Tool
API rate limit planner
Turn a batch job into requests and tokens per minute, find the published rate-limit tier that finishes it before your deadline, and see what batch processing and rented GPUs change. 50,000 L1 tasks (a 1,500-token document in, a 200-token answer out) in 3 hours need 278 requests and 416,667 input tokens a minute. With Claude Sonnet 5.5 the published limits allow it from the Start tier, where the job takes 50 min at those limits. At list prices it costs $250.00, or $125.00 with batch processing if the work can wait up to 24 hours.
Calculated Runs in your browser. Your inputs stay on this device. All calculators · Rate limits by tier · Task cost calculator
Result Calculated Claude Sonnet 5.5 · 50,000 tasks in 3 hours · data of 9 Nov 2026
Fits the deadline from the Start tier of Anthropic.
- Requests per minute, at an even pace
- 278
- Input tokens per minute
- 416,667
- Output tokens per minute
- 55,556
- On your tier it takes
- 50 min
- Requests in flight at once
- 13
- Job cost at list prices
- $250.00
50,000 requests · 75M input tokens (0 from cache) · 10M output tokens in the whole job.
| Tier | requests per minute | input tokens per minute | output tokens per minute | spend per month | This job takes | Within 3 hours |
|---|---|---|---|---|---|---|
| Start | 1,000 | 2M | 400,000 | $500 | 50 min | yes |
| Build | 5,000 | 5M | 1M | $1,000 | 15 min | yes |
| Scale | 10,000 | 10M | 2M | $200,000 | 8 min | yes |
Published Limits: Anthropic rate limits page, read · Anthropic batch page, read . Price: $2.00 per 1M input, $10.00 per 1M output, Anthropic price list, as listed in the United States on 9 Nov 2026.
- Cache: Cached input does not count toward input tokens per minute here, as Anthropic states: “For most Claude models, only uncached input tokens count toward your ITPM rate limits.”
- Seconds per request: From our own test requests: a short prompt, not your inputs; long inputs take longer.
- Tiers: “Limits are defined by usage tier. Organizations are placed on a tier automatically based on usage history and account standing and can move to a higher tier over time as they use the API.”
Batch processing
- Batch processing: 50% off, $125.00 for this job.
- The batch window is 24 hours and your deadline is 3 hours: batch results are not promised in time. Anthropic says: “The system processes each batch as fast as possible, with most batches completing within 1 hour.”
Every model on this job
Calculated 50,000 tasks of the size above in 3 hours, at each model’s lowest published tier and at list prices. Batch cost: the provider’s published discount; off-peak at DeepSeek.
| Model | Lowest tier for the deadline | Time on the lowest tier | Job cost | With batch |
|---|---|---|---|---|
| Claude Opus 5.5Anthropic | Start | Start: 50 min, by requests per minute | $500.00Anthropic price list | $250.00 |
| Claude Sonnet 5.5Anthropic | Start | Start: 50 min, by requests per minute | $250.00Anthropic price list | $125.00 |
| Claude Haiku 5.5Anthropic | Start | Start: 50 min, by requests per minute | $12.50Anthropic price list | $6.25 |
| GPT-6 AstraOpenAI | Launch | Build: 76.0 days, by spend per month | $1,250.00OpenAI price list | $625.00 |
| GPT-6 SolOpenAI | Build | Build: 1 h 25 min, by tokens per minute | $250.00OpenAI price list | $125.00 |
| GPT-6 LunaOpenAI | Build | Build: 42 min, by tokens per minute | $12.50OpenAI price list | $6.25 |
| Gemini 4 ArgonGoogle | Tier 2 | Tier 1: 4 h 10 min, by spend per 10 minutes | $250.00Google price list | $125.00 |
| Gemini 3.8 FlashGoogle | Tier 1 | Tier 1: 1 h 34 min, by spend per 10 minutes | $93.75Google price list | $46.88 |
| Grok 4.7xAI | Tier 0 | Tier 0: 6 min, by requests per second | $210.00xAI price list | — |
| DeepSeek V4 ProDeepSeek | Account | Account: 6 min, by requests at once | $138.60DeepSeek price list | $69.30 |
| DeepSeek FlashDeepSeek | Account | Account: 42 s | Unknownprice not in our data | — |
| gpt-oss-120b on GroqGroq | above published tiers | Developer: 14.0 months, by tokens per day | $17.25Groq price list | $8.62 |
| Qwen3.8 27B on GroqGroq | above published tiers | Developer: 14.0 months, by tokens per day | $100.00Groq price list | $50.00 |
| gpt-oss-120b on CerebrasCerebras | Developer | Free Trial: 85.0 days, by tokens per day | Unknownprice not in our data | — |
| Qwen3.8 27B on CerebrasCerebras | above published tiers | Developer: 9 h 27 min, by uncached tokens per minute | Unknownprice not in our data | — |
| Mistral Large 4Mistral | no numbers published | — | $73.50Mistral price list | $36.75 |
The same output on rented GPUs
Calculated The job writes 10M output tokens. Qwen3.8 27B with 8-bit weights and 16 requests per server is cheapest on NVIDIA RTX A4000: 3 servers of 4 cards to finish in 3 hours, 25.7 GPU-hours, $3.86 for the busy time at $0.15 per card-hour (Hyperstack, lowest listed on-demand price).
| GPU | Cards per server | Output tokens/s per server | Servers to finish in time | GPU-hours | Per card-hour | Cost of busy time | Renting them for the deadline |
|---|---|---|---|---|---|---|---|
| NVIDIA RTX A4000 | 4 | 432 | 3 | 25.7 | $0.15 | $3.86 | $5.40 |
| NVIDIA GeForce RTX 3090 | 4 | 902 | 2 | 12.3 | $0.50 | $6.16 | $12.00 |
| NVIDIA GH200 Grace Hopper 96 GB | 1 | 964 | 1 | 2.9 | $2.59 | $7.46 | $7.77 |
| NVIDIA RTX A6000 | 1 | 185 | 6 | 15.0 | $0.50 | $7.50 | $9.00 |
| NVIDIA A100 PCIe 80 GB | 1 | 466 | 2 | 6.0 | $1.35 | $8.04 | $8.10 |
| NVIDIA B200 | 1 | 1,928 | 1 | 1.4 | $6.00 | $8.64 | $18.00 |
| NVIDIA A100 SXM 80 GB | 1 | 491 | 2 | 5.7 | $1.60 | $9.04 | $9.60 |
| NVIDIA H200 141 GB | 1 | 1,157 | 1 | 2.4 | $3.99 | $9.58 | $11.97 |
| NVIDIA A40 | 1 | 168 | 6 | 16.6 | $0.59 | $9.77 | $10.62 |
| NVIDIA GeForce RTX 4090 | 4 | 972 | 1 | 11.4 | $0.89 | $10.18 | $10.68 |
The 10 cheapest of 21 GPU types that fit. Prices as shown in the United States on . Formula of the VRAM calculator: memory for weights and the KV cache of the longest request (1,700 tokens), output speed from memory bandwidth. The estimate leaves out reading the input, so time and cost here are a floor; setup, idle time, storage and network are extra.
How the planner works
- Tokens per task: for each step, input = input at step 1 + growth × (step − 1) + tool schema; output = visible output + hidden reasoning. The cached share of step 1 input is read from cache, and later steps read the earlier context from cache, as in the task cost calculator.
- Needed per minute: the whole job spread evenly over the deadline: requests, input and output tokens.
- Time on a tier: for every limit the provider publishes, the job’s total divided by the limit; the longest one is the time. Daily and hourly limits count only when the job is bigger than one day’s or hour’s allowance. Cached input is left out of a limit only where the provider says it does not count. A task’s requests run one after another, so the job also takes at least steps × seconds per request. Without seconds per request the time comes from the limits alone and is shown as a floor (“at least”).
- Fits the deadline: the first tier whose time is within the deadline.
- Requests in flight: requests per second × seconds per request, rounded up, and never more than the number of tasks.
- Cost: cost per task from the provider’s own price list × tasks; batch applies the provider’s published discount. Without the provider’s own price the cost is unknown: a price from another seller comes with that seller’s limits, not these tiers.
Assumptions
| What | Value | Why |
|---|---|---|
| Pace Assumption | even over the deadline | Limits are token buckets that refill continuously; a burst at the start needs more. |
| Retries Assumption | none | Requests rejected with 429 and repeated are not added. |
| Other traffic Assumption | none | Limits are per organization or project; anything else you run shares them. |
| Uncached tokens Assumption | input not read from cache, plus output | Cerebras names an uncached token limit without listing what it covers. |
| Model families Assumption | OpenAI’s table | OpenAI lists limits by family (“Astra, Sol, Terra”, “Luna”), not by version; applied here to GPT-6 models of those families. |
| Seconds per request Assumption | our measurement or yours | Measured on a short request; long inputs and long answers take longer. |
| GPU speed Assumption | formula, not measured | Memory-bandwidth estimate of the VRAM calculator; it leaves out reading the input. |
Sources
Published Every limit, tier rule and batch term is quoted from these provider pages; the full tables are on API rate limits by tier.
- Anthropic rate limits page, read
- Anthropic batch page, read
- OpenAI rate limits page, read
- OpenAI batch page, read
- Google gemini rate limits page, read
- Google gemini batch page, read
- xAI rate limits page, read
- DeepSeek rate limit page, read
- DeepSeek pricing page, read
- Groq rate limits page, read
- Groq batch page, read
- Cerebras rate limits page, read
- Mistral usage limits page, read
- Mistral batch page, read