diff --git a/openhands-sdk/openhands/sdk/conversation/__init__.py b/openhands-sdk/openhands/sdk/conversation/__init__.py
index d89d3040d..82a518483 100644
--- a/openhands-sdk/openhands/sdk/conversation/__init__.py
+++ b/openhands-sdk/openhands/sdk/conversation/__init__.py
@@ -8,6 +8,12 @@
 from openhands.sdk.conversation.secret_registry import SecretRegistry
 from openhands.sdk.conversation.state import ConversationState
 from openhands.sdk.conversation.stuck_detector import StuckDetector
+from openhands.sdk.conversation.token_display import (
+    TokenDisplay,
+    TokenDisplayMode,
+    compute_token_display,
+    get_default_mode_from_env,
+)
 from openhands.sdk.conversation.types import ConversationCallbackType
 from openhands.sdk.conversation.visualizer import ConversationVisualizer
 
@@ -25,4 +31,9 @@
     "RemoteConversation",
     "EventsListBase",
     "get_agent_final_response",
+    # Token display utilities (public SDK API)
+    "TokenDisplay",
+    "TokenDisplayMode",
+    "compute_token_display",
+    "get_default_mode_from_env",
 ]
diff --git a/openhands-sdk/openhands/sdk/conversation/token_display.py b/openhands-sdk/openhands/sdk/conversation/token_display.py
new file mode 100644
index 000000000..96a733a4a
--- /dev/null
+++ b/openhands-sdk/openhands/sdk/conversation/token_display.py
@@ -0,0 +1,124 @@
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass
+from enum import Enum
+
+from openhands.sdk.conversation.conversation_stats import ConversationStats
+from openhands.sdk.llm.utils.metrics import Metrics, TokenUsage
+
+
+class TokenDisplayMode(str, Enum):
+    PER_CONTEXT = "per_context"  # show metrics for the latest request only
+    ACCUMULATED = "accumulated"  # show accumulated tokens across all requests
+
+    @classmethod
+    def from_str(cls, value: str | None) -> TokenDisplayMode:
+        if not value:
+            return cls.PER_CONTEXT
+        v = value.strip().lower().replace("-", "_")
+        if v in {"per_context", "per_request", "latest", "current"}:
+            return cls.PER_CONTEXT
+        if v in {"accumulated", "total", "sum"}:
+            return cls.ACCUMULATED
+        # default to per-context to match current visual default and tests
+        return cls.PER_CONTEXT
+
+
+@dataclass(frozen=True)
+class TokenDisplay:
+    # Raw counts (not abbreviated)
+    input_tokens: int
+    output_tokens: int
+    cache_read_tokens: int
+    reasoning_tokens: int
+    context_window: int
+    # Rate [0.0, 1.0] or None if undefined
+    cache_hit_rate: float | None
+    # Total accumulated cost in USD
+    total_cost: float
+    # Optional delta of input tokens compared to previous request
+    since_last_input_tokens: int | None = None
+
+
+def _get_combined_metrics(stats: ConversationStats | None) -> Metrics | None:
+    if not stats:
+        return None
+    try:
+        return stats.get_combined_metrics()
+    except Exception:
+        return None
+
+
+def compute_token_display(
+    stats: ConversationStats | None,
+    mode: TokenDisplayMode = TokenDisplayMode.PER_CONTEXT,
+    include_since_last: bool = False,
+) -> TokenDisplay | None:
+    """Compute token display values from conversation stats.
+
+    Args:
+        stats: ConversationStats to read metrics from
+        mode: Whether to show per-context (latest request) or accumulated values
+        include_since_last: If True, include the delta of input tokens compared to
+            the previous request (when available)
+
+    Returns:
+        TokenDisplay with raw numeric values, or None if metrics are unavailable
+    """
+    combined = _get_combined_metrics(stats)
+    if not combined:
+        return None
+
+    # No token usage recorded yet
+    if not combined.token_usages:
+        return None
+
+    total_cost = combined.accumulated_cost or 0.0
+
+    if mode == TokenDisplayMode.ACCUMULATED:
+        usage: TokenUsage | None = combined.accumulated_token_usage
+        if usage is None:
+            return None
+        input_tokens = usage.prompt_tokens or 0
+        output_tokens = usage.completion_tokens or 0
+        cache_read = usage.cache_read_tokens or 0
+        reasoning_tokens = usage.reasoning_tokens or 0
+        context_window = usage.context_window or 0
+        cache_hit_rate = (cache_read / input_tokens) if input_tokens > 0 else None
+        since_last: int | None = None
+    else:  # PER_CONTEXT
+        usage = combined.token_usages[-1]
+        input_tokens = usage.prompt_tokens or 0
+        output_tokens = usage.completion_tokens or 0
+        cache_read = usage.cache_read_tokens or 0
+        reasoning_tokens = usage.reasoning_tokens or 0
+        context_window = usage.context_window or 0
+        cache_hit_rate = (cache_read / input_tokens) if input_tokens > 0 else None
+        since_last = None
+        if include_since_last and len(combined.token_usages) >= 2:
+            prev = combined.token_usages[-2]
+            since_last = max(0, (usage.prompt_tokens or 0) - (prev.prompt_tokens or 0))
+
+    return TokenDisplay(
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        cache_read_tokens=cache_read,
+        reasoning_tokens=reasoning_tokens,
+        context_window=context_window,
+        cache_hit_rate=cache_hit_rate,
+        total_cost=total_cost,
+        since_last_input_tokens=since_last,
+    )
+
+
+def get_default_mode_from_env() -> TokenDisplayMode:
+    """Resolve default token display mode from env var.
+
+    Env var: OH_TOKENS_VIEW_MODE
+      - "per_context" (default)
+      - "accumulated"
+      - also accepts aliases: per_request/latest/current and total/sum
+    """
+    value = os.environ.get("OH_TOKENS_VIEW_MODE")
+    return TokenDisplayMode.from_str(value)
diff --git a/openhands-sdk/openhands/sdk/conversation/visualizer.py b/openhands-sdk/openhands/sdk/conversation/visualizer.py
index df12c129c..8150d8025 100644
--- a/openhands-sdk/openhands/sdk/conversation/visualizer.py
+++ b/openhands-sdk/openhands/sdk/conversation/visualizer.py
@@ -1,3 +1,4 @@
+import os
 import re
 from typing import TYPE_CHECKING
 
@@ -5,6 +6,10 @@
 from rich.panel import Panel
 from rich.text import Text
 
+from openhands.sdk.conversation.token_display import (
+    compute_token_display,
+    get_default_mode_from_env,
+)
 from openhands.sdk.event import (
     ActionEvent,
     AgentErrorEvent,
@@ -269,19 +274,27 @@ def _create_event_panel(self, event: Event) -> Panel | None:
             )
 
     def _format_metrics_subtitle(self) -> str | None:
-        """Format LLM metrics as a visually appealing subtitle string with icons,
-        colors, and k/m abbreviations using conversation stats."""
+        """Format LLM metrics subtitle based on conversation stats.
+
+        Uses TokenDisplay utility to compute values and supports env-configured
+        mode (per-context vs accumulated) and optional since-last delta.
+        """
+        display_mode = get_default_mode_from_env()
+        include_since_last = os.environ.get(
+            "OH_TOKENS_VIEW_DELTA", "false"
+        ).lower() in {"1", "true", "yes"}
+
         if not self._conversation_stats:
             return None
 
-        combined_metrics = self._conversation_stats.get_combined_metrics()
-        if not combined_metrics or not combined_metrics.accumulated_token_usage:
+        data = compute_token_display(
+            stats=self._conversation_stats,
+            mode=display_mode,
+            include_since_last=include_since_last,
+        )
+        if not data:
             return None
 
-        usage = combined_metrics.accumulated_token_usage
-        cost = combined_metrics.accumulated_cost or 0.0
-
-        # helper: 1234 -> "1.2K", 1200000 -> "1.2M"
         def abbr(n: int | float) -> str:
             n = int(n or 0)
             if n >= 1_000_000_000:
@@ -294,26 +307,25 @@ def abbr(n: int | float) -> str:
                 return str(n)
             return s.replace(".0", "")
 
-        input_tokens = abbr(usage.prompt_tokens or 0)
-        output_tokens = abbr(usage.completion_tokens or 0)
-
-        # Cache hit rate (prompt + cache)
-        prompt = usage.prompt_tokens or 0
-        cache_read = usage.cache_read_tokens or 0
-        cache_rate = f"{(cache_read / prompt * 100):.2f}%" if prompt > 0 else "N/A"
-        reasoning_tokens = usage.reasoning_tokens or 0
-
-        # Cost
-        cost_str = f"{cost:.4f}" if cost > 0 else "0.00"
+        cache_rate = (
+            f"{(data.cache_hit_rate * 100):.2f}%"
+            if data.cache_hit_rate is not None
+            else "N/A"
+        )
+        cost_str = f"{data.total_cost:.4f}"
 
-        # Build with fixed color scheme
         parts: list[str] = []
-        parts.append(f"[cyan]↑ input {input_tokens}[/cyan]")
+        input_part = f"[cyan]↑ input {abbr(data.input_tokens)}"
+        if include_since_last and data.since_last_input_tokens is not None:
+            input_part += f" (+{abbr(data.since_last_input_tokens)})"
+        input_part += "[/cyan]"
+        parts.append(input_part)
+
         parts.append(f"[magenta]cache hit {cache_rate}[/magenta]")
-        if reasoning_tokens > 0:
-            parts.append(f"[yellow] reasoning {abbr(reasoning_tokens)}[/yellow]")
-        parts.append(f"[blue]↓ output {output_tokens}[/blue]")
-        parts.append(f"[green]$ {cost_str}[/green]")
+        if data.reasoning_tokens > 0:
+            parts.append(f"[yellow] reasoning {abbr(data.reasoning_tokens)}[/yellow]")
+        parts.append(f"[blue]↓ output {abbr(data.output_tokens)}[/blue]")
+        parts.append(f"[green]$ {cost_str} (total)[/green]")
 
         return "Tokens: " + " • ".join(parts)
 
diff --git a/tests/sdk/conversation/test_token_display.py b/tests/sdk/conversation/test_token_display.py
new file mode 100644
index 000000000..6357242d7
--- /dev/null
+++ b/tests/sdk/conversation/test_token_display.py
@@ -0,0 +1,104 @@
+import pytest
+
+from openhands.sdk.conversation import (
+    TokenDisplayMode,
+    compute_token_display,
+)
+from openhands.sdk.conversation.conversation_stats import ConversationStats
+from openhands.sdk.conversation.visualizer import ConversationVisualizer
+from openhands.sdk.llm.utils.metrics import Metrics
+
+
+@pytest.fixture(autouse=True)
+def _clear_env(monkeypatch):
+    # Ensure env vars do not leak between tests
+    monkeypatch.delenv("OH_TOKENS_VIEW_MODE", raising=False)
+    monkeypatch.delenv("OH_TOKENS_VIEW_DELTA", raising=False)
+
+
+def _make_stats_with_two_requests():
+    stats = ConversationStats()
+    m = Metrics(model_name="test-model")
+    # First call
+    m.add_cost(0.1)
+    m.add_token_usage(
+        prompt_tokens=100,
+        completion_tokens=25,
+        cache_read_tokens=10,
+        cache_write_tokens=0,
+        reasoning_tokens=5,
+        context_window=8000,
+        response_id="first",
+    )
+    # Second call
+    m.add_cost(0.05)
+    m.add_token_usage(
+        prompt_tokens=220,
+        completion_tokens=80,
+        cache_read_tokens=44,
+        cache_write_tokens=0,
+        reasoning_tokens=0,
+        context_window=8000,
+        response_id="second",
+    )
+    stats.usage_to_metrics["usage-1"] = m
+    return stats
+
+
+def test_compute_token_display_per_context_with_delta():
+    stats = _make_stats_with_two_requests()
+
+    td = compute_token_display(
+        stats=stats, mode=TokenDisplayMode.PER_CONTEXT, include_since_last=True
+    )
+    assert td is not None
+
+    # Latest request values
+    assert td.input_tokens == 220
+    assert td.output_tokens == 80
+    assert td.reasoning_tokens == 0
+    assert td.cache_read_tokens == 44
+    assert td.context_window == 8000
+    assert td.total_cost == pytest.approx(0.15)
+    assert td.cache_hit_rate == pytest.approx(44 / 220)
+
+    # Delta vs previous
+    assert td.since_last_input_tokens == 120  # 220 - 100
+
+
+def test_compute_token_display_accumulated():
+    stats = _make_stats_with_two_requests()
+
+    td = compute_token_display(stats=stats, mode=TokenDisplayMode.ACCUMULATED)
+    assert td is not None
+
+    # Accumulated values: sums of prompt/completion/cache_read; max context_window
+    assert td.input_tokens == 100 + 220
+    assert td.output_tokens == 25 + 80
+    assert td.cache_read_tokens == 10 + 44
+    assert td.reasoning_tokens == 5 + 0
+    assert td.context_window == 8000
+    assert td.total_cost == pytest.approx(0.15)
+    assert td.cache_hit_rate == pytest.approx((10 + 44) / (100 + 220))
+
+    # No since-last in accumulated mode
+    assert td.since_last_input_tokens is None
+
+
+def test_visualizer_env_vars_toggle_delta(monkeypatch):
+    stats = _make_stats_with_two_requests()
+
+    # Force per-context and delta
+    monkeypatch.setenv("OH_TOKENS_VIEW_MODE", "per_context")
+    monkeypatch.setenv("OH_TOKENS_VIEW_DELTA", "true")
+
+    viz = ConversationVisualizer(conversation_stats=stats)
+    subtitle = viz._format_metrics_subtitle()
+    assert subtitle is not None
+    assert "(+" in subtitle  # shows since-last delta
+
+    # Force accumulated mode: should hide delta even if env says true
+    monkeypatch.setenv("OH_TOKENS_VIEW_MODE", "accumulated")
+    subtitle2 = viz._format_metrics_subtitle()
+    assert subtitle2 is not None
+    assert "(+" not in subtitle2
diff --git a/tests/sdk/conversation/test_visualizer.py b/tests/sdk/conversation/test_visualizer.py
index 7b479c2c0..5e1a1c9e0 100644
--- a/tests/sdk/conversation/test_visualizer.py
+++ b/tests/sdk/conversation/test_visualizer.py
@@ -337,7 +337,47 @@ def test_metrics_formatting():
     assert "500" in subtitle  # Output tokens
     assert "20.00%" in subtitle  # Cache hit rate
     assert "200" in subtitle  # Reasoning tokens
-    assert "0.0234" in subtitle  # Cost
+    assert "$ 0.0234 (total)" in subtitle
+
+
+def test_metrics_formatting_uses_latest_request():
+    """Tokens should reflect the latest request while cost stays cumulative."""
+    from openhands.sdk.conversation.conversation_stats import ConversationStats
+    from openhands.sdk.llm.utils.metrics import Metrics
+
+    conversation_stats = ConversationStats()
+    metrics = Metrics(model_name="test-model")
+    metrics.add_cost(0.1)
+    metrics.add_token_usage(
+        prompt_tokens=120,
+        completion_tokens=40,
+        cache_read_tokens=12,
+        cache_write_tokens=0,
+        reasoning_tokens=5,
+        context_window=8000,
+        response_id="first",
+    )
+    metrics.add_cost(0.05)
+    metrics.add_token_usage(
+        prompt_tokens=200,
+        completion_tokens=75,
+        cache_read_tokens=25,
+        cache_write_tokens=0,
+        reasoning_tokens=0,
+        context_window=8000,
+        response_id="second",
+    )
+    conversation_stats.service_to_metrics["test_service"] = metrics
+
+    visualizer = ConversationVisualizer(conversation_stats=conversation_stats)
+
+    subtitle = visualizer._format_metrics_subtitle()
+    assert subtitle is not None
+    assert "input 200" in subtitle
+    assert "output 75" in subtitle
+    assert "cache hit 10.00%" not in subtitle  # ensure using latest cache values
+    assert "cache hit 12.50%" in subtitle
+    assert "$ 0.1500 (total)" in subtitle
 
 
 def test_event_base_fallback_visualize():