diff --git a/Cargo.lock b/Cargo.lock
index f05d75f..d5b0b9a 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -693,7 +693,7 @@ checksum = "460fbee9c2c2f33933d720630a6a0bac33ba7053db5344fac858d4b8952d77d5"
 
 [[package]]
 name = "crw-browse"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -723,7 +723,7 @@ dependencies = [
 
 [[package]]
 name = "crw-cli"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "anyhow",
  "axum",
@@ -757,7 +757,7 @@ dependencies = [
 
 [[package]]
 name = "crw-core"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "config",
  "crw-mcp-proto",
@@ -775,7 +775,7 @@ dependencies = [
 
 [[package]]
 name = "crw-crawl"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "crw-core",
  "crw-diff",
@@ -800,7 +800,7 @@ dependencies = [
 
 [[package]]
 name = "crw-diff"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "crw-core",
  "hex",
@@ -815,7 +815,7 @@ dependencies = [
 
 [[package]]
 name = "crw-extract"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "crw-core",
  "ego-tree",
@@ -844,7 +844,7 @@ dependencies = [
 
 [[package]]
 name = "crw-mcp"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "clap",
  "crw-core",
@@ -859,7 +859,7 @@ dependencies = [
 
 [[package]]
 name = "crw-mcp-proto"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "serde",
  "serde_json",
@@ -868,7 +868,7 @@ dependencies = [
 
 [[package]]
 name = "crw-monitor"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "crw-core",
  "crw-crawl",
@@ -891,7 +891,7 @@ dependencies = [
 
 [[package]]
 name = "crw-renderer"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "async-trait",
  "axum",
@@ -918,7 +918,7 @@ dependencies = [
 
 [[package]]
 name = "crw-search"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "crw-core",
  "futures",
@@ -934,7 +934,7 @@ dependencies = [
 
 [[package]]
 name = "crw-server"
-version = "0.10.0"
+version = "0.11.0"
 dependencies = [
  "axum",
  "axum-test",
diff --git a/Cargo.toml b/Cargo.toml
index 9ce2f1e..2c0aba1 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -16,7 +16,7 @@ members = [
 resolver = "2"
 
 [workspace.package]
-version = "0.10.0"
+version = "0.11.0"
 edition = "2024"
 license = "AGPL-3.0"
 repository = "https://github.com/us/crw"
diff --git a/crates/crw-browse/Cargo.toml b/crates/crw-browse/Cargo.toml
index 880bed5..b7406ea 100644
--- a/crates/crw-browse/Cargo.toml
+++ b/crates/crw-browse/Cargo.toml
@@ -11,8 +11,8 @@ description = "MCP server for interactive browser automation over CDP"
 publish = false
 
 [dependencies]
-crw-core = { path = "../crw-core", version = "0.10.0" }
-crw-renderer = { path = "../crw-renderer", version = "0.10.0", features = ["cdp"] }
+crw-core = { path = "../crw-core", version = "0.11.0" }
+crw-renderer = { path = "../crw-renderer", version = "0.11.0", features = ["cdp"] }
 
 rmcp = { version = "1.5", features = ["server", "macros", "transport-io"] }
 clap = { workspace = true }
diff --git a/crates/crw-cli/Cargo.toml b/crates/crw-cli/Cargo.toml
index b452ddb..44039bf 100644
--- a/crates/crw-cli/Cargo.toml
+++ b/crates/crw-cli/Cargo.toml
@@ -28,17 +28,17 @@ browse = ["dep:crw-browse", "dep:rmcp", "dep:anyhow"]
 
 [dependencies]
 # Core crates
-crw-core = { path = "../crw-core", version = "0.10.0" }
-crw-renderer = { path = "../crw-renderer", version = "0.10.0", features = ["auto-browser", "cdp"] }
-crw-extract = { path = "../crw-extract", version = "0.10.0" }
-crw-crawl = { path = "../crw-crawl", version = "0.10.0" }
-crw-search = { path = "../crw-search", version = "0.10.0" }
+crw-core = { path = "../crw-core", version = "0.11.0" }
+crw-renderer = { path = "../crw-renderer", version = "0.11.0", features = ["auto-browser", "cdp"] }
+crw-extract = { path = "../crw-extract", version = "0.11.0" }
+crw-crawl = { path = "../crw-crawl", version = "0.11.0" }
+crw-search = { path = "../crw-search", version = "0.11.0" }
 
 # Server (for serve + mcp-embedded + setup commands)
-crw-server = { path = "../crw-server", version = "0.10.0", optional = true, features = ["cdp"] }
+crw-server = { path = "../crw-server", version = "0.11.0", optional = true, features = ["cdp"] }
 
 # Browse (for browse command)
-crw-browse = { path = "../crw-browse", version = "0.10.0", optional = true }
+crw-browse = { path = "../crw-browse", version = "0.11.0", optional = true }
 
 # Web framework (for serve command)
 axum = { workspace = true, optional = true }
diff --git a/crates/crw-core/Cargo.toml b/crates/crw-core/Cargo.toml
index 13ead64..8d709d9 100644
--- a/crates/crw-core/Cargo.toml
+++ b/crates/crw-core/Cargo.toml
@@ -19,7 +19,7 @@ description = "Core types, config, and error handling for the CRW web scraper"
 cdp = []
 
 [dependencies]
-crw-mcp-proto = { path = "../crw-mcp-proto", version = "0.10.0" }
+crw-mcp-proto = { path = "../crw-mcp-proto", version = "0.11.0" }
 serde = { workspace = true }
 serde_json = { workspace = true }
 toml = { workspace = true }
diff --git a/crates/crw-core/src/types.rs b/crates/crw-core/src/types.rs
index ca850eb..5012eff 100644
--- a/crates/crw-core/src/types.rs
+++ b/crates/crw-core/src/types.rs
@@ -322,6 +322,24 @@ pub struct LlmUsage {
     pub truncated: bool,
     #[serde(default = "one_u32", skip_serializing_if = "is_one_u32")]
     pub calls: u32,
+
+    // ── Wave 4 (R1) additions: SaaS billing correlation across legs ──
+    //
+    // The SaaS-side managed pricing path needs to know exactly how many
+    // summary calls executed AND whether the answer leg ran. The 5-branch
+    // fail-closed dispatch keys off these counters:
+    //   - executedSummaries > 0 OR answerExecuted ⇒ engine did work
+    //   - inputTokens == 0 AND outputTokens == 0 ⇒ no upstream cost
+    // Without the counters the SaaS cannot disambiguate "no work" from
+    // "work but missing telemetry" and would refund or charge wrong.
+    //
+    // Always serialized (no skip_serializing_if) so the always-present
+    // R1 invariant holds: when /v1/search returns llmUsage, both fields
+    // are explicitly visible.
+    #[serde(default)]
+    pub executed_summaries: u32,
+    #[serde(default)]
+    pub answer_executed: bool,
 }
 
 fn one_u32() -> u32 {
diff --git a/crates/crw-crawl/Cargo.toml b/crates/crw-crawl/Cargo.toml
index 76eeb63..3844ae0 100644
--- a/crates/crw-crawl/Cargo.toml
+++ b/crates/crw-crawl/Cargo.toml
@@ -10,10 +10,10 @@ categories.workspace = true
 description = "Async BFS web crawler with rate limiting and robots.txt support for CRW"
 
 [dependencies]
-crw-core = { path = "../crw-core", version = "0.10.0" }
-crw-diff = { path = "../crw-diff", version = "0.10.0" }
-crw-renderer = { path = "../crw-renderer", version = "0.10.0" }
-crw-extract = { path = "../crw-extract", version = "0.10.0" }
+crw-core = { path = "../crw-core", version = "0.11.0" }
+crw-diff = { path = "../crw-diff", version = "0.11.0" }
+crw-renderer = { path = "../crw-renderer", version = "0.11.0" }
+crw-extract = { path = "../crw-extract", version = "0.11.0" }
 reqwest = { workspace = true }
 serde_json = { workspace = true }
 scraper = { workspace = true }
diff --git a/crates/crw-diff/Cargo.toml b/crates/crw-diff/Cargo.toml
index 2bf4df5..6d1fe02 100644
--- a/crates/crw-diff/Cargo.toml
+++ b/crates/crw-diff/Cargo.toml
@@ -13,7 +13,7 @@ description = "Stateless change-tracking diff engine for the CRW web scraper"
 # Shared types only (ChangeTrackingOptions/Result, DiffAst, etc.). This crate
 # MUST NOT depend on crw-extract — judging is injected upstream so the diff
 # engine stays pure (no LLM, no HTTP, no I/O).
-crw-core = { path = "../crw-core", version = "0.10.0" }
+crw-core = { path = "../crw-core", version = "0.11.0" }
 serde = { workspace = true }
 serde_json = { workspace = true }
 similar = { workspace = true }
diff --git a/crates/crw-extract/Cargo.toml b/crates/crw-extract/Cargo.toml
index 2765ec3..45f3570 100644
--- a/crates/crw-extract/Cargo.toml
+++ b/crates/crw-extract/Cargo.toml
@@ -10,7 +10,7 @@ categories.workspace = true
 description = "HTML extraction and markdown conversion engine for the CRW web scraper"
 
 [dependencies]
-crw-core = { path = "../crw-core", version = "0.10.0" }
+crw-core = { path = "../crw-core", version = "0.11.0" }
 lol_html = { workspace = true }
 scraper = { workspace = true }
 htmd = { workspace = true }
diff --git a/crates/crw-extract/src/llm.rs b/crates/crw-extract/src/llm.rs
index 0176850..0fe5988 100644
--- a/crates/crw-extract/src/llm.rs
+++ b/crates/crw-extract/src/llm.rs
@@ -418,6 +418,11 @@ fn parse_anthropic_usage(payload: &serde_json::Value, model: &str) -> Option<Llm
         cache_miss_input_tokens,
         truncated: false,
         calls: 1,
+        // R1 counters are scoped to /v1/search aggregation; single-call
+        // sites always emit defaults. Aggregation happens in the caller
+        // (crw-server::routes::search::search_inner).
+        executed_summaries: 0,
+        answer_executed: false,
     })
 }
 
@@ -474,6 +479,11 @@ fn parse_openai_usage(
         cache_miss_input_tokens,
         truncated: false,
         calls: 1,
+        // R1 counters are scoped to /v1/search aggregation; single-call
+        // sites always emit defaults. Aggregation happens in the caller
+        // (crw-server::routes::search::search_inner).
+        executed_summaries: 0,
+        answer_executed: false,
     })
 }
 
diff --git a/crates/crw-extract/src/structured.rs b/crates/crw-extract/src/structured.rs
index 6381dbb..c263d63 100644
--- a/crates/crw-extract/src/structured.rs
+++ b/crates/crw-extract/src/structured.rs
@@ -295,6 +295,10 @@ pub(crate) async fn call_anthropic(
             cache_miss_input_tokens: cache_miss,
             truncated: false,
             calls: 1,
+            // R1 counters are aggregated in the /v1/search caller;
+            // single-call sites always emit defaults.
+            executed_summaries: 0,
+            answer_executed: false,
         }
     });
 
@@ -498,6 +502,10 @@ pub(crate) async fn call_openai(
             cache_miss_input_tokens: cache_miss,
             truncated: false,
             calls: 1,
+            // R1 counters are aggregated in the /v1/search caller;
+            // single-call sites always emit defaults.
+            executed_summaries: 0,
+            answer_executed: false,
         }
     });
 
diff --git a/crates/crw-mcp/Cargo.toml b/crates/crw-mcp/Cargo.toml
index 329b785..e8e67de 100644
--- a/crates/crw-mcp/Cargo.toml
+++ b/crates/crw-mcp/Cargo.toml
@@ -18,8 +18,8 @@ default = ["embedded"]
 embedded = ["dep:crw-server"]
 
 [dependencies]
-crw-core = { path = "../crw-core", version = "0.10.0" }
-crw-renderer = { path = "../crw-renderer", version = "0.10.0", features = ["auto-browser"] }
+crw-core = { path = "../crw-core", version = "0.11.0" }
+crw-renderer = { path = "../crw-renderer", version = "0.11.0", features = ["auto-browser"] }
 serde_json = { workspace = true }
 tokio = { workspace = true }
 reqwest = { workspace = true }
@@ -28,4 +28,4 @@ tracing-subscriber = { workspace = true }
 clap = { workspace = true }
 
 # Embedded mode: pulls in full scraping engine
-crw-server = { path = "../crw-server", version = "0.10.0", optional = true, features = ["cdp"] }
+crw-server = { path = "../crw-server", version = "0.11.0", optional = true, features = ["cdp"] }
diff --git a/crates/crw-monitor/Cargo.toml b/crates/crw-monitor/Cargo.toml
index d7b63c9..57c9acc 100644
--- a/crates/crw-monitor/Cargo.toml
+++ b/crates/crw-monitor/Cargo.toml
@@ -14,11 +14,11 @@ description = "Optional self-host monitor mode for the CRW web scraper (SQLite-b
 # `crw-crawl` provides the scrape/crawl primitives; `crw-extract` provides the
 # LLM judge. None of these pull a DB dependency — the SQLite/cron/hmac stack is
 # local to this crate and feature-gated, never compiled into the default server.
-crw-core = { path = "../crw-core", version = "0.10.0" }
-crw-diff = { path = "../crw-diff", version = "0.10.0" }
-crw-crawl = { path = "../crw-crawl", version = "0.10.0" }
-crw-extract = { path = "../crw-extract", version = "0.10.0" }
-crw-renderer = { path = "../crw-renderer", version = "0.10.0" }
+crw-core = { path = "../crw-core", version = "0.11.0" }
+crw-diff = { path = "../crw-diff", version = "0.11.0" }
+crw-crawl = { path = "../crw-crawl", version = "0.11.0" }
+crw-extract = { path = "../crw-extract", version = "0.11.0" }
+crw-renderer = { path = "../crw-renderer", version = "0.11.0" }
 
 serde = { workspace = true }
 serde_json = { workspace = true }
diff --git a/crates/crw-renderer/Cargo.toml b/crates/crw-renderer/Cargo.toml
index e366859..bc0c21c 100644
--- a/crates/crw-renderer/Cargo.toml
+++ b/crates/crw-renderer/Cargo.toml
@@ -15,8 +15,8 @@ cdp = ["tokio-tungstenite", "crw-core/cdp"]
 auto-browser = ["dep:dirs"]
 
 [dependencies]
-crw-core = { path = "../crw-core", version = "0.10.0" }
-crw-extract = { path = "../crw-extract", version = "0.10.0" }
+crw-core = { path = "../crw-core", version = "0.11.0" }
+crw-extract = { path = "../crw-extract", version = "0.11.0" }
 reqwest = { workspace = true }
 tokio = { workspace = true }
 tokio-tungstenite = { version = "0.28", features = ["rustls-tls-native-roots"], optional = true }
diff --git a/crates/crw-search/Cargo.toml b/crates/crw-search/Cargo.toml
index c2b8194..5baee0d 100644
--- a/crates/crw-search/Cargo.toml
+++ b/crates/crw-search/Cargo.toml
@@ -10,7 +10,7 @@ categories.workspace = true
 description = "SearXNG-backed search client and result transforms for the CRW web scraper"
 
 [dependencies]
-crw-core = { path = "../crw-core", version = "0.10.0" }
+crw-core = { path = "../crw-core", version = "0.11.0" }
 futures = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
diff --git a/crates/crw-server/Cargo.toml b/crates/crw-server/Cargo.toml
index 50879d3..a997028 100644
--- a/crates/crw-server/Cargo.toml
+++ b/crates/crw-server/Cargo.toml
@@ -19,14 +19,14 @@ test-utils = []
 monitor = ["dep:crw-monitor"]
 
 [dependencies]
-crw-core = { path = "../crw-core", version = "0.10.0" }
-crw-diff = { path = "../crw-diff", version = "0.10.0" }
-crw-renderer = { path = "../crw-renderer", version = "0.10.0" }
-crw-extract = { path = "../crw-extract", version = "0.10.0" }
-crw-crawl = { path = "../crw-crawl", version = "0.10.0" }
-crw-search = { path = "../crw-search", version = "0.10.0" }
+crw-core = { path = "../crw-core", version = "0.11.0" }
+crw-diff = { path = "../crw-diff", version = "0.11.0" }
+crw-renderer = { path = "../crw-renderer", version = "0.11.0" }
+crw-extract = { path = "../crw-extract", version = "0.11.0" }
+crw-crawl = { path = "../crw-crawl", version = "0.11.0" }
+crw-search = { path = "../crw-search", version = "0.11.0" }
 # Optional self-host monitor mode (default OFF — see the `monitor` feature).
-crw-monitor = { path = "../crw-monitor", version = "0.10.0", optional = true }
+crw-monitor = { path = "../crw-monitor", version = "0.11.0", optional = true }
 axum = { workspace = true }
 tower = { workspace = true }
 tower-http = { workspace = true }
diff --git a/crates/crw-server/src/routes/search.rs b/crates/crw-server/src/routes/search.rs
index 6a45bad..c66aa07 100644
--- a/crates/crw-server/src/routes/search.rs
+++ b/crates/crw-server/src/routes/search.rs
@@ -5,7 +5,7 @@ use crw_core::Deadline;
 use crw_core::config::LlmConfig;
 use crw_core::error::CrwError;
 use crw_core::types::{
-    ApiResponse, OutputFormat, ScrapeData, ScrapeRequest, SearchData, SearchRequest,
+    ApiResponse, LlmUsage, OutputFormat, ScrapeData, ScrapeRequest, SearchData, SearchRequest,
     SearchResponse, SearchResponseData, SearchResult, SearchScrapeOptions,
 };
 use crw_crawl::single::scrape_url;
@@ -21,6 +21,14 @@ const DEFAULT_ANSWER_TOP_N: u32 = 5;
 const MAX_ANSWER_TOP_N: u32 = 10;
 const DEFAULT_MAX_CHARS_PER_SOURCE: usize = 8192;
 
+/// Wave 4 (R2): hard cap on `max_tokens` per LLM leg (one summary call OR
+/// the answer call). Independent of the user's configured `cfg.max_tokens`
+/// because the SaaS-side `estimateMaxCreditCostForSearch` uses this number
+/// to pre-reserve credits; a per-leg cap higher than this would let real
+/// usage exceed the reservation. Mirror in
+/// `crw-saas/src/lib/llm-pricing.ts::legCost` (default 1024).
+const SEARCH_LLM_MAX_TOKENS_PER_LEG: u32 = 1024;
+
 use crate::error::AppError;
 use crate::state::AppState;
 
@@ -102,6 +110,43 @@ pub async fn search_inner(
 
     let wants_summaries = req.summarize_results.unwrap_or(false);
     let wants_answer = req.answer.unwrap_or(false);
+    // Wave 4 (R1): once we enter LLM mode the response MUST carry a
+    // non-null llmUsage object (the always-present invariant the SaaS
+    // 5-branch dispatch relies on). We aggregate summary + answer counts
+    // into this builder and emit it at every return path below.
+    let mut llm_attempted = false;
+    let mut agg_input_tokens: u32 = 0;
+    let mut agg_output_tokens: u32 = 0;
+    let mut agg_cache_hit: u32 = 0;
+    let mut agg_cache_miss: u32 = 0;
+    let mut agg_calls: u32 = 0;
+    let mut agg_executed_summaries: u32 = 0;
+    let mut agg_answer_executed = false;
+    let mut agg_provider: String = String::new();
+    let mut agg_model: String = String::new();
+    let mut agg_truncated = false;
+    let merge_usage = |agg_input_tokens: &mut u32,
+                       agg_output_tokens: &mut u32,
+                       agg_cache_hit: &mut u32,
+                       agg_cache_miss: &mut u32,
+                       agg_calls: &mut u32,
+                       agg_provider: &mut String,
+                       agg_model: &mut String,
+                       agg_truncated: &mut bool,
+                       u: &LlmUsage| {
+        *agg_input_tokens = agg_input_tokens.saturating_add(u.input_tokens);
+        *agg_output_tokens = agg_output_tokens.saturating_add(u.output_tokens);
+        *agg_cache_hit = agg_cache_hit.saturating_add(u.cache_hit_input_tokens.unwrap_or(0));
+        *agg_cache_miss = agg_cache_miss.saturating_add(u.cache_miss_input_tokens.unwrap_or(0));
+        *agg_calls = agg_calls.saturating_add(u.calls.max(1));
+        if agg_provider.is_empty() {
+            *agg_provider = u.provider.clone();
+        }
+        if agg_model.is_empty() {
+            *agg_model = u.model.clone();
+        }
+        *agg_truncated = *agg_truncated || u.truncated;
+    };
 
     if (wants_summaries || wants_answer) && req.scrape_options.is_none() {
         warnings.push(
@@ -115,15 +160,35 @@ pub async fn search_inner(
                     .into(),
             ),
             Some(llm) => {
+                llm_attempted = true;
+                // Wave 4 (R2): cap max_tokens at SEARCH_LLM_MAX_TOKENS_PER_LEG so
+                // a single leg can never exceed the SaaS pre-reservation in
+                // estimateMaxCreditCostForSearch.
+                let mut leg_cfg = llm.clone();
+                leg_cfg.max_tokens = leg_cfg.max_tokens.min(SEARCH_LLM_MAX_TOKENS_PER_LEG);
                 if wants_summaries {
-                    let count = attach_result_summaries(
+                    let (count, usages) = attach_result_summaries(
                         &mut data,
-                        llm,
-                        llm.max_concurrency,
+                        &leg_cfg,
+                        leg_cfg.max_concurrency,
                         req.summary_prompt.as_deref(),
                         req.max_content_chars,
                     )
                     .await;
+                    agg_executed_summaries = count.ok as u32;
+                    for u in usages.into_iter().flatten() {
+                        merge_usage(
+                            &mut agg_input_tokens,
+                            &mut agg_output_tokens,
+                            &mut agg_cache_hit,
+                            &mut agg_cache_miss,
+                            &mut agg_calls,
+                            &mut agg_provider,
+                            &mut agg_model,
+                            &mut agg_truncated,
+                            &u,
+                        );
+                    }
                     if count.failed > 0 {
                         warnings.push(format!(
                             "{} of {} per-result summaries failed",
@@ -133,14 +198,41 @@ pub async fn search_inner(
                     }
                 }
                 if wants_answer {
-                    match synthesize_answer(&req, &data, llm).await {
-                        Ok((ans, cites, usage, mut ans_warns)) => {
+                    match synthesize_answer(&req, &data, &leg_cfg).await {
+                        Ok((ans, cites, ans_usage, mut ans_warns)) => {
                             warnings.append(&mut ans_warns);
+                            agg_answer_executed = true;
+                            if let Some(ref u) = ans_usage {
+                                merge_usage(
+                                    &mut agg_input_tokens,
+                                    &mut agg_output_tokens,
+                                    &mut agg_cache_hit,
+                                    &mut agg_cache_miss,
+                                    &mut agg_calls,
+                                    &mut agg_provider,
+                                    &mut agg_model,
+                                    &mut agg_truncated,
+                                    u,
+                                );
+                            }
+                            let aggregated = build_aggregated_usage(
+                                agg_input_tokens,
+                                agg_output_tokens,
+                                agg_cache_hit,
+                                agg_cache_miss,
+                                agg_calls,
+                                agg_executed_summaries,
+                                agg_answer_executed,
+                                agg_provider.clone(),
+                                agg_model.clone(),
+                                agg_truncated,
+                                &leg_cfg,
+                            );
                             let wrapped = SearchResponseData {
                                 results: data,
                                 answer: Some(ans),
                                 citations: cites,
-                                llm_usage: usage,
+                                llm_usage: Some(aggregated),
                                 warnings,
                             };
                             let mut resp = ApiResponse::ok(wrapped);
@@ -157,11 +249,51 @@ pub async fn search_inner(
         }
     }
 
+    // R1 always-present invariant: if we attempted LLM work, emit the
+    // aggregated usage even when zero tokens were consumed (e.g. all
+    // summaries failed and no answer leg ran). The SaaS dispatch maps
+    // (executedSummaries == 0 && answerExecuted == false && tokens == 0)
+    // to Branch 1 (no-op refund); anything else routes correctly.
+    let final_usage = if llm_attempted {
+        Some(build_aggregated_usage(
+            agg_input_tokens,
+            agg_output_tokens,
+            agg_cache_hit,
+            agg_cache_miss,
+            agg_calls,
+            agg_executed_summaries,
+            agg_answer_executed,
+            if agg_provider.is_empty() {
+                effective_llm
+                    .map(|c| c.provider.clone())
+                    .unwrap_or_default()
+            } else {
+                agg_provider
+            },
+            if agg_model.is_empty() {
+                effective_llm.map(|c| c.model.clone()).unwrap_or_default()
+            } else {
+                agg_model
+            },
+            agg_truncated,
+            effective_llm
+                .map(|c| {
+                    let mut c = c.clone();
+                    c.max_tokens = c.max_tokens.min(SEARCH_LLM_MAX_TOKENS_PER_LEG);
+                    c
+                })
+                .as_ref()
+                .unwrap_or(&crw_core::config::LlmConfig::default()),
+        ))
+    } else {
+        None
+    };
+
     let wrapped = SearchResponseData {
         results: data,
         answer: None,
         citations: Vec::new(),
-        llm_usage: None,
+        llm_usage: final_usage,
         warnings,
     };
     let mut resp = ApiResponse::ok(wrapped);
@@ -169,6 +301,52 @@ pub async fn search_inner(
     Ok(resp)
 }
 
+#[allow(clippy::too_many_arguments)]
+fn build_aggregated_usage(
+    input_tokens: u32,
+    output_tokens: u32,
+    cache_hit: u32,
+    cache_miss: u32,
+    calls: u32,
+    executed_summaries: u32,
+    answer_executed: bool,
+    provider: String,
+    model: String,
+    truncated: bool,
+    fallback_cfg: &crw_core::config::LlmConfig,
+) -> LlmUsage {
+    LlmUsage {
+        input_tokens,
+        output_tokens,
+        total_tokens: input_tokens.saturating_add(output_tokens),
+        estimated_cost_usd: None,
+        model: if model.is_empty() {
+            fallback_cfg.model.clone()
+        } else {
+            model
+        },
+        provider: if provider.is_empty() {
+            fallback_cfg.provider.clone()
+        } else {
+            provider
+        },
+        cache_hit_input_tokens: if cache_hit == 0 {
+            None
+        } else {
+            Some(cache_hit)
+        },
+        cache_miss_input_tokens: if cache_miss == 0 {
+            None
+        } else {
+            Some(cache_miss)
+        },
+        truncated,
+        calls: calls.max(1),
+        executed_summaries,
+        answer_executed,
+    }
+}
+
 #[derive(Default)]
 struct SummaryFanoutCount {
     ok: usize,
@@ -177,18 +355,22 @@ struct SummaryFanoutCount {
 
 /// Fan-out summary calls across all results that have markdown. Bounded by
 /// `max_concurrency`. Pattern mirrors `crates/crw-crawl/src/sitemap.rs`.
+///
+/// Wave 4 (R1): returns the per-call `Option<LlmUsage>` for every job
+/// alongside the ok/failed count so the caller can aggregate token totals
+/// across summaries + answer.
 async fn attach_result_summaries(
     data: &mut SearchData,
     cfg: &LlmConfig,
     max_concurrency: usize,
     user_prompt: Option<&str>,
     max_content_chars: Option<usize>,
-) -> SummaryFanoutCount {
+) -> (SummaryFanoutCount, Vec<Option<LlmUsage>>) {
     let targets: &mut Vec<SearchResult> = match data {
         SearchData::Flat(v) => v,
         SearchData::Grouped(g) => match g.web.as_mut() {
             Some(v) if !v.is_empty() => v,
-            _ => return SummaryFanoutCount::default(),
+            _ => return (SummaryFanoutCount::default(), Vec::new()),
         },
     };
     // Capture markdown + index pairs first so we don't hold a borrow of
@@ -199,12 +381,13 @@ async fn attach_result_summaries(
         .filter_map(|(idx, r)| r.markdown.as_ref().map(|md| (idx, md.clone())))
         .collect();
     if jobs.is_empty() {
-        return SummaryFanoutCount::default();
+        return (SummaryFanoutCount::default(), Vec::new());
     }
     let cfg_owned = cfg.clone();
     let user_prompt_owned: Option<String> = user_prompt.map(str::to_owned);
     let concurrency = max_concurrency.max(1);
-    let results: Vec<(usize, Result<String, String>)> = stream::iter(jobs)
+    type SummaryOutcome = (usize, Result<(String, Option<LlmUsage>), String>);
+    let results: Vec<SummaryOutcome> = stream::iter(jobs)
         .map(|(idx, md)| {
             let cfg = cfg_owned.clone();
             let user_prompt = user_prompt_owned.clone();
@@ -212,7 +395,7 @@ async fn attach_result_summaries(
                 let outcome =
                     summary::summarize(&md, &cfg, user_prompt.as_deref(), max_content_chars)
                         .await
-                        .map(|r| r.content)
+                        .map(|r| (r.content, r.usage))
                         .map_err(|e| e.to_string());
                 (idx, outcome)
             }
@@ -222,18 +405,23 @@ async fn attach_result_summaries(
         .await;
 
     let mut count = SummaryFanoutCount::default();
+    let mut usages: Vec<Option<LlmUsage>> = Vec::with_capacity(results.len());
     for (idx, res) in results {
         match res {
-            Ok(text) => {
+            Ok((text, usage)) => {
                 if let Some(slot) = targets.get_mut(idx) {
                     slot.summary = Some(text);
                     count.ok += 1;
+                    usages.push(usage);
                 }
             }
-            Err(_) => count.failed += 1,
+            Err(_) => {
+                count.failed += 1;
+                usages.push(None);
+            }
         }
     }
-    count
+    (count, usages)
 }
 
 async fn synthesize_answer(
diff --git a/crates/crw-server/tests/search_route.rs b/crates/crw-server/tests/search_route.rs
index eaf14ca..53f467d 100644
--- a/crates/crw-server/tests/search_route.rs
+++ b/crates/crw-server/tests/search_route.rs
@@ -28,6 +28,29 @@ timeout_ms = 5000
     TestServer::new(app)
 }
 
+fn test_app_with_searxng_and_llm(url: &str) -> TestServer {
+    // Sibling of `test_app_with_searxng` that also wires an `[extraction.llm]`
+    // leg so provider/model can fall back to server config. `api_key` has no
+    // serde default (required field) — must be present for the block to parse.
+    let toml = format!(
+        r#"
+[search]
+enabled = true
+searxng_url = "{url}"
+timeout_ms = 5000
+
+[extraction.llm]
+provider = "anthropic"
+api_key = "sk-test"
+model = "claude-sonnet-4-20250514"
+"#
+    );
+    let config: AppConfig = toml::from_str(&toml).unwrap();
+    let state = AppState::new(config).expect("AppState::new failed");
+    let app = create_app(state);
+    TestServer::new(app)
+}
+
 fn test_app_search_disabled() -> TestServer {
     // Default config has no searxng_url → state.searxng = None.
     let config: AppConfig = toml::from_str("").unwrap();
@@ -101,6 +124,53 @@ fn searxng_mixed_response() -> Value {
     })
 }
 
+#[tokio::test]
+async fn search_llm_usage_always_present_on_zero_results() {
+    // Wave 4 (R1) invariant: once LLM mode is entered (summarizeResults +
+    // scrapeOptions present), `/v1/search` ALWAYS returns a non-null
+    // `llmUsage` object — even with ZERO search results and no LLM call
+    // actually running. The SaaS 5-branch credit dispatch relies on this.
+    let mock = MockServer::start().await;
+    Mock::given(method("GET"))
+        .and(path("/search"))
+        .respond_with(
+            ResponseTemplate::new(200)
+                .set_body_json(json!({"results": [], "number_of_results": 0})),
+        )
+        .mount(&mock)
+        .await;
+
+    let server = test_app_with_searxng_and_llm(&mock.uri());
+    let resp = server
+        .post("/v1/search")
+        .json(&json!({
+            "query": "rust async",
+            "summarizeResults": true,
+            "scrapeOptions": {"formats": ["markdown"]}
+        }))
+        .await;
+    resp.assert_status_ok();
+    let body: Value = resp.json();
+    assert_eq!(body["success"], true);
+
+    let usage = &body["data"]["llmUsage"];
+    assert!(usage.is_object(), "llmUsage must be present, got: {usage}");
+    assert_eq!(usage["executedSummaries"], 0);
+    assert_eq!(usage["answerExecuted"], false);
+    assert_eq!(usage["inputTokens"], 0);
+    assert_eq!(usage["outputTokens"], 0);
+    assert!(
+        usage["provider"].is_string(),
+        "provider must be a string, got: {}",
+        usage["provider"]
+    );
+    assert!(
+        usage["model"].is_string(),
+        "model must be a string, got: {}",
+        usage["model"]
+    );
+}
+
 #[tokio::test]
 async fn search_returns_flat_results() {
     let mock = MockServer::start().await;
diff --git a/npm/crw-mcp-darwin-arm64/package.json b/npm/crw-mcp-darwin-arm64/package.json
index 75fdac0..c73a6c9 100644
--- a/npm/crw-mcp-darwin-arm64/package.json
+++ b/npm/crw-mcp-darwin-arm64/package.json
@@ -1,6 +1,6 @@
 {
   "name": "crw-mcp-darwin-arm64",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "CRW MCP server binary for darwin arm64",
   "license": "AGPL-3.0",
   "homepage": "https://github.com/us/crw",
diff --git a/npm/crw-mcp-darwin-x64/package.json b/npm/crw-mcp-darwin-x64/package.json
index 9dd7bce..0cf0796 100644
--- a/npm/crw-mcp-darwin-x64/package.json
+++ b/npm/crw-mcp-darwin-x64/package.json
@@ -1,6 +1,6 @@
 {
   "name": "crw-mcp-darwin-x64",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "CRW MCP server binary for darwin x64",
   "license": "AGPL-3.0",
   "homepage": "https://github.com/us/crw",
diff --git a/npm/crw-mcp-linux-arm64/package.json b/npm/crw-mcp-linux-arm64/package.json
index 7390e8c..4dc9306 100644
--- a/npm/crw-mcp-linux-arm64/package.json
+++ b/npm/crw-mcp-linux-arm64/package.json
@@ -1,6 +1,6 @@
 {
   "name": "crw-mcp-linux-arm64",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "CRW MCP server binary for linux arm64",
   "license": "AGPL-3.0",
   "homepage": "https://github.com/us/crw",
diff --git a/npm/crw-mcp-linux-x64/package.json b/npm/crw-mcp-linux-x64/package.json
index 5f5a837..1f77fb0 100644
--- a/npm/crw-mcp-linux-x64/package.json
+++ b/npm/crw-mcp-linux-x64/package.json
@@ -1,6 +1,6 @@
 {
   "name": "crw-mcp-linux-x64",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "CRW MCP server binary for linux x64",
   "license": "AGPL-3.0",
   "homepage": "https://github.com/us/crw",
diff --git a/npm/crw-mcp-win32-arm64/package.json b/npm/crw-mcp-win32-arm64/package.json
index d845765..b618dc0 100644
--- a/npm/crw-mcp-win32-arm64/package.json
+++ b/npm/crw-mcp-win32-arm64/package.json
@@ -1,6 +1,6 @@
 {
   "name": "crw-mcp-win32-arm64",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "CRW MCP server binary for win32 arm64",
   "license": "AGPL-3.0",
   "homepage": "https://github.com/us/crw",
diff --git a/npm/crw-mcp-win32-x64/package.json b/npm/crw-mcp-win32-x64/package.json
index 71346ba..62e80f5 100644
--- a/npm/crw-mcp-win32-x64/package.json
+++ b/npm/crw-mcp-win32-x64/package.json
@@ -1,6 +1,6 @@
 {
   "name": "crw-mcp-win32-x64",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "CRW MCP server binary for win32 x64",
   "license": "AGPL-3.0",
   "homepage": "https://github.com/us/crw",
diff --git a/npm/crw-mcp/package.json b/npm/crw-mcp/package.json
index 3283fef..48ba1a2 100644
--- a/npm/crw-mcp/package.json
+++ b/npm/crw-mcp/package.json
@@ -1,6 +1,6 @@
 {
   "name": "crw-mcp",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "MCP server for CRW web scraper — scrape, crawl, and map tools for AI agents",
   "license": "AGPL-3.0",
   "homepage": "https://github.com/us/crw",
@@ -27,11 +27,11 @@
     "skills/SKILL.md"
   ],
   "optionalDependencies": {
-    "crw-mcp-darwin-x64": "0.10.0",
-    "crw-mcp-darwin-arm64": "0.10.0",
-    "crw-mcp-linux-x64": "0.10.0",
-    "crw-mcp-linux-arm64": "0.10.0",
-    "crw-mcp-win32-x64": "0.10.0",
-    "crw-mcp-win32-arm64": "0.10.0"
+    "crw-mcp-darwin-x64": "0.11.0",
+    "crw-mcp-darwin-arm64": "0.11.0",
+    "crw-mcp-linux-x64": "0.11.0",
+    "crw-mcp-linux-arm64": "0.11.0",
+    "crw-mcp-win32-x64": "0.11.0",
+    "crw-mcp-win32-arm64": "0.11.0"
   }
 }
diff --git a/python/pyproject.toml b/python/pyproject.toml
index aa3b82b..c13634e 100644
--- a/python/pyproject.toml
+++ b/python/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "crw"
-version = "0.10.0"
+version = "0.11.0"
 description = "Python SDK for CRW web scraper — scrape, crawl, and map any website from Python"
 readme = "README.md"
 requires-python = ">=3.9"
diff --git a/server.json b/server.json
index 5e943c6..8149c0f 100644
--- a/server.json
+++ b/server.json
@@ -3,7 +3,7 @@
   "name": "io.github.us/crw",
   "title": "CRW Web Scraper",
   "description": "Open-source web scraper for AI agents with scrape, crawl, and map tools",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "websiteUrl": "https://us.github.io/crw",
   "repository": {
     "url": "https://github.com/us/crw",