Concept renderA cloud model charges you per token for the life of every device, while an edge model charges you once in bill of materials, battery, heat and certification. This guide lists the official per-token prices as of October 31, 2026, gives a break-even formula, and shows which inputs you must measure on your own board.
For an AI device, the model choice is a financial decision as much as a technical one. A cloud model adds a bill that grows with every call from every unit you have sold. An edge model moves that cost into the product itself, and adds engineering risk. This guide gives you the published prices to start from and a formula to finish with your own numbers. It does not contain a Duobay price or a measured result.
What do the published cloud prices look like?
These are the official list prices per million tokens from each provider's own pricing page. Each row was read on October 31, 2026. Prices change, so re-read the page before you decide.
| Model | Input, per million tokens | Output, per million tokens | Note | Source |
|---|---|---|---|---|
| Gemini 2.5 Flash-Lite | $0.10 (text, image, video); $0.30 (audio) | $0.40 | Standard tier. | Google pricing page, as of October 31, 2026 |
| Gemini 3.8 Flash | $0.75 through December 31, 2026; $1.50 from January 1, 2027 | $3.75 through December 31, 2026; $7.50 from January 1, 2027 | The price changes on a fixed date inside your first year. | Google pricing page, as of October 31, 2026 |
| Claude Haiku 5.5, prompts up to 100,000 tokens | $0.10 | $0.50 | Cache hit input is $0.01. | Anthropic pricing page, as of October 31, 2026 |
| Claude Haiku 5.5, prompts over 100,000 tokens | $0.50 | $2.50 | Five times the short-prompt rate on both sides. | Anthropic pricing page, as of October 31, 2026 |
| Claude Haiku 4.5 | $1 | $5 | Listed on the same page. | Anthropic pricing page, as of October 31, 2026 |
| Claude Sonnet 5.5 | $2 | $10 | Listed on the same page. | Anthropic pricing page, as of October 31, 2026 |
| gpt-5-nano | $0.05 | $0.40 | Batch is about half the standard price. | OpenAI pricing page, as of October 31, 2026 |
| gpt-4.1-mini | $0.40 | $1.60 | Batch is about half the standard price. | OpenAI pricing page, as of October 31, 2026 |
| gpt-4o-mini | $0.15 | $0.60 | Batch is about half the standard price. | OpenAI pricing page, as of October 31, 2026 |
Voice can cost more per token
A device that listens and speaks may use an audio model. Audio models bill audio tokens. The gap to text depends on the model you compare. gpt-realtime-2.1-mini audio input is $10 per million tokens, against $0.05 for gpt-5-nano text input, which is 200 times higher. Claude Sonnet 5.5 text input is $2, the same as the Gemini 3.5 Transcribe audio input rate of $2.00, so against that text row the gap is zero. Audio tokens per second of speech differ from text tokens per word, so compare cost per minute of real sessions.
| Model | Input, per million tokens | Output, per million tokens | Source |
|---|---|---|---|
| gpt-realtime-2.1-mini | Audio: $10 | Audio: $20 | OpenAI pricing page, as of October 31, 2026 |
| gpt-realtime-2.1 | Audio: $32 | Audio: $64 | OpenAI pricing page, as of October 31, 2026 |
| Gemini 3.5 Transcribe | Audio: $2.00, or $0.003 per minute | Not applicable (transcription) | Google pricing page, as of October 31, 2026 |
The per-minute figure for the transcription model assumes 25 audio tokens per second, which is how the conversion from tokens to minutes works. Check the provider page for the current conversion.
What is a public reference for edge hardware?
The Raspberry Pi AI HAT+ 2 is a useful public data point because its price is on the manufacturer page. It uses a Hailo-10H accelerator, lists 40 TOPS (INT4) and carries 8GB of RAM. Its list price is $200 (Raspberry Pi product page, as of October 31, 2026).
Use it only as a reference. It is a developer accessory. A production design needs its own quote. The page does not publish power draw and does not name a specific language model that runs on it. Your own edge cost comes from the chip or module you choose, quoted at your volume.
How do you compute the five-year cost per device?
Cloud cost per device per year is the number of calls times the cost of one call.
cloud_per_device_per_year = N × 365 × (T_in × p_in + T_out × p_out) / 1,000,000
cloud_fleet_five_years = D × (Y1 + Y2 + Y3 + Y4 + Y5)
where each Yn uses the price in force in that year
edge_five_years = D × (ΔBOM + ΔOther) + 5 × M
E = ΔBOM + ΔOther (one-time, per device)
m = M / D (yearly maintenance, per device)
C = cloud_per_device_per_year (per device, at a fixed price)
break_even_months = 12 × E / (C − m) valid only when C > mUnits: D, N, T_in, T_out are counts. E, m and C are in dollars per device (E once, m and C per year). M, cloud_fleet_five_years and edge_five_years are in dollars for the whole fleet. The break-even line comes from setting the two fleet totals equal after t months: D × E + M × t/12 = D × C × t/12. Divide by D and solve for t. If C is not larger than m, the edge option never pays back its extra cost on these inputs.
Worked example with your own assumptions. Suppose C − m comes out equal to E: the yearly cloud cost per device, minus the per-device share of maintenance, equals the one-time extra edge cost. Then break_even_months = 12 × E / E = 12. If C − m is half of E, the result is 12 × E / (E / 2) = 24 months. If C − m is twice E, it is 6 months. Replace E, M, D, N, T_in and T_out with your own figures, compute C and m, and apply the same line. C uses one year's price, so check the answer against the five-year fleet totals when prices change.
Compute the cloud side once per year, because prices change and usage grows. The edge side is mostly paid before the device ships, so it is the one you cannot adjust later. If the edge design also needs a cloud call for hard requests, add that residual cloud cost to the edge side with its own N.
Inputs, and which are yours to replace
Only the prices come from a source. Every other input is an assumption for you to replace. We do not supply example values, because a made-up number would look like a fact.
| Variable | Meaning | Status | What to do |
|---|---|---|---|
| D | Active devices in the field | Your assumption | Replace with your own plan. |
| N | Cloud model calls per device per day | Your assumption | Replace after you log real sessions on a prototype. |
| T_in, T_out | Input and output tokens per call | Your assumption | Replace after you measure real prompts, including any system prompt and history you resend. |
| p_in, p_out | Price per million input and output tokens | Published price | Take from the table above. Check the date and the prompt-length tier. |
| ΔBOM | Extra bill-of-materials cost of the edge design, per device | Your assumption | Replace with a quoted figure for your chosen silicon at your volume. The $200 board in this article is a public reference. It is a developer accessory price and says nothing about a mass-production quote. |
| ΔOther | Battery, thermal, enclosure and certification increments, per device | Your assumption | Replace after a power and heat measurement on your own board. |
| M | Edge model maintenance for the whole fleet: OTA delivery, model updates, regression testing, per year | Your assumption | Replace with your engineering and hosting estimate for the whole fleet. Divide by D to get the per-device share used in the break-even formula. |
Run the formula at least three ways: today's prices, the higher price where one is published (for example the Gemini 3.8 Flash rate from January 1, 2027), and a case where your calls per day double. If the answer flips between cases, the decision depends on a number you have not measured yet.
What are the common cost traps?
These follow from the published prices and specs above. They come from engineering reasoning about how the numbers behave. None describes an incident at any company.
- A price with an end date. The Gemini 3.8 Flash input rate doubles on January 1, 2027, and the output rate doubles too. A device that ships in late 2026 and models year-one cost at the launch price understates every later year.
- A long-prompt tier. Claude Haiku 5.5 charges five times more per token once a prompt passes 100,000 tokens. A device that keeps appending conversation history can cross that line without any change to the product.
- Audio priced like text. The OpenAI realtime audio rates are $10 input and $20 output per million tokens for the mini model. The text input rate for gpt-5-nano is $0.05, so the audio input rate is 200 times higher. The gap depends on which models you compare, so do not reuse a text estimate for a voice product.
- Spec-sheet TOPS as a speed promise. The reference board page lists 40 TOPS (INT4) and 8GB of RAM. It does not list power draw or any named language model. A TOPS figure does not tell you tokens per second for your model, so you cannot compute edge speed or battery life from it.
- Cache discounts you cannot count on. A cache-hit input price such as $0.01 per million tokens applies only when the request actually hits the cache. Model both the hit and miss cases, and treat the hit rate as an assumption.
- A hosted model that gets retired. A cloud model your firmware calls can be replaced or withdrawn by the provider. Budget engineering time for a migration and for retesting prompts, and plan how an installed device is updated.
What should you measure on your own board?
The published sources do not give edge speed, power, heat or answer quality for your model on your silicon. We have not measured these for you, and no result appears in this article. Measure them yourself before you commit to the edge path.
| What to measure | How | How to set the pass line |
|---|---|---|
| Tokens per second on the target chip | Run your chosen small model on the real module with a fixed prompt set. Record tokens per second for the first token and for steady output, at room temperature and again after a long continuous run. | Set the pass line from the slowest reply your product can tolerate. Write it down before you run the test. |
| Power draw during inference | Put a current meter or power analyzer on the battery rail. Record idle, one inference and a repeated-use session. Do not rely on a development-board figure, because the board vendor page may publish none. | Compute battery hours from your real usage profile, then compare with your product promise. |
| Heat in the enclosure | Run a repeated-use session in the real housing, held in a hand or worn. Log surface temperature at the skin contact point and the chip temperature, and watch for throttling. | Pick a surface temperature limit from the applicable safety standard for your category before you test. Fail if the chip throttles below your speed line. |
| Battery impact | Compare battery life over a full day of your usage profile with the edge model on and with it off (cloud call instead). | The extra battery capacity or volume needed to hold your promised runtime becomes a line in ΔOther. |
| Answer quality of the small model | Build a test set from real user requests (log them on a prototype). Score the edge model and the cloud model on the same set with a rubric written before scoring. | Set the minimum acceptable score and the maximum gap to the cloud model in advance. Include the hard cases as well as the easy ones. |
| Tokens and calls per real interaction | Log prompts and replies from prototype sessions, including resent history and system prompts. Count tokens with the provider tokenizer. | The averages and the 95th percentile replace T_in, T_out and N in the formula. |
Fix every pass line before you run the test. A threshold chosen after you see the result tends to land just under the result.
What should you ask before you commit?
Take this list to your team, and to any supplier who proposes a cloud or an edge design.
- Which exact model name and pricing tier does our firmware call, and on what date was that price read?
- What happens to our unit cost when a promotional price ends? Have we modeled each year separately?
- What is the 95th percentile prompt length in our logs, and does any tier change near it?
- Is the voice path priced by audio tokens or by minutes, and which assumption converts one to the other?
- Who pays if the provider retires the model, and how many weeks of firmware and prompt retesting does a swap take?
- For the edge option, what is the quoted price of the chip or module at our volume, and where is the datasheet that states power draw?
- Does the edge option need a larger battery, a heat spreader or a different enclosure, and what does each cost?
- What is the update path for the edge model after launch, and who tests each new version?
- Does the recurring cost fall on us as a subscription we must charge for, or on a margin we have already set?
- Which assumptions in the model are measured on our own board, and which are still guesses?
Questions about cloud and edge cost
Is an edge model always cheaper over five years?
No, and this article does not claim a result. The answer depends on your fleet size, how often each device calls the model, how long each call is, and what the edge hardware adds. The break-even formula above lets you test your own numbers.
Why not just use the $200 board price as the edge cost?
That price is the public list price of one developer accessory from the manufacturer page (source, as of October 31, 2026). It is a reference point. A production design uses different parts at different prices and adds power, thermal and certification changes. Replace it with a quote.
Do I need to model the price rising?
You should model both directions. Some published prices have an end date, and long prompts can move you to a higher tier. Run the model with the current price and with at least one higher case.
How do I price a voice device?
Start from the audio token rates or per-minute rates on the provider page. Audio tokens per second differ from text tokens per word, so compare cost per minute of real sessions. Then measure how many minutes of audio a real session uses.
Can Duobay tell me the cost for my device?
We do not publish prices. A Gate 0 Feasibility study produces a costed bill of materials and an architecture choice for your product. Tell us about your device and we will explain how it works.
If you want a costed bill of materials and an architecture recommendation for your device, tell us about your device. You can also read how our engineering gates handle the cloud-or-edge decision.
Sources
All pages accessed October 31, 2026. Prices and model lists change, so re-read them before you decide.
- Google, Gemini API pricing: https://ai.google.dev/gemini-api/docs/pricing (accessed 2026-10-31)
- Anthropic, Claude API pricing: https://platform.claude.com/docs/en/about-claude/pricing (accessed 2026-10-31)
- OpenAI, API pricing: https://developers.openai.com/api/docs/pricing (accessed 2026-10-31)
- Raspberry Pi, AI HAT+ 2 product page: https://www.raspberrypi.com/products/ai-hat-plus-2/ (accessed 2026-10-31)
Planning a device like this? Tell us what it needs to do. A Duobay engineer replies within 1 business day. Send a project brief →