Fixed reasoning tokens inflating the exact token count

app/MindWork AI Studio/Chat/ChatThread.cs

@@ -416,22 +416,25 @@ public sealed record ChatThread
     /// is a lot of machinery for a number which corrects itself one answer later.
     /// </remarks>
     /// <param name="model">The model the next request would go to.</param>
-    /// <returns>What the provider reported, or unknown when no report describes this conversation.</returns>
+    /// <returns>
-    public TokenUsage ReportedUsageFor(Model model)
+    /// What the provider reported, together with the answer which followed it, or
+    /// ReportedHistory.UNKNOWN when no report describes this conversation.
+    /// </returns>
+    public ReportedHistory ReportedHistoryFor(Model model)
     {
         if (this.Blocks.Count is 0)
-            return TokenUsage.UNKNOWN;
+            return ReportedHistory.UNKNOWN;
 
-        if (this.Blocks[^1].Content is not ContentText { IsStreaming: false, ReportedUsage: { } reported })
+        if (this.Blocks[^1].Content is not ContentText { IsStreaming: false, ReportedUsage: { } reported } answer)
-            return TokenUsage.UNKNOWN;
+            return ReportedHistory.UNKNOWN;
 
         if (reported.BlockCount != this.Blocks.Count)
-            return TokenUsage.UNKNOWN;
+            return ReportedHistory.UNKNOWN;
 
         if (!string.Equals(reported.ModelId, model.Id, StringComparison.Ordinal))
-            return TokenUsage.UNKNOWN;
+            return ReportedHistory.UNKNOWN;
 
-        return reported.ToTokenUsage();
+        return ReportedHistory.Of(reported.ToTokenUsage(), answer.Text);
     }
 
     private static void DeleteManagedAttachments(ContentBlock block)

app/MindWork AI Studio/Chat/ContentText.cs

@@ -62,7 +62,7 @@ public sealed class ContentText : IContent
     /// answer written before this was recorded, and at every provider which reports nothing.
     ///
     /// Having a report is not the same as the report still being true. Whether it still describes
-    /// the conversation is decided by ChatThread.ReportedUsageFor, which looks at the thread around
+    /// the conversation is decided by ChatThread.ReportedHistoryFor, which looks at the thread around
     /// the answer, not just at the answer.
     /// </remarks>
     public ReportedTokenUsage? ReportedUsage { get; set; }
@@ -132,7 +132,6 @@ public sealed class ContentText : IContent
         this.ReportedUsage = new()
         {
             PromptTokens = usage.PromptTokens,
-            CompletionTokens = usage.CompletionTokens,
             ModelId = modelId,
             BlockCount = blockCount,
         };

app/MindWork AI Studio/Chat/ConversationTokens.cs

@@ -61,8 +61,11 @@ public readonly record struct ConversationTokens
     /// </summary>
     /// <remarks>
     /// True once a provider has stated what a request of this conversation cost, which makes
-    /// everything but the draft an exact number. It says nothing about the draft, which stays an
+    /// everything up to the last answer an exact number. That answer is counted by this app, since
-    /// estimate either way -- nobody has charged for that one yet.
+    /// the provider's number for it includes reasoning which no request carries; next to the rest
+    /// of the conversation, its share of the error is small enough to still call the history exact.
+    /// It says nothing about the draft, which stays an estimate either way -- nobody has charged
+    /// for that one yet.
     /// </remarks>
     public bool HistoryIsReported { get; init; }
 

app/MindWork AI Studio/Chat/ReportedHistory.cs

@@ -0,0 +1,59 @@
+using AIStudio.Provider;
+
+namespace AIStudio.Chat;
+
+/// <summary>
+/// What a conversation carries into its next request, as far as a provider has stated it.
+/// </summary>
+/// <remarks>
+/// Two parts, and only one of them is the provider's statement. PromptTokens is what the provider
+/// counted for the request behind the last answer: the system prompt, the tools, and every message
+/// up to the question. The answer travels in the next request as well, though not in the shape the
+/// provider charged for: its completion included the model's reasoning and whatever the model wrote
+/// between think tags. So the answer comes along as the text it will be sent as, and is counted the
+/// same way every other text is.
+///
+/// That text is the answer alone. Reasoning never belongs to it, whether it was thrown away or kept
+/// to be read next to the answer: it is there for a person, and no request carries it. An answer
+/// which consists of reasoning only has no text at all, is not sent, and adds nothing to the prompt.
+///
+/// What is estimated that way is one answer, next to a prompt which holds the whole conversation
+/// before it. The error is the tokenizer's error on that one answer, not on the chat.
+/// </remarks>
+public sealed record ReportedHistory
+{
+    /// <summary>
+    /// The history of a conversation no report describes.
+    /// </summary>
+    public static readonly ReportedHistory UNKNOWN = new();
+
+    /// <summary>
+    /// Whether a report describes the conversation. When false, nothing else here means anything.
+    /// </summary>
+    public bool IsKnown { get; private init; }
+
+    /// <summary>
+    /// What the provider counted for the request behind the last answer.
+    /// </summary>
+    public int PromptTokens { get; private init; }
+
+    /// <summary>
+    /// The text of the last answer, as the next request will carry it, without any reasoning.
+    /// </summary>
+    public string LastAnswer { get; private init; } = string.Empty;
+
+    /// <summary>
+    /// States what a report says about the conversation.
+    /// </summary>
+    /// <param name="usage">What the provider reported for the request behind the last answer.</param>
+    /// <param name="lastAnswer">The text of that answer.</param>
+    /// <returns>The history, or UNKNOWN when the usage states nothing.</returns>
+    public static ReportedHistory Of(TokenUsage usage, string lastAnswer) => usage.IsKnown
+        ? new()
+        {
+            IsKnown = true,
+            PromptTokens = usage.PromptTokens,
+            LastAnswer = lastAnswer,
+        }
+        : UNKNOWN;
+}

app/MindWork AI Studio/Chat/ReportedTokenUsage.cs

@@ -6,10 +6,11 @@ namespace AIStudio.Chat;
 /// What a provider said the request behind one answer cost, as it is stored on that answer.
 /// </summary>
 /// <remarks>
-/// The numbers and the model they were charged for travel as one value. Each of them is meaningless
+/// The number, the model it was charged for, and the conversation it was counted on travel as one
-/// without the others, and loose fields on the answer could be set, cleared, or copied apart.
+/// value. Each of them is meaningless without the others, and loose fields on the answer could be
+/// set, cleared, or copied apart.
 ///
-/// Plain numbers rather than a TokenUsage: that type only ever comes out of its own factory, which
+/// A plain number rather than a TokenUsage: that type only ever comes out of its own factory, which
 /// is what keeps an impossible usage from existing, while a stored value has to be readable back by
 /// the serializer. ToTokenUsage is the way back, and it treats a chat file somebody edited by hand
 /// the same way the reading side treats a provider's JSON.
@@ -22,12 +23,7 @@ public sealed record ReportedTokenUsage
     public int PromptTokens { get; init; }
 
     /// <summary>
-    /// What the model wrote in answer.
+    /// Which model the number was charged for.
-    /// </summary>
-    public int CompletionTokens { get; init; }
-
-    /// <summary>
-    /// Which model the numbers were charged for.
     /// </summary>
     /// <remarks>
     /// A token count belongs to the tokenizer which produced it. Switch the model of a chat, and
@@ -52,8 +48,8 @@ public sealed record ReportedTokenUsage
     public int BlockCount { get; init; }
 
     /// <summary>
-    /// States the stored numbers as a usage again.
+    /// States the stored number as a usage again.
     /// </summary>
-    /// <returns>The usage, or TokenUsage.UNKNOWN when the stored numbers state nothing usable.</returns>
+    /// <returns>The usage, or TokenUsage.UNKNOWN when the stored number states nothing usable.</returns>
-    public TokenUsage ToTokenUsage() => TokenUsage.OfReported(this.PromptTokens, this.CompletionTokens);
+    public TokenUsage ToTokenUsage() => TokenUsage.OfReported(this.PromptTokens);
 }

app/MindWork AI Studio/Components/ChatComponent.razor.cs

@@ -1595,7 +1595,7 @@ public partial class ChatComponent : MSGComponentBase
     {
         var provider = AIStudio.Settings.Provider.NONE;
         var parts = ConversationParts.NOTHING;
-        var reported = TokenUsage.UNKNOWN;
+        var reported = ReportedHistory.UNKNOWN;
 
         //
         // Collected on the render thread, counted off it. Counting may take an IPC call per text,
@@ -1614,7 +1614,7 @@ public partial class ChatComponent : MSGComponentBase
             var toolDefinitions = this.GetRunnableToolDefinitions();
             provider = this.Provider;
             parts = ConversationParts.Of(thread, this.BuildSystemPromptFor(thread, toolDefinitions), this.UserInput, this.ComposerState.FileAttachments, provider.SupportsImageInput(), toolDefinitions);
-            reported = thread.ReportedUsageFor(provider.Model);
+            reported = thread.ReportedHistoryFor(provider.Model);
         });
 
         var counted = await this.ConversationTokenCounter.CountAsync(provider, parts, reported, token);

app/MindWork AI Studio/Provider/BaseProvider.cs

@@ -1128,7 +1128,7 @@ public abstract class BaseProvider : IProvider, ISecretId
                 {
                     yield return providerResponse.ContainsContent()
                         ? providerResponse.GetContent() with { Usage = usage }
-                        : new(string.Empty, [], usage);
+                        : new(string.Empty, [], Usage: usage);
 
                     continue;
                 }

app/MindWork AI Studio/Provider/OpenAI/ChatCompletionUsage.cs

@@ -5,9 +5,12 @@ namespace AIStudio.Provider.OpenAI;
 /// What an OpenAI-compatible provider reports a chat completion cost.
 /// </summary>
 /// <remarks>
-/// Every number is optional because this is somebody else's JSON: the block arrives only when the
+/// The number is optional because this is somebody else's JSON: the block arrives only when the
 /// request asked for it, and the providers which follow the shape loosely leave fields out. Reading
 /// it is one thing, believing it another -- TokenUsage.OfReported decides that.
+///
+/// The block states more than this, the completion and its reasoning share among it. Those are left
+/// unread on purpose, for the reason given at TokenUsage: no later request carries them.
 /// </remarks>
 public sealed record ChatCompletionUsage
 {
@@ -16,11 +19,6 @@ public sealed record ChatCompletionUsage
     /// </summary>
     public int? PromptTokens { get; init; }
 
-    /// <summary>
-    /// What the model wrote in answer.
-    /// </summary>
-    public int? CompletionTokens { get; init; }
-
     /// <summary>
     /// States what this block reports, as far as it can be believed.
     /// </summary>
@@ -29,5 +27,5 @@ public sealed record ChatCompletionUsage
     /// so that what counts as believable is decided in a single place.
     /// </remarks>
     /// <returns>The usage, or TokenUsage.UNKNOWN when the block states nothing usable.</returns>
-    public TokenUsage ToTokenUsage() => TokenUsage.OfReported(this.PromptTokens, this.CompletionTokens);
+    public TokenUsage ToTokenUsage() => TokenUsage.OfReported(this.PromptTokens);
 }

app/MindWork AI Studio/Provider/TokenUsage.cs

@@ -1,13 +1,18 @@
 namespace AIStudio.Provider;
 
 /// <summary>
-/// What a provider said one request actually cost, in tokens.
+/// What a provider said one request actually carried, in tokens.
 /// </summary>
 /// <remarks>
 /// The counterpart to what the app counts for itself: the app estimates what the next request will
-/// cost, while this is what the provider charged for the last one. Two different statements, and
+/// cost, while this is what the provider counted for the last one. Two different statements, and
 /// this one is the only exact one of the two.
 ///
+/// Only the prompt is kept. Providers state what the answer cost as well, but that number includes
+/// the model's reasoning and whatever the model wrote between think tags, and neither of them ever
+/// becomes part of the answer's text. No later request carries them, so the number has no place in
+/// a statement about those requests.
+///
 /// Nothing here says "unknown" with a zero. The default value of this type is unknown, which is the
 /// right answer for every provider which reports nothing, and a counted request can never cost zero
 /// prompt tokens because the factory below refuses to build one.
@@ -20,7 +25,7 @@ public readonly record struct TokenUsage
     public static readonly TokenUsage UNKNOWN = new();
 
     /// <summary>
-    /// Whether a provider reported anything at all. When false, the numbers are meaningless.
+    /// Whether a provider reported anything at all. When false, the number is meaningless.
     /// </summary>
     public bool IsKnown { get; private init; }
 
@@ -30,37 +35,23 @@ public readonly record struct TokenUsage
     /// </summary>
     public int PromptTokens { get; private init; }
 
-    /// <summary>
-    /// What the model wrote in answer.
-    /// </summary>
-    public int CompletionTokens { get; private init; }
-
-    /// <summary>
-    /// What the whole exchange cost, which is what the next request carries as its history.
-    /// </summary>
-    public int TotalTokens => this.PromptTokens + this.CompletionTokens;
-
     /// <summary>
     /// States what a provider reported.
     /// </summary>
     /// <remarks>
-    /// A completion of zero tokens is a real answer: a model which was cut off before writing
+    /// A prompt of zero is not a report, because there is no request without one, and a provider
-    /// anything still cost its prompt. A prompt of zero is not, because there is no request
+    /// sending it means we read the wrong field.
-    /// without one, and a provider sending it means we read the wrong field.
     /// </remarks>
     /// <param name="promptTokens">What the request carried. Has to be greater than zero.</param>
-    /// <param name="completionTokens">What the answer cost. Zero or more.</param>
     /// <returns>The usage.</returns>
-    public static TokenUsage Of(int promptTokens, int completionTokens)
+    public static TokenUsage Of(int promptTokens)
     {
         ArgumentOutOfRangeException.ThrowIfNegativeOrZero(promptTokens);
-        ArgumentOutOfRangeException.ThrowIfNegative(completionTokens);
 
         return new()
         {
             IsKnown = true,
             PromptTokens = promptTokens,
-            CompletionTokens = completionTokens,
         };
     }
 
@@ -68,14 +59,10 @@ public readonly record struct TokenUsage
     /// States what a provider reported, or unknown when it reported nothing usable.
     /// </summary>
     /// <remarks>
-    /// For the reading side, where the numbers come out of somebody else's JSON: a missing field, a
+    /// For the reading side, where the number comes out of somebody else's JSON: a missing field, a
-    /// null, or a zero prompt all mean the same thing there, and none of them is worth an exception.
+    /// null, or a zero all mean the same thing there, and none of them is worth an exception.
     /// </remarks>
     /// <param name="promptTokens">What the request carried, as the provider stated it.</param>
-    /// <param name="completionTokens">What the answer cost, as the provider stated it.</param>
     /// <returns>The usage, or UNKNOWN.</returns>
-    public static TokenUsage OfReported(int? promptTokens, int? completionTokens) =>
+    public static TokenUsage OfReported(int? promptTokens) => promptTokens is > 0 ? Of(promptTokens.Value) : UNKNOWN;
-        promptTokens is > 0
-            ? Of(promptTokens.Value, completionTokens is > 0 ? completionTokens.Value : 0)
-            : UNKNOWN;
 }

app/MindWork AI Studio/Tools/Services/ConversationTokenCounter.cs

@@ -69,12 +69,13 @@ public sealed class ConversationTokenCounter(RustService rustService, ILogger<Co
     /// <param name="provider">The configured provider, which decides both the tokenizer and the window.</param>
     /// <param name="parts">What the conversation would send, collected beforehand.</param>
     /// <param name="reported">
-    /// What a provider said the last request of this conversation cost, where one did. It replaces
+    /// What a provider counted for the request behind the last answer, where its report still
-    /// everything the app would otherwise estimate about the conversation so far.
+    /// describes this conversation. Together with that answer, it replaces everything the app would
+    /// otherwise estimate about the conversation so far.
     /// </param>
     /// <param name="token">Ends the counting when nobody needs the answer anymore.</param>
     /// <returns>What the conversation costs, or that nothing could be counted.</returns>
-    public async Task<ConversationTokens> CountAsync(Provider provider, ConversationParts parts, TokenUsage reported = default, CancellationToken token = default)
+    public async Task<ConversationTokens> CountAsync(Provider provider, ConversationParts parts, ReportedHistory reported, CancellationToken token = default)
     {
         if (provider.UsedLLMProvider is LLMProviders.NONE)
             return ConversationTokens.UNAVAILABLE;
@@ -83,6 +84,7 @@ public sealed class ConversationTokenCounter(RustService rustService, ILogger<Co
         var previouslyGrowing = this.stillGrowing;
         var growing = new Dictionary<string, int>(StringComparer.Ordinal);
         var tokens = 0;
+        var reportedTokens = 0;
 
         try
         {
@@ -106,6 +108,14 @@ public sealed class ConversationTokenCounter(RustService rustService, ILogger<Co
 
             foreach (var document in parts.Documents)
                 tokens += await this.CountDocumentAsync(provider, document, token);
+
+            //
+            // The last answer is counted as the text it is sent as, not taken from the report:
+            // ReportedHistory says why the provider's number for it is the wrong one. It is a
+            // finished text of the conversation, so the cache has it already.
+            //
+            if (reported.IsKnown)
+                reportedTokens = reported.PromptTokens + await this.CountTextAsync(provider, reported.LastAnswer, token);
         }
         catch (OperationCanceledException)
         {
@@ -146,14 +156,14 @@ public sealed class ConversationTokenCounter(RustService rustService, ILogger<Co
 
         //
         // Where a provider has said what this conversation cost, that number replaces everything
-        // but the draft. It is the only exact one of the two, and it covers the same ground: the
+        // but the last answer and the draft. It is the only exact one of the two, and it covers
-        // system prompt, the tools, and every message which has been sent.
+        // the same ground: the system prompt, the tools, and every message which has been sent.
         //
         var historyIsReported = reported.IsKnown;
         return new()
         {
             IsKnown = true,
-            Tokens = historyIsReported ? reported.TotalTokens + draftTokens : tokens,
+            Tokens = historyIsReported ? reportedTokens + draftTokens : tokens,
             IsEstimate = string.IsNullOrWhiteSpace(provider.TokenizerPath),
             DraftTokens = draftTokens,
             HistoryIsReported = historyIsReported,

app/Tests/Chat/ChatThreadReportedHistoryTests.cs

@@ -13,7 +13,7 @@ namespace AIStudio.Tests.Chat;
 /// exactly the state the report was taken in, and the report counts again.
 /// </remarks>
 [TestFixture]
-public sealed class ChatThreadReportedUsageTests
+public sealed class ChatThreadReportedHistoryTests
 {
     private static readonly DateTimeOffset START = new(2026, 9, 23, 10, 0, 0, TimeSpan.Zero);
 
@@ -26,12 +26,39 @@ public sealed class ChatThreadReportedUsageTests
     {
         var thread = Thread(Question(1), Answer(2, promptTokens: 1200, blockCount: 2));
 
-        var usage = thread.ReportedUsageFor(MODEL);
+        var history = thread.ReportedHistoryFor(MODEL);
 
         Assert.Multiple(() =>
         {
-            Assert.That(usage.IsKnown, Is.True);
+            Assert.That(history.IsKnown, Is.True);
-            Assert.That(usage.PromptTokens, Is.EqualTo(1200));
+            Assert.That(history.PromptTokens, Is.EqualTo(1200));
+
+            //
+            // The answer comes along as text rather than as the provider's number for it, which
+            // would include the reasoning the next request never carries:
+            //
+            Assert.That(history.LastAnswer, Is.EqualTo("Answer 2"));
+        });
+    }
+
+    [Test]
+    public void AnAnswerWithoutTextAddsNothingToThePrompt()
+    {
+        //
+        // An answer which consists of reasoning only. It may be kept to be read, but without any
+        // text it is never sent, so what the provider counted is all the next request carries of
+        // the conversation so far.
+        //
+        var reasoningOnly = new ContentText { Text = string.Empty };
+        reasoningOnly.RecordReportedUsage(TokenUsage.Of(1200), MODEL.Id, 2);
+        var thread = Thread(Question(1), Block(ChatRole.AI, reasoningOnly, 2));
+
+        var history = thread.ReportedHistoryFor(MODEL);
+        Assert.Multiple(() =>
+        {
+            Assert.That(history.IsKnown, Is.True);
+            Assert.That(history.PromptTokens, Is.EqualTo(1200));
+            Assert.That(history.LastAnswer, Is.Empty);
         });
     }
 
@@ -45,7 +72,7 @@ public sealed class ChatThreadReportedUsageTests
         //
         var thread = Thread(Question(1), Answer(2, promptTokens: 1200, blockCount: 2), Question(3));
 
-        Assert.That(thread.ReportedUsageFor(MODEL).IsKnown, Is.False);
+        Assert.That(thread.ReportedHistoryFor(MODEL).IsKnown, Is.False);
     }
 
     [Test]
@@ -54,7 +81,7 @@ public sealed class ChatThreadReportedUsageTests
         var streaming = Block(ChatRole.AI, new ContentText { Text = "Half an ans", IsStreaming = true }, 4);
         var thread = Thread(Question(1), Answer(2, promptTokens: 1200, blockCount: 2), Question(3), streaming);
 
-        Assert.That(thread.ReportedUsageFor(MODEL).IsKnown, Is.False);
+        Assert.That(thread.ReportedHistoryFor(MODEL).IsKnown, Is.False);
     }
 
     [Test]
@@ -67,7 +94,7 @@ public sealed class ChatThreadReportedUsageTests
         var withoutReport = Block(ChatRole.AI, new ContentText { Text = "Second answer" }, 4);
         var thread = Thread(Question(1), Answer(2, promptTokens: 1200, blockCount: 2), Question(3), withoutReport);
 
-        Assert.That(thread.ReportedUsageFor(MODEL).IsKnown, Is.False);
+        Assert.That(thread.ReportedHistoryFor(MODEL).IsKnown, Is.False);
     }
 
     [Test]
@@ -82,7 +109,7 @@ public sealed class ChatThreadReportedUsageTests
 
         thread.Remove(firstQuestion.Content!);
 
-        Assert.That(thread.ReportedUsageFor(MODEL).IsKnown, Is.False);
+        Assert.That(thread.ReportedHistoryFor(MODEL).IsKnown, Is.False);
     }
 
     [Test]
@@ -90,7 +117,7 @@ public sealed class ChatThreadReportedUsageTests
     {
         var thread = Thread(Question(1), Answer(2, promptTokens: 1200, blockCount: 2));
 
-        Assert.That(thread.ReportedUsageFor(OTHER_MODEL).IsKnown, Is.False);
+        Assert.That(thread.ReportedHistoryFor(OTHER_MODEL).IsKnown, Is.False);
     }
 
     [Test]
@@ -107,11 +134,12 @@ public sealed class ChatThreadReportedUsageTests
         thread.Remove(lastQuestion.Content!);
         thread.Remove(lastAnswer.Content!);
 
-        var usage = thread.ReportedUsageFor(MODEL);
+        var history = thread.ReportedHistoryFor(MODEL);
         Assert.Multiple(() =>
         {
-            Assert.That(usage.IsKnown, Is.True, "The first answer is the last block again, with exactly the blocks it was reported for.");
+            Assert.That(history.IsKnown, Is.True, "The first answer is the last block again, with exactly the blocks it was reported for.");
-            Assert.That(usage.PromptTokens, Is.EqualTo(1200));
+            Assert.That(history.PromptTokens, Is.EqualTo(1200));
+            Assert.That(history.LastAnswer, Is.EqualTo("Answer 2"));
         });
     }
 
@@ -123,11 +151,12 @@ public sealed class ChatThreadReportedUsageTests
 
         thread.RollBackTo(firstAnswer.Content!);
 
-        var usage = thread.ReportedUsageFor(MODEL);
+        var history = thread.ReportedHistoryFor(MODEL);
         Assert.Multiple(() =>
         {
-            Assert.That(usage.IsKnown, Is.True);
+            Assert.That(history.IsKnown, Is.True);
-            Assert.That(usage.PromptTokens, Is.EqualTo(1200));
+            Assert.That(history.PromptTokens, Is.EqualTo(1200));
+            Assert.That(history.LastAnswer, Is.EqualTo("Answer 2"));
         });
     }
 
@@ -141,7 +170,7 @@ public sealed class ChatThreadReportedUsageTests
     private static ContentBlock Answer(int minute, int promptTokens, int blockCount)
     {
         var answer = new ContentText { Text = $"Answer {minute}" };
-        answer.RecordReportedUsage(TokenUsage.Of(promptTokens, 100), MODEL.Id, blockCount);
+        answer.RecordReportedUsage(TokenUsage.Of(promptTokens), MODEL.Id, blockCount);
         return Block(ChatRole.AI, answer, minute);
     }
 

app/Tests/Provider/ChatCompletionUsageTests.cs

@@ -43,8 +43,6 @@ public sealed class ChatCompletionUsageTests
         {
             Assert.That(line!.GetUsage().IsKnown, Is.True);
             Assert.That(line.GetUsage().PromptTokens, Is.EqualTo(1200));
-            Assert.That(line.GetUsage().CompletionTokens, Is.EqualTo(345));
-            Assert.That(line.GetUsage().TotalTokens, Is.EqualTo(1545));
 
             //
             // The line which carries the usage carries no answer, which is why it has to be read
@@ -109,9 +107,10 @@ public sealed class ChatCompletionUsageTests
     /// The line a real LM Studio server sends, taken off the wire.
     /// </summary>
     /// <remarks>
-    /// It carries fields the shape above does not name -- the reasoning share of the completion,
+    /// It carries far more than we read, and it shows why the prompt is all we take: 102 of the 116
-    /// among them -- and a server which sends more than we read must not stop us from reading what
+    /// completion tokens are the model's reasoning, which the next request never carries. Counting
-    /// we came for.
+    /// the completion would have put the history at 133 tokens, when what travels on is the prompt
+    /// and an answer of a handful of tokens.
     /// </remarks>
     [Test]
     public void ARealServerLineIsRead()
@@ -124,22 +123,7 @@ public sealed class ChatCompletionUsageTests
         Assert.Multiple(() =>
         {
             Assert.That(line!.GetUsage().IsKnown, Is.True);
-            Assert.That(line.GetUsage().TotalTokens, Is.EqualTo(133));
+            Assert.That(line.GetUsage().PromptTokens, Is.EqualTo(17));
-        });
-    }
-
-    /// <summary>
-    /// An answer cut off before it wrote anything still cost its prompt.
-    /// </summary>
-    [Test]
-    public void AnAnswerOfNoTokensIsStillACost()
-    {
-        var usage = TokenUsage.OfReported(900, 0);
-
-        Assert.Multiple(() =>
-        {
-            Assert.That(usage.IsKnown, Is.True);
-            Assert.That(usage.TotalTokens, Is.EqualTo(900));
         });
     }
 }