Cheapest AI APIs: Token Rates and Real Workload Costs in 2026
When you're pricing out an AI feature, it's tempting to grab the lowest number on a provider's landing page and call it a day. I get it, we've all done it. But a cheap headline rate doesn't mean much if long prompts push you into higher pricing tiers, or if the model's answers need constant retries.
For the baseline workload we compare here, Qwen Flash on Alibaba's Singapore International tariff comes out to $1.30 for 10 million input tokens and two million output tokens. GPT-6 Luna and Claude Haiku 5.5 each come to $2.00 at their standard context tiers. These are raw inference charges. They aren't a promise of identical answer quality, and they aren't the cash you'll need to fund the account upfront.
The real question is how much it costs to get your actual job done correctly. In this guide, we walk through concrete token rates alongside the rules that change them in production: context thresholds, prompt caching, batch processing, and output reliability.
The rate card: specify the model and the tariff
All amounts below are USD per one million billed tokens. Standard text inference is the baseline, with no cached input, batch discounts, free trials, tools or tax. Monthly totals can get big even when each individual request is short, because the context tier is selected per request, not from monthly token volume.
| Model and billing condition | Input / 1M | Output / 1M |
|---|---|---|
| Qwen Flash, Singapore International, input ≤256K | $0.05 | $0.40 |
| GPT-6 Luna, Standard, input ≤272K | $0.10 | $0.50 |
| Claude Haiku 5.5, prompt ≤100K | $0.10 | $0.50 |
| DeepSeek V4.1 Flash, off-peak | $0.15 | $0.60 |
| DeepSeek V4.1 Flash, peak | $0.30 | $1.20 |
| Gemini 3.8 Flash, introductory Standard tariff | $0.75 | $3.75 |
| Claude Sonnet 5.5, Standard | $2.00 | $10.00 |
Sources: Alibaba model pricing, GPT-6 Luna, Haiku 5.5, DeepSeek pricing, Gemini pricing and Sonnet 5.5.
What to keep in mind: Qwen's row reflects one specific Singapore International Flash tariff, not every regional deployment or model family variant. DeepSeek's current request ID is deepseek-flash, serving DeepSeek-V4.1-Flash-1017. Its peak periods run weekdays 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. If your requests run around the clock, your effective bill will blend peak and off-peak rates.
Gemini's introductory rates run through December 31, 2026. Scheduled Standard rates from January 1, 2027 are $1.50 input and $7.50 output. That makes the same workload twice as expensive without a single extra token, so budget for the renewal period, not just today's promotion.
Free AI API Tiers & Developer Credits Tracker
Every legitimate way to run LLMs, image generation, and agent workflows for $0 in 2026. Filter by rate limits, models, and whether credit card verification is required.
Google AI Studio
Generous always-free tier for flagship Gemini models
- RPM Rate
- 15 RPM
- Tokens/Min
- 1,000,000 TPM
- Daily Limit
- 1,500 RPD
Groq Console
Near-instant LPU inference on leading open models
- RPM Rate
- 30 RPM
- Tokens/Min
- 6,000 to 14,400 TPM
- Daily Limit
- 14,400 RPD
Cloudflare Workers AI
Serverless edge GPU inference with 10k daily neurons
- RPM Rate
- Unlimited
- Tokens/Min
- Burst to 10k Neurons
- Daily Limit
- 10,000 Neurons/day
GitHub Models (Azure AI)
Direct access to OpenAI and Meta models with a GitHub account
- RPM Rate
- 15 RPM
- Tokens/Min
- 8,000 TPM
- Daily Limit
- 150 RPD
Together AI
$5.00 free credit to run open-weights models and FLUX images
- RPM Rate
- 60 RPM
- Tokens/Min
- Standard Tier
- Daily Limit
- Capped by Balance
Cerebras Inference
World record token speeds on wafer-scale AI hardware
- RPM Rate
- 30 RPM
- Tokens/Min
- 60,000 TPM
- Daily Limit
- 1,000,000 TPD
SambaNova Cloud
Enterprise-grade SN40L Reconfigurable Dataflow chip API
- RPM Rate
- 20 RPM
- Tokens/Min
- 40,000 TPM
- Daily Limit
- Reasonable Use
OpenRouter Free Models
Aggregated free community endpoints across 20+ models
- RPM Rate
- 20 RPM
- Tokens/Min
- Shared Pool
- Daily Limit
- 200 RPD
Mistral La Plateforme
Official free experimentation tier for Codestral and Mistral Small
- RPM Rate
- 1 RPS (60 RPM)
- Tokens/Min
- 1,000,000 TPM/mo
- Daily Limit
- Monthly Cap
DeepSeek Platform
5,000,000 input & output token grant on phone registration
- RPM Rate
- Standard Tier
- Tokens/Min
- Standard Tier
- Daily Limit
- Grant Balance
Cohere API
Free evaluation keys for enterprise RAG, Embeddings & Rerank
- RPM Rate
- 10 RPM
- Tokens/Min
- 10,000 TPM
- Daily Limit
- 1,000 calls/mo
RunPod Serverless
On-demand cloud GPU seconds with initial developer promo
- RPM Rate
- Auto-scaling
- Tokens/Min
- GPU VRAM Bound
- Daily Limit
- Balance Bound
Indie Developer Pro-Tip: The Triple-Tier Free Stack
You can build and test complete production-grade applications without paying a single dollar by combining Google AI Studio (for massive 1M token context reasoning), Groq (for sub-second streaming and voice response), and Cloudflare Workers AI (for serverless edge execution). Once you exceed rate limits, inspect the unit economics in our rate card below before upgrading.
Two workloads you can reproduce
Assume lots of individual requests, each with at most 50K input tokens, so everything stays within every listed lower pricing tier, with no chargeable extras. The baseline formula is simple: input millions × input rate + output millions × output rate.
Our smaller scenario represents roughly 5,000 requests averaging 2,000 input tokens and 400 output tokens apiece (10 million input and two million output tokens). The second scenario scales that demand tenfold to 100 million input and 20 million output tokens.
| Model / tariff | 10M input + 2M output | 100M input + 20M output |
|---|---|---|
| Qwen Flash, stated International tariff | $1.30 | $13.00 |
| GPT-6 Luna, lower context tier | $2.00 | $20.00 |
| Claude Haiku 5.5, lower context tier | $2.00 | $20.00 |
| DeepSeek V4.1 Flash, all off-peak | $2.70 | $27.00 |
| DeepSeek V4.1 Flash, all peak | $5.40 | $54.00 |
| Gemini 3.8 Flash, introductory Standard | $15.00 | $150.00 |
| Claude Sonnet 5.5, Standard | $40.00 | $400.00 |
Here's the math, so you can check it on a napkin. For Qwen, 10 × $0.05 + 2 × $0.40 = $0.50 + $0.80 = $1.30 (or $13.00 at 10x volume). Luna and Haiku each give 10 × $0.10 + 2 × $0.50 = $2.00 (or $20.00). DeepSeek off-peak comes to $1.50 + $1.20 = $2.70 (or $27.00), and at peak rates it reaches $5.40 (or $54.00). If your traffic splits evenly between peak and off-peak hours, you land on the average: $4.05 for the smaller batch.
If you run those same Gemini tokens under the scheduled rates starting January 1, 2027, plan for $30.00 and $300.00 respectively. Claude Sonnet 5.5 is here as a higher-priced quality reference at $40.00 and $400.00. It's not a suggestion that every basic automation needs a flagship model.
Context changes the result before monthly volume does
Thresholds apply per request, not across your monthly total. For example, Claude Haiku 5.5's prompt tier above 100K jumps to $0.50 input and $2.50 output. If every request in our 10M/2M workload carries an oversized prompt that crosses that 100K line, the bill leaps to $10.00 instead of $2.00.
Qwen Flash prompts above 256K rise to $0.25 input and $2.00 output, which totals $6.50 for the same token volume. GPT-6 Luna above 272K doubles the input rate to $0.20 and increases output by 1.5x to $0.75, yielding $3.50.
The takeaway: keep an eye on individual request lengths in your logs. You won't pay a long-context penalty just because your monthly volume reaches 100 million tokens. But shoving unnecessary boilerplate into single prompts can quietly multiply your bill.
Caching: count hits and writes separately
Prompt caching can deliver major savings, but only if you track writes and reads as separate meters. On Luna's standard tier, cached reads cost $0.01 per million, while cache writes cost $0.125 per million.
Say you run a monthly workload of 10 million input and two million output tokens, where seven million input tokens hit the cache and three million miss. The read-side charge is 3 × $0.10 ($0.30) + 7 × $0.01 ($0.07) + 2 × $0.50 ($1.00) = $1.37. If you also make a separate one-million-token cache-write request, add $0.125 for a total of $1.495 (about $1.50).
That extended scenario has 11 million input tokens in total, including the additional write, so don't charge a write again as ordinary input. For a fixed 10-million-input workload split into 2M ordinary input, 1M writes and 7M reads, the bill is $0.20 + $0.125 + $0.07 + $1.00 output = $1.395. Looking only at the read rate hides the write cost.
Caching also depends on eviction policies and prompt consistency. If prompts vary slightly at the start, or the cache expires between calls, requests fall back to uncached miss rates. Check the provider's reported usage categories in the response body before assuming a discount, because field names and write policies differ across APIs.
One more thing: some providers, like Google for Gemini, bill an explicit token-hour storage fee for keeping cached context warm. If your workload rarely hits that stored context, the ongoing storage fees can easily erase your read-side savings.
Batch processing: useful when the job can wait
If your task isn't interactive and can wait several hours for a response, batch APIs offer a dependable half-price discount. Official batch endpoints for GPT-6 Luna, Claude Haiku, and Gemini offer qualifying batch inference at 50% off standard rates, bringing our 10M/2M workload down to $1.00, $1.00, and $7.50 respectively.
Qwen's Singapore International Flash tariff offers batch processing at $0.025 input/$0.20 output, dropping the workload to $0.65.
The practical catch: treat batch as an operational tradeoff, not a blanket coupon. Alibaba explicitly notes that batch and context-cache discounts can't be combined on the same request. And if human users or live customer automations are waiting on answers, cheaper overnight processing solves the wrong problem.
The cost per accepted result matters more than the cheapest row
A lower token rate can lose its edge when it takes more attempts to deliver an acceptable result. So measure the complete job rather than treating the cheapest table row as a quality ranking.
Suppose 5,000 tasks produce 4,500 usable results on the first attempt. At $2.00 of total inference, that first-pass cost looks tiny, about $0.000444 per accepted result. But you still have 500 failed or malformed outputs. Retrying those 500 tasks adds another $0.20 of compute, and if they fail again, someone on your team has to review them manually. Count that review time in the comparison instead of assuming retries alone fix every failure.
Before committing your architecture to the cheapest row in a table, run a small test suite of 50 to 100 representative examples from your real workload. Measure format compliance, factual accuracy, and how often answers need retries.
A common, cost-effective design is a tiered fallback. Send routine, predictable requests to an ultra-cheap model like Qwen Flash, GPT-6 Luna, or DeepSeek Flash, and route validated failures or complex edge cases to a stronger model like Claude Sonnet. Our AI gateway comparison covers routing architectures, and our token budget guide walks through hard spending caps.
Consumption, account funding and infrastructure
An estimated $1.30 usage charge doesn't mean your credit card gets charged $1.30. A provider might require prepaid funds, a minimum credit purchase, or billing in a specific currency. So it helps to keep your deposited balance and your actual token consumption separate in your head (and your spreadsheet).
Self-hosting an open model works the same way: you swap token fees for fixed server rentals, GPU memory overhead, and maintenance hours. A low per-hour GPU rate can sound appealing, but if your application only handles requests sporadically throughout the day, paying for an idle server running 24/7 will quickly outpace a small token bill.
Common questions
Which is the cheapest AI API?
Among the stated tariffs and workload assumptions, Qwen Flash's Singapore International row has the lowest uncached charge. Availability, request shape and acceptable results can change your choice, so think of this as a defined comparison, not an exhaustive market ranking.
Are GPT-6 Luna, Haiku 5.5 and DeepSeek V4.1 real endpoints?
Yes. Their official documentation is linked above. Check model identifiers and release notes rather than substituting an older model's tariff because a name looks unfamiliar.
Can I use monthly totals to choose a context tier?
No. Check each request's applicable prompt/context threshold. Thousands of short requests can consume millions of tokens without ever crossing a long-context boundary.
How should I choose the first model to test?
Start with a model that meets your language, output format, context and data-handling requirements. Run the same representative tasks through it, then compare the full cost per accepted result. The lowest token bill is great, as long as the output actually does the job.