What was confirmed, and what was not
AWS's weekly roundup of 9 November 2026 announces Web Answer on Amazon Bedrock AgentCore. It describes a single tool call that returns a synthesized, cited answer to a natural-language question, in place of a search, retrieval and citation pipeline you build yourself. It says Web Answer plans its own searches and runs several in parallel, and it calls Web Answer a tool you attach to a gateway you already run 1. The excerpt I could read cut off mid-sentence while describing how Web Answer decides it has enough evidence.
Three things were not visible in the pages I could read, and this article does not guess them:
- The Web Answer rate and its billing unit (per answer, per search it runs, per token, or a mix).
- Whether the fee includes the model that writes the answer, or whether those tokens bill separately.
- Whether the status is general availability. The announcement is real; the GA wording was not on a page I read.
Bedrock's pricing page lists Web Search as its own tab beside Guardrails, Knowledge Bases and Model Evaluation, but the rate did not appear in the text returned 2. INFRO's earlier tool-cost article, which relays AWS's pages, records that Bedrock Web Search is charged per query and that its observations count as input tokens 3. Treat that as second-hand until you have the rate on screen.
So this is a comparison of cost structure, with a template where the unknowns are inputs. Fill them from AWS's pricing page before anyone signs off a budget.
Two shapes of call
The difference is who does the loop. With Web Search your agent's model decides to search, receives observations, reads them, and may search again. With Web Answer that loop runs inside the tool, and your model receives text with citations.
| Web Search | Web Answer | |
|---|---|---|
| What your agent sends | One query per call | A natural-language question |
| Who plans the searches | Your model | The tool 1 |
| What comes back | Observations (5, 11 or 25 per search, set by search_context_size) 3 | A synthesized, cited answer 1 |
| Fee unit | Per query 3 | Not confirmed |
| Tokens your model pays for | Observations as input, on this turn and every later turn they stay in context | The answer text, plus whatever your model writes around it |
| Native cap on call count | AWS says the context-size setting does not limit how many searches the model makes 3 | Not confirmed |
The practical consequence: Web Search cost scales with your agent's behaviour, which you can only bound with your own guards. Web Answer cost scales with how many questions you ask, which is easier to forecast if its unit is per answer, and harder if the unit hides internal searches.
Pricing units: what to find on the page
Before you build a model, write down the unit for each line. AgentCore itself is consumption-priced with no upfront commitments or minimum fees, and each feature can be used on its own or together 4. That tells you the charges stack by feature, so a web answer can involve several lines.
Lines to look for, in this order:
- The tool fee: per query for Web Search, and whatever Web Answer uses.
- Gateway charges, since Web Answer attaches to an AgentCore Gateway. The FAQ describes Gateway as the way agents reach tools, by turning APIs and Lambda functions into agent-compatible tools and connecting existing MCP servers 5. Check whether invoking a tool through Gateway has its own unit; I did not see it in the text returned.
- Runtime time. Runtime on microVMs bills per second on consumed CPU and memory, with a one-second minimum, and CPU scales to zero during I/O wait such as waiting for LLM responses or tool calls 4. A longer web wait is therefore mostly a memory cost, not a CPU one.
- Model tokens for your agent's own turns.
- Logging and data transfer. Network data transfer is billed at standard EC2 rates for Runtime 4.
AgentCore Browser is a different tool, billed on CPU and memory over the session 4. It is for interacting with sites, not for answering a question, so keep it out of this comparison unless you use it too.
Model calls and tokens behind each request
For Web Search, count model calls in your own loop. A question that needs three searches is at least four model calls: one to issue each search and one to answer, and each call re-sends the context, including earlier observations. INFRO's earlier article, citing Anthropic and OpenAI's pricing pages, makes the same point for other providers: search results are billed as input tokens and re-read on later turns 3.
For Web Answer, the searches and the synthesis happen inside the tool, and your model sees the result. Two cases matter for the forecast:
- If the Web Answer fee covers the whole thing, your cost is the fee plus the tokens for one question going out and one cited answer coming back in.
- If the fee covers search only, add the synthesis model's tokens as a second line.
Which case applies is a reading task, and it can change the answer to this whole comparison. Ask AWS or read the pricing page for the "what's included" text, and record the date you read it.
One further effect cuts in Web Answer's favour on tokens: raw observations never enter your context, so they cannot be re-read on later turns. It can cut the other way on quality and control: you see less of what was retrieved, so you have less to inspect when an answer is wrong.
Cost per answered question
The unit finance should track is cost per answered question: everything the agent spends on web access in a run, divided by the number of questions it actually answered. Questions that return nothing useful still cost money, so divide by answered, not asked.
Notice that in the Web Search path the tokens are the largest line, while the per-query fee is smaller. That matches INFRO's earlier worked example, where search results cost three times their own fee in input tokens 3.
A forecast template
Use one function for both paths so finance can change a single input and see the effect. Fill rates from AWS's pricing page and your model's rate, never from this article.
from dataclasses import dataclass
@dataclass
class Rates:
search_per_query: float # Web Search fee, from AWS pricing page
answer_per_call: float # Web Answer fee, from AWS pricing page
synth_in_per_m: float # 0 if the Web Answer fee includes synthesis
synth_out_per_m: float
model_in_per_m: float # your agent's own model
model_out_per_m: float
def search_path(r: Rates, queries, calls, in_tokens, out_tokens):
return (queries * r.search_per_query
+ in_tokens / 1e6 * r.model_in_per_m
+ out_tokens / 1e6 * r.model_out_per_m)
def answer_path(r: Rates, in_tokens, out_tokens, synth_in=0, synth_out=0):
return (r.answer_per_call
+ synth_in / 1e6 * r.synth_in_per_m
+ synth_out / 1e6 * r.synth_out_per_m
+ in_tokens / 1e6 * r.model_in_per_m
+ out_tokens / 1e6 * r.model_out_per_m)
def per_answered(total_cost, asked, answered_rate):
return total_cost * asked / (asked * answered_rate)Then forecast monthly spend with three inputs per workload: questions per month, share answered, and the average per-question cost from your measured runs. Keep a range, not a point: run each path on a sample of real questions, record the p50 and p90 of tokens and tool calls, and budget on the p90 for the first month. Add Runtime seconds and Gateway lines separately once you have their units.
Controls and measurement
Whichever path you pick, cap it. For Web Search the cap has to be yours, since AWS notes the context-size setting does not limit the number of searches the model makes 3. A step limit and a run-level cost limit in your loop are the controls; INFRO's article on bounding agent loops covers the pattern in /blog/bound-ai-agent-loops.
For Web Answer, check whether the call has a limit on the searches it runs, and whether it exposes a per-call count. If it does not, your control is the number of Web Answer calls per run and per user, so enforce that in code.
For measurement, AWS records Web Search activity as CloudTrail data events when that logging is on 3, which gives you a call count independent of your own logs. Compare it with the billed count at month end. Scope matters too: a domain filter limits where results come from, not how many searches run, so it does not cap cost 6.
If your agent's own model calls go through INFRO, per-request usage and estimated provider cost are recorded, and budgets and attribution can cap and group them by project or key; see the INFRO docs (opens in a new tab) 7. The web tool fee is billed by AWS on your AWS account, so it appears in your AWS cost data, and you join the two on your own run identifier.
Which to pick
Choose Web Answer when the task is "answer this question with sources" and you want a predictable per-question cost, and once its fee and inclusions are confirmed. Choose Web Search when your agent needs to read specific results, apply its own ranking, or combine search with other tools in one reasoning loop. In that case budget the tokens and add a loop guard.
Either way, run both on 50 to 100 real questions, record cost per answered question and how often a reviewer accepts the answer, and decide from that. The rate cards matter less than the token and call counts your own workload produces.
