The Bedrock tier ladder: batch and flex at 50% off, priority at +75%

Published October 15, 2026

If nobody is waiting on the response, you’re probably paying double. A month of Nova Pro usage (10 million input tokens, 2 million output) costs $7.20 on batch, $14.40 on standard, and $25.20 on priority. Same model, same answers, and 3.5x between the cheapest way to buy them and the most expensive. Standard is what you pay for not deciding.

A service tier is the pricing lane you submit a request through, and batch and flex are half price. The catch is availability. Only 76 of the 149 models Bedrock prices in us-east-1 offer batch, and the newest Claude models offer no discounted tier at all.

Batch and flex halve the bill. Priority adds three quarters.

Every tier is a multiple of the standard rate. This table shows three models; look at the right-hand columns, where the multiples barely move.

Rates are us-east-1 from our Bedrock dataset. Batch and flex carry identical prices on all 96 model and token-type combinations where both appear. Two names, one price. Two lanes that match to the cent on every combination look like policy, not cost accounting. Our read is that AWS is selling spare capacity to anyone who can wait. What separates them is how you submit. Batch is an asynchronous job whose results land in S3. Flex is a synchronous call that accepts higher latency. Priority is the premium lane.

The tier usually moves the bill more than the model does. Swap Nova Pro for Llama 3.3 70B at standard rates and the same 10M-input, 2M-output month costs $8.64. That is 40% below Nova Pro standard and still 20% above Nova Pro on batch. Batching the model you already evaluated beats switching to a cheaper one you haven’t.

A frontier model plays in a different league. Claude Sonnet 5 runs $40.00 for the same month, and no Nova Pro tier gets near it.

Llama’s low bill comes from its output price. It charges the same $0.72 for input and output. Nova Pro charges 4x more for output than input, Claude Sonnet 5 charges 5x, and Mistral Large 3 charges 3x. At a 5:1 input-to-output mix, Llama is 40% cheaper than Nova Pro even though its input rate is only 10% lower.

About half the catalog offers a discount tier

A tier you can’t buy saves nothing. In us-east-1, 149 models carry a price, and every one but Llama 3.1 405B has a standard rate. The rest depends on the model.

Add it up and batch covers 76 models, while flex and priority cover 49 each. Seventy-two models have standard pricing and nothing else. By provider:

The total row covers all 18 providers; the ten left out of the table account for the difference.

The full ladder sits mostly in open-weight families (Qwen, Gemma, Mistral, DeepSeek, MiniMax, Moonshot, GLM) plus Nova. Anthropic has 15 models with batch and none with flex or priority. The newest Claude models, including Sonnet 5, Sonnet 5.5, Opus 4.7 and Opus 4.8, have no discounted tier in our data. If your shortlist is the newest Claude, the tier decision was made for you, at full price.

Llama 3.1 405B is the reverse case. It has only a batch rate in us-east-1, so in this dataset it can’t be called interactively. The committed rows in the file are provisioned-throughput and image-generation commitments, not a token tier.

Location isn’t the constraint. All four tiers appear in all 36 priced regions, so the model decides what you can buy. The compare page shows the tier gaps next to the rates.

Caching pays for itself on the second use

Prompt caching lets Bedrock hold a long prompt you send repeatedly, so you don’t pay full price to resend it. Think of it as a deposit. You pay a premium the first time, then each reuse costs a fraction of the normal rate.

Where a model prices it, a 5-minute cache write costs 1.25x the input rate and a 1-hour write costs 2x. A read costs 0.10x input on 22 of the 34 models that offer one, Claude Sonnet 5 among them. Nova reads at 0.25x and Grok 4.3 at 0.16x. In us-east-1, 34 models carry a cache-read rate, 31 a 5-minute write rate, and 17 a 1-hour write rate. All 17 of the 1-hour models are Anthropic.

The break-even takes one formula. Let w be the write price divided by the input price, r the read price divided by the input price, and N the number of requests that use the prompt, one of them writing it:

N* = (w − r) / (1 − r)

On Sonnet 5’s 5-minute cache that is (1.25 − 0.10) / 0.90 = 1.28 uses, so two uses clear it. The 1-hour write is (2.00 − 0.10) / 0.90 = 2.11, so it needs three.

Here is what that looks like for a 20,000-token prompt on Sonnet 5, where input costs $2.00 per million. Uncached, each use costs $0.04. A 5-minute write costs $0.05, each read costs $0.004, and a 1-hour write costs $0.08.

The 5-minute column assumes each hit refreshes the window, so requests spaced under five minutes apart keep one entry alive. Space them further apart and every burst pays for a new write. The 1-hour write suits prompts reused across a slower workflow. It needs three uses instead of two, then settles at the same 0.10x read rate. A hundred uses spread across a month is a hundred writes unless the traffic is bursty enough to stay inside a window.

Nova is the odd one. Every Nova model in our data carries a $0.00 cache-write rate in every region that prices one, which makes the numerator negative and means caching pays from the first reuse. A $0.00 line on a price list is worth finding on an invoice before you build a plan around it. Nova’s reads also run at 0.25x input, so a hit saves 75% per token rather than 90%.

Caching trims about 8% off the month. The tier trims 50%.

Back to the Nova Pro month, costed per tier:

Now give 100 of those requests the same 20,000-token system prompt. That is 2M of the 10M input tokens. On standard with Nova Pro’s cache, the arithmetic is:

uncached:  2M tokens × $0.80/M                          = $1.60
cached:    1 write ($0.00) + 99 reads (1.98M × $0.20/M) = $0.40
month:     $14.40 − $1.60 + $0.40                       = $13.20  (−8%)

Flex supports the same cache. Its input is $0.40 and its cache read is $0.10, so the reads cost $0.20 instead of $0.80 and the month lands at $6.60. That is 8% below plain flex and 54% below standard. Nova Pro’s batch tier has no cache-read rate, so the cached batch month stays at $7.20.

AWS sells two large discounts here and, on 33 of the 34 cache-capable models, won’t let you use both. Only Grok 4.3 has a batch cache-read rate in us-east-1. The others price cache reads on standard, and sometimes on flex and priority, but not batch. A batch job that reuses a 20k prompt on every request pays full batch input for every copy.

On Sonnet 5, the same 2M cached tokens cost about $0.45 against $4.00 uncached. That takes $3.55 off a $40.00 month, or 9%. Caching moves the bill by under a tenth here, and the batch tier, where a model offers one, moves it by half.

Caveats in the data

• Regions move the dollars, not the multiples. Nova Pro’s standard input rate runs $0.80 in us-east-1 and $1.28 in eu-south-1. The percentages in this post hold in every region.
• Priority is only nominally 1.75x. Sub-cent rates round. A $0.06 input token becomes $0.11 under priority (1.83x), and Qwen3 Next 80B’s $0.14 becomes $0.26 (1.86x).
• Two models have more than one row per tier and token type. Claude Sonnet 5 prices a global endpoint at $2.00 input and $10.00 output and a regional endpoint at $2.20 and $11.00. The tables and cache math use the global rate, which matches the lower figure the model page reports. Mistral Large 3 lists two usage types per rate, both at the same dollar, so either resolves to the same number.

What to do

• Offline scoring, classification, enrichment, synthetic data: batch. It’s half price and nobody is waiting. Confirm the model has a batch row first; 76 of 149 do.
• Scheduled jobs with a loose latency budget: flex. Same 50% discount as batch, and 7 of the 49 flex models also price flex cache reads.
• Interactive chat and agents with a stable system prompt: standard plus caching. The 5-minute write breaks even at two uses, so any prompt that appears twice inside five minutes should be cached.
• Latency-critical endpoints: priority, if the model offers it, which rules out every Anthropic model in our data. Nova Pro Latency Optimized is the fallback at $1.00 and $4.00, 25% over standard against priority’s 75%.

The tier decides which models are candidates at all, so filter on it before you compare prices. Model pages on the AI explorer list input, output, cache and batch rates side by side, which is the quickest way to see whether a model can take the discount you’ve budgeted.

Bottom line: the service tier is a 0.5x to 1.75x multiplier on every token, and most cache reads bill at a tenth of the input rate. Neither is available everywhere. 72 of 149 models in us-east-1 are standard-only, and 33 of 34 cache-capable models can’t cache inside a batch job. Price the tier before you price the model.