the same 880 calls are $35.20. Offline work that can wait belongs on batch. Interactive coding should not wait for the half-price queue.

OpenRouter can route one model across hosts. DeepSeek prices here often come from a third-party host, not from DeepSeek’s own peak and off-peak table. If the invoice has to match this worksheet, pin the model and the provider in the console and set a spend cap. Free models are rate-limited. Leave them out of a production monthly budget.

The API bill has no Mac

None of the bands above includes compiling for iOS, running the simulator, or signing. OpenRouter is model inference. xcodebuild still needs a Mac that stays awake. On the day this was written, the homepage price pills showed M4 · 16GB / 256GB from $101.7 a month and M4 · 24GB / 512GB from $200.3 a month. Day, week, month, and quarter are all available. Checkout follows the plan page. This note does not turn a starting price into a contract price.

Add the 16GB starting price to the “ninety percent Sol” band: $98.56 of model, about $104 after credits, plus $101.7 for the Mac, about $206 together. Remove the Mac and the API side is still about $104. The two payments are not for the same thing.

FAQ

Does OpenRouter add a markup on tokens?

On the public list for 30 September 2026, standard input and output prices match the vendors. There is no second per-token markup. The extra 5.5% on the standard plan is the fee when you buy credits.

When do you pay the 5.5%?

When you buy credits on the standard plan by card or crypto. BYOK skips that fee on the first $25,000 of list-price inference each month, then charges 5%. The Business platform fee is 8%.

Which model should a monthly budget default to?

A trial uses deepseek/deepseek-v4-flash-0731 or openai/gpt-6-luna. Daily coding uses gpt-6-sol, claude-sonnet-5.5, or deepseek-v4.1-flash. Keep Astra and Fable for cases you have measured. On the same 880 calls, all-Astra is $352 and ninety percent Sol is $98.56.

Can an alias go into the budget?

No. ~deepseek/deepseek-v4-flash-latest can change versions. On the day this was written its output price was $1.25, while 0731’s output price was $0.32. Pin a dated model ID.

Does this bill include a Mac?

No. Inference is on OpenRouter. xcodebuild, the simulator, and signing are on macOS. A 16GB Cloud Mac starts near $101.7 a month on the homepage, paid separately from tokens.

Pin the model, then multiply your tokens

The default model sets the order of magnitude. A personal trial starts on Flash 0731 or Luna. Daily coding stays on Sol, Sonnet 5.5, or V4.1, and hard cases move up to Opus or Astra. Buying credits multiplies by 1.055. BYOK under $25,000 does not. If you need macOS, add a Cloud Mac on its own line. Do not merge it into a “model monthly fee.”

Further reading

• OpenAI Agents API in practice: build an agent, deploy it, and price the API and the server
• Claude Opus 5.5 vs Claude Code: how to choose a 2026 coding model — price, tokens, coding ability, and agent workflow
• What Is GPT-6 Astra? OpenAI’s Latest GPT-6 Features, Pricing, API, Release Date, and How to Use It in ChatGPT