diff --git a/teammapper-backend/benchmark/run-benchmark.ts b/teammapper-backend/benchmark/run-benchmark.ts
index c1297611..411d0ff7 100644
--- a/teammapper-backend/benchmark/run-benchmark.ts
+++ b/teammapper-backend/benchmark/run-benchmark.ts
@@ -6,7 +6,6 @@ import type { LlmUsageCounting } from '../src/map/services/llm-usage-counter.ser
 const noUsageCounter: LlmUsageCounting = {
   reserve: async () => ({ tokensUsed: 0, requestsCount: 0 }),
   adjustTokens: async () => {},
-  release: async () => {},
 }
 
 async function main(): Promise<void> {
diff --git a/teammapper-backend/src/map/services/ai.service.spec.ts b/teammapper-backend/src/map/services/ai.service.spec.ts
index 081a22de..e9f27150 100644
--- a/teammapper-backend/src/map/services/ai.service.spec.ts
+++ b/teammapper-backend/src/map/services/ai.service.spec.ts
@@ -1,13 +1,15 @@
 import { jest } from '@jest/globals'
 
+import { AiService, DEFAULT_MAX_OUTPUT_TOKENS } from './ai.service'
 import {
-  AiService,
-  DEFAULT_MAX_OUTPUT_TOKENS,
-  SYSTEM_PROMPT_TOKEN_OVERHEAD,
-} from './ai.service'
+  DEFAULT_AI_MAP_SHAPE,
+  systemPrompt,
+  userPrompt,
+} from '../utils/prompts'
+import type { AiMapShape } from '@teammapper/shared'
 import type { LlmUsageCounting } from './llm-usage-counter.service'
 import { RateLimitExceededException } from '../controllers/rate-limit.exception'
-import { generateText } from 'ai'
+import { APICallError, generateText } from 'ai'
 import * as aiProvider from '../utils/aiProvider'
 import configService from '../../config.service'
 import type { LLMProps } from '../../config.service'
@@ -18,10 +20,32 @@ type GetLLMConfigMock = jest.MockedFunction<typeof configService.getLLMConfig>
 
 type MockGenerateTextReturn = Awaited<ReturnType<typeof generateText>>
 
-jest.mock('ai')
+// Only generateText is faked; APICallError keeps its real isInstance check.
+jest.mock('ai', () => ({
+  ...jest.requireActual<typeof import('ai')>('ai'),
+  generateText: jest.fn(),
+}))
 jest.mock('../utils/aiProvider')
 jest.mock('../../config.service')
 
+// One token per UTF-8 byte of the prompts sent, plus the full output cap.
+const estimateFor = (
+  description: string,
+  maxOutputTokens = DEFAULT_MAX_OUTPUT_TOKENS,
+  shape: AiMapShape = DEFAULT_AI_MAP_SHAPE
+) =>
+  Buffer.byteLength(systemPrompt(shape), 'utf8') +
+  Buffer.byteLength(userPrompt(description, 'en'), 'utf8') +
+  maxOutputTokens
+
+const apiError = (statusCode: number | undefined) =>
+  new APICallError({
+    message: 'provider error',
+    url: 'localhost:3000',
+    requestBodyValues: {},
+    statusCode,
+  })
+
 interface FakeUsageState {
   tokensUsed: number
   requestsCount: number
@@ -43,10 +67,6 @@ const buildUsageCounterMock = (
   adjustTokens: jest.fn(async (_dateUsage: string, delta: number) => {
     state.tokensUsed = Math.max(0, state.tokensUsed + delta)
   }),
-  release: jest.fn(async (_dateUsage: string, tokens: number) => {
-    state.tokensUsed = Math.max(0, state.tokensUsed - tokens)
-    state.requestsCount = Math.max(0, state.requestsCount - 1)
-  }),
 })
 
 describe('AiService', () => {
@@ -106,22 +126,32 @@ describe('AiService', () => {
     jest.useRealTimers()
   })
 
-  describe('estimateTokens', () => {
-    it('estimates tokens for short input', () => {
-      expect(aiService.estimateTokens('hello')).toBe(
-        Math.ceil(5 / 4) +
-          SYSTEM_PROMPT_TOKEN_OVERHEAD +
-          DEFAULT_MAX_OUTPUT_TOKENS
-      )
+  describe('token estimate', () => {
+    const reserved = () => usageCounter.reserve.mock.calls[0][1]
+
+    it('counts every byte of both prompts plus the output cap', async () => {
+      await aiService.generateMermaid('hello', 'en')
+
+      expect(reserved()).toBe(estimateFor('hello'))
     })
 
-    it('estimates tokens for empty input', () => {
-      expect(aiService.estimateTokens('')).toBe(
-        SYSTEM_PROMPT_TOKEN_OVERHEAD + DEFAULT_MAX_OUTPUT_TOKENS
+    it('counts a multi-byte character as several tokens', async () => {
+      await aiService.generateMermaid('漢字', 'en')
+
+      // Each of the two characters takes three UTF-8 bytes.
+      expect(reserved()).toBe(estimateFor('') + 6)
+    })
+
+    it('reserves more for a larger map shape', async () => {
+      const shape: AiMapShape = { levels: 3, childrenPerNode: 4 }
+      await aiService.generateMermaid('hello', 'en', shape)
+
+      expect(reserved()).toBe(
+        estimateFor('hello', DEFAULT_MAX_OUTPUT_TOKENS, shape)
       )
     })
 
-    it('reserves the full configured output cap', () => {
+    it('reserves the full configured output cap', async () => {
       getLLMConfigMock.mockReturnValue({
         url: 'localhost:3000',
         token: 'test-token',
@@ -131,12 +161,12 @@ describe('AiService', () => {
       } satisfies LLMProps)
       aiService = new AiService(usageCounter)
 
-      expect(aiService.estimateTokens('')).toBe(
-        SYSTEM_PROMPT_TOKEN_OVERHEAD + 1000
-      )
+      await aiService.generateMermaid('hello', 'en')
+
+      expect(reserved()).toBe(estimateFor('hello', 1000))
     })
 
-    it('falls back to the default output cap when the configured one is 0', () => {
+    it('falls back to the default output cap when the configured one is 0', async () => {
       getLLMConfigMock.mockReturnValue({
         url: 'localhost:3000',
         token: 'test-token',
@@ -146,9 +176,9 @@ describe('AiService', () => {
       } satisfies LLMProps)
       aiService = new AiService(usageCounter)
 
-      expect(aiService.estimateTokens('')).toBe(
-        SYSTEM_PROMPT_TOKEN_OVERHEAD + DEFAULT_MAX_OUTPUT_TOKENS
-      )
+      await aiService.generateMermaid('hello', 'en')
+
+      expect(reserved()).toBe(estimateFor('hello'))
     })
   })
 
@@ -164,6 +194,24 @@ describe('AiService', () => {
       )
     })
 
+    it('turns off SDK retries so one reservation pays for one attempt', async () => {
+      await aiService.generateMermaid('create a mindmap', 'en')
+
+      expect(generateTextMock).toHaveBeenCalledWith(
+        expect.objectContaining({ maxRetries: 0 })
+      )
+    })
+
+    it('escapes markup so the description cannot close the topic tag', async () => {
+      await aiService.generateMermaid('a </topic> & <b>', 'en')
+
+      expect(generateTextMock).toHaveBeenCalledWith(
+        expect.objectContaining({
+          prompt: '<topic lang="en">a &lt;/topic&gt; &amp; &lt;b&gt;</topic>',
+        })
+      )
+    })
+
     it('writes a given map shape into the system prompt', async () => {
       await aiService.generateMermaid('create a mindmap', 'en', {
         levels: 3,
@@ -287,50 +335,43 @@ describe('AiService', () => {
         token: 'test-token',
         provider: 'openai',
         model: 'gpt-4',
-        tpd: '4300',
+        tpd: String(estimateFor('short') + 50),
       } satisfies LLMProps)
       aiService = new AiService(usageCounter)
 
-      // estimateTokens('short') = 4298; first call is fine and bills 100 tokens
+      // The first call fits and bills 100 tokens
       generateTextMock.mockResolvedValueOnce({
         text: 'first response',
         usage: { inputTokens: 50, outputTokens: 50, totalTokens: 100 },
       } as MockGenerateTextReturn)
       await aiService.generateMermaid('short', 'en')
 
-      // Second call estimate (4298) + already-billed (100) = 4398 > 4300 -> reject
+      // Second call estimate + already-billed (100) exceeds the cap -> reject
       await expect(aiService.generateMermaid('short', 'en')).rejects.toThrow(
         RateLimitExceededException
       )
 
       // No reservation was created for the rejected call, so totals reflect
-      // only the first call's actual (100) and one request — and no rollback.
-      expect({
-        ...usageState,
-        releaseCalls: usageCounter.release.mock.calls.length,
-      }).toEqual({
-        tokensUsed: 100,
-        requestsCount: 1,
-        releaseCalls: 0,
-      })
+      // only the first call's actual (100) and one request.
+      expect(usageState).toEqual({ tokensUsed: 100, requestsCount: 1 })
     })
 
-    it('allows a reservation that lands exactly at the TPD cap', async () => {
+    it('allows a reservation that reaches the TPD cap exactly', async () => {
       getLLMConfigMock.mockReturnValue({
         url: 'localhost:3000',
         token: 'test-token',
         provider: 'openai',
         model: 'gpt-4',
-        tpd: '4298',
+        tpd: String(estimateFor('short')),
       } satisfies LLMProps)
       aiService = new AiService(usageCounter)
 
-      // estimateTokens('short') = 4298; equal to the cap, so this must succeed.
+      // The estimate equals the cap, so this must succeed.
       await aiService.generateMermaid('short', 'en')
       expect(usageCounter.reserve).toHaveBeenCalledWith(
         expect.any(String),
-        4298,
-        4298
+        estimateFor('short'),
+        estimateFor('short')
       )
     })
 
@@ -381,13 +422,22 @@ describe('AiService', () => {
       } as MockGenerateTextReturn)
       await aiService.generateMermaid('short', 'en')
 
-      // estimated = 4298, actual = 100 -> delta = -4198
       expect(usageCounter.adjustTokens).toHaveBeenCalledWith(
         expect.any(String),
-        -4198
+        100 - estimateFor('short')
       )
     })
 
+    it('keeps the estimate when the provider reports no usage', async () => {
+      generateTextMock.mockResolvedValueOnce({
+        text: 'response',
+        usage: {},
+      } as MockGenerateTextReturn)
+      await aiService.generateMermaid('short', 'en')
+
+      expect(usageState.tokensUsed).toBe(estimateFor('short'))
+    })
+
     it('keeps the conservative reservation when adjustTokens fails after a successful LLM call', async () => {
       getLLMConfigMock.mockReturnValue({
         url: 'localhost:3000',
@@ -398,36 +448,88 @@ describe('AiService', () => {
       aiService = new AiService(usageCounter)
       usageCounter.adjustTokens.mockRejectedValueOnce(new Error('db hiccup'))
 
-      // Reconciliation failure must not propagate, must not release.
+      // Reconciliation failure must not propagate.
       await aiService.generateMermaid('short', 'en')
-      expect(usageCounter.release).not.toHaveBeenCalled()
-      // Reservation persists at the conservative estimate (4298), not actual.
-      expect(usageState.tokensUsed).toBe(4298)
+      // Reservation persists at the conservative estimate, not actual.
+      expect(usageState.tokensUsed).toBe(estimateFor('short'))
     })
 
-    it('does not mask the original LLM error when release fails during rollback', async () => {
+    it('keeps the estimate booked when generateText fails', async () => {
+      generateTextMock.mockRejectedValueOnce(new Error('boom'))
+      await expect(aiService.generateMermaid('short', 'en')).rejects.toThrow(
+        'boom'
+      )
+
+      // The provider may have billed the failed call.
+      expect(usageState).toEqual({
+        tokensUsed: estimateFor('short'),
+        requestsCount: 1,
+      })
+    })
+
+    it('keeps the estimate when the provider reports a total of 0', async () => {
+      generateTextMock.mockResolvedValueOnce({
+        text: 'response',
+        usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0 },
+      } as MockGenerateTextReturn)
+      await aiService.generateMermaid('short', 'en')
+
+      expect(usageState.tokensUsed).toBe(estimateFor('short'))
+    })
+
+    it('books input plus output when the reported total is lower', async () => {
+      generateTextMock.mockResolvedValueOnce({
+        text: 'response',
+        usage: { inputTokens: 300, outputTokens: 500, totalTokens: 10 },
+      } as MockGenerateTextReturn)
+      await aiService.generateMermaid('short', 'en')
+
+      expect(usageState.tokensUsed).toBe(800)
+    })
+
+    it.each([
+      ['a 401', 401],
+      ['a 400', 400],
+      ['a failed connection', undefined],
+    ])('refunds the reservation after %s', async (_label, statusCode) => {
       getLLMConfigMock.mockReturnValue({
         url: 'localhost:3000',
         token: 'test-token',
         provider: 'openai',
         model: 'gpt-4',
+        rpm: '1',
       } satisfies LLMProps)
       aiService = new AiService(usageCounter)
-      usageCounter.release.mockRejectedValueOnce(new Error('db unreachable'))
 
-      generateTextMock.mockRejectedValueOnce(new Error('boom'))
+      generateTextMock.mockRejectedValueOnce(apiError(statusCode))
       await expect(aiService.generateMermaid('short', 'en')).rejects.toThrow(
-        'boom'
+        APICallError
       )
+      expect(usageState.tokensUsed).toBe(0)
+
+      // The refund frees the per-minute request slot.
+      await aiService.generateMermaid('short', 'en')
     })
 
+    it.each([408, 429, 500, 503])(
+      'keeps the reservation after a %i',
+      async (statusCode) => {
+        generateTextMock.mockRejectedValueOnce(apiError(statusCode))
+        await expect(aiService.generateMermaid('short', 'en')).rejects.toThrow(
+          APICallError
+        )
+
+        expect(usageState.tokensUsed).toBe(estimateFor('short'))
+      }
+    )
+
     it('throws an error if the tokens per minute limit is reached', async () => {
       getLLMConfigMock.mockReturnValue({
         url: 'localhost:3000',
         token: 'test-token',
         provider: 'openai',
         model: 'gpt-4',
-        tpm: '5000',
+        tpm: String(estimateFor('short') + 500),
         rpm: undefined,
         tpd: undefined,
       } satisfies LLMProps)
@@ -478,13 +580,12 @@ describe('AiService', () => {
         token: 'test-token',
         provider: 'openai',
         model: 'gpt-4',
-        tpm: '8000',
+        tpm: String(2 * estimateFor('short') - 1),
       } satisfies LLMProps)
       aiService = new AiService(usageCounter)
 
-      // estimateTokens('short') = 4298 and the long input estimates 4596. Both
-      // together need 8894 reserved up-front; with TPM=8000, only the first
-      // should succeed.
+      // Both calls together need two estimates reserved up-front; with a TPM
+      // one below that, only the first should succeed.
       let release!: () => void
       const block = new Promise<void>((resolve) => {
         release = resolve
@@ -498,16 +599,16 @@ describe('AiService', () => {
       }) as unknown as typeof generateText)
 
       const first = aiService.generateMermaid('short', 'en')
-      // Second call must observe first's pre-charge of 4298 already in the
-      // per-minute window, blocking it instead of racing through the precheck.
-      await expect(
-        aiService.generateMermaid('s'.repeat(1200), 'en')
-      ).rejects.toThrow(RateLimitExceededException)
+      // Second call must observe first's pre-charge already in the per-minute
+      // window, blocking it instead of racing through the precheck.
+      await expect(aiService.generateMermaid('short', 'en')).rejects.toThrow(
+        RateLimitExceededException
+      )
       release()
       await first
     })
 
-    it('releases the per-minute reservation when generateText fails', async () => {
+    it('counts a failed call against the per-minute limits', async () => {
       getLLMConfigMock.mockReturnValue({
         url: 'localhost:3000',
         token: 'test-token',
@@ -522,8 +623,9 @@ describe('AiService', () => {
         'boom'
       )
 
-      // After release, RPM=1 should still allow one more call.
-      await aiService.generateMermaid('short', 'en')
+      await expect(aiService.generateMermaid('short', 'en')).rejects.toThrow(
+        RateLimitExceededException
+      )
     })
 
     it('resets token count after one minute', async () => {
@@ -532,7 +634,7 @@ describe('AiService', () => {
         token: 'test-token',
         provider: 'openai',
         model: 'gpt-4',
-        tpm: '5000',
+        tpm: String(estimateFor('short') + 500),
         rpm: undefined,
         tpd: undefined,
       } satisfies LLMProps)
@@ -563,7 +665,7 @@ describe('AiService', () => {
         token: 'test-token',
         provider: 'openai',
         model: 'gpt-4',
-        tpm: '4500',
+        tpm: String(estimateFor('') + 1000),
         rpm: undefined,
         tpd: undefined,
       } satisfies LLMProps)
diff --git a/teammapper-backend/src/map/services/ai.service.ts b/teammapper-backend/src/map/services/ai.service.ts
index 3a1915e0..fa56c902 100644
--- a/teammapper-backend/src/map/services/ai.service.ts
+++ b/teammapper-backend/src/map/services/ai.service.ts
@@ -1,5 +1,10 @@
 import { Inject, Injectable, Logger } from '@nestjs/common'
-import { generateText, LanguageModel, LanguageModelUsage } from 'ai'
+import {
+  APICallError,
+  generateText,
+  LanguageModel,
+  LanguageModelUsage,
+} from 'ai'
 import {
   AiMapShape,
   MermaidCreateResult,
@@ -19,7 +24,6 @@ import {
   LlmUsageCounterService,
 } from './llm-usage-counter.service'
 
-export const SYSTEM_PROMPT_TOKEN_OVERHEAD = 200
 // A reasoning model spends its thinking from the same output budget as the
 // answer, so both defaults leave room for long reasoning output.
 export const DEFAULT_MAX_OUTPUT_TOKENS = 4096
@@ -82,37 +86,33 @@ export class AiService {
       return { mermaid: '', truncated: false }
     }
 
-    const estimated = this.estimateTokens(mindmapDescription)
+    const system = systemPrompt(shape)
+    const prompt = userPrompt(mindmapDescription, language)
+    const estimated = this.estimateTokens(system, prompt)
     const reservation = await this.reserveBudget(estimated)
-    let generated: MermaidCreateResult
-    let usage: LanguageModelUsage
-    try {
-      const result = await this.callLlm(
-        provider(this.llmConfig.model),
-        mindmapDescription,
-        language,
-        shape
-      )
-      generated = {
-        mermaid: pruneMindmap(result.text, shape),
-        truncated: result.finishReason === 'length',
-      }
-      usage = result.usage
-    } catch (err) {
-      await this.releaseReservation(reservation)
-      throw err
+    const result = await this.callLlmOrRefund(
+      reservation,
+      provider(this.llmConfig.model),
+      system,
+      prompt
+    )
+    const booked = AiService.bookedTokens(result.usage, estimated)
+    await this.tryCommitReservation(reservation, booked)
+    this.logUsage(booked, estimated, result.usage)
+    return {
+      mermaid: pruneMindmap(result.text, shape),
+      truncated: result.finishReason === 'length',
     }
-    // Reconciliation is best-effort: a failure here must not roll back a
-    // successful LLM call (would silently under-bill the budget).
-    await this.tryCommitReservation(reservation, usage.totalTokens ?? 0)
-    this.logUsage(estimated, usage)
-    return generated
   }
 
   /** Logs the token counts of one LLM call. */
-  private logUsage(estimated: number, usage: LanguageModelUsage): void {
+  private logUsage(
+    booked: number,
+    estimated: number,
+    usage: LanguageModelUsage
+  ): void {
     this.logger.debug(
-      `LLM call billed ${usage.totalTokens ?? 0} tokens (estimated ${estimated}, ` +
+      `LLM call booked ${booked} tokens (estimated ${estimated}, ` +
         `input ${usage.inputTokens ?? 0}, output ${usage.outputTokens ?? 0}, ` +
         `reasoning ${usage.outputTokenDetails?.reasoningTokens ?? 0}, ` +
         `cap ${this.limits.maxOutputTokens})`
@@ -120,18 +120,46 @@ export class AiService {
   }
 
   /**
-   * Estimates the most tokens one call can bill: the input, the system prompt
-   * and the full output cap. The reservation holds this estimate until the
-   * call reports its real usage, so parallel calls cannot overshoot a limit.
+   * Estimates the most tokens one call can bill: one token per UTF-8 byte of
+   * the prompts sent, because a byte-level tokenizer emits at most one token
+   * per byte, plus the full output cap. The reservation holds this estimate
+   * until the call reports its real usage, so parallel calls cannot overshoot
+   * a limit.
    */
-  estimateTokens(input: string): number {
+  private estimateTokens(system: string, prompt: string): number {
     return (
-      Math.ceil(input.length / 4) +
-      SYSTEM_PROMPT_TOKEN_OVERHEAD +
+      Buffer.byteLength(system, 'utf8') +
+      Buffer.byteLength(prompt, 'utf8') +
       this.limits.maxOutputTokens
     )
   }
 
+  /**
+   * Picks the tokens to book for a finished call: the larger of the reported
+   * total and the reported input plus output. A provider that reports nothing
+   * keeps the estimate booked.
+   */
+  private static bookedTokens(
+    usage: LanguageModelUsage,
+    estimated: number
+  ): number {
+    const parts = (usage.inputTokens ?? 0) + (usage.outputTokens ?? 0)
+    const reported = Math.max(usage.totalTokens ?? 0, parts)
+    return reported > 0 ? reported : estimated
+  }
+
+  /**
+   * Reports whether the provider rejected a call before running the model and
+   * so billed nothing: no connection, or a 4xx other than 408 and 429. A
+   * timeout or a 5xx may have run the model and counts as billed.
+   */
+  private static isUnbilled(err: unknown): boolean {
+    if (!APICallError.isInstance(err)) return false
+    const status = err.statusCode
+    if (status === undefined) return true
+    return status >= 400 && status < 500 && status !== 408 && status !== 429
+  }
+
   /** Reads a rate limit. A limit of 0 blocks every call. */
   private static parseLimit(raw: string | undefined): number | undefined {
     if (!raw) return undefined
@@ -145,21 +173,44 @@ export class AiService {
     return value !== undefined && value > 0 ? value : undefined
   }
 
-  private async callLlm(
+  /**
+   * Calls the LLM. A failed call keeps its reservation, because the provider
+   * may have billed it, unless the provider rejected the call unbilled.
+   */
+  private async callLlmOrRefund(
+    reservation: Reservation,
     model: LanguageModel,
-    description: string,
-    language: SupportedLanguage,
-    shape: AiMapShape
+    system: string,
+    prompt: string
   ) {
+    try {
+      return await this.callLlm(model, system, prompt)
+    } catch (err) {
+      if (AiService.isUnbilled(err)) await this.refundReservation(reservation)
+      throw err
+    }
+  }
+
+  // One reservation pays for one attempt, so the SDK must not retry.
+  private async callLlm(model: LanguageModel, system: string, prompt: string) {
     return await generateText({
       model,
-      system: systemPrompt(shape),
-      prompt: userPrompt(description, language),
+      system,
+      prompt,
       maxOutputTokens: this.limits.maxOutputTokens,
+      maxRetries: 0,
       abortSignal: AbortSignal.timeout(this.limits.timeoutMs),
     })
   }
 
+  /** Removes the per-minute entry and books 0 tokens for the day. */
+  private async refundReservation(reservation: Reservation): Promise<void> {
+    this.tokensUsedPerMinute = this.tokensUsedPerMinute.filter(
+      (e) => e !== reservation.entry
+    )
+    await this.tryCommitReservation(reservation, 0)
+  }
+
   private async reserveBudget(estimated: number): Promise<Reservation> {
     this.checkPerMinuteLimits(estimated)
     const entry: PerMinuteEntry = { time: Date.now(), count: estimated }
@@ -228,25 +279,9 @@ export class AiService {
         actual - reservation.estimated
       )
     } catch (err) {
+      const reason = err instanceof Error ? err.message : String(err)
       this.logger.warn(
-        `Failed to reconcile actual tokens for ${reservation.dateUsage}; keeping conservative reservation. ${(err as Error).message}`
-      )
-    }
-  }
-
-  private async releaseReservation(reservation: Reservation): Promise<void> {
-    this.tokensUsedPerMinute = this.tokensUsedPerMinute.filter(
-      (e) => e !== reservation.entry
-    )
-    try {
-      await this.usageCounter.release(
-        reservation.dateUsage,
-        reservation.estimated
-      )
-    } catch (err) {
-      // Don't mask the original error from the caller's catch path.
-      this.logger.error(
-        `Failed to release reserved tokens for ${reservation.dateUsage}: ${(err as Error).message}`
+        `Failed to reconcile actual tokens for ${reservation.dateUsage}; keeping conservative reservation. ${reason}`
       )
     }
   }
diff --git a/teammapper-backend/src/map/services/llm-usage-counter.service.spec.ts b/teammapper-backend/src/map/services/llm-usage-counter.service.spec.ts
index 4682ff1f..d900b50e 100644
--- a/teammapper-backend/src/map/services/llm-usage-counter.service.spec.ts
+++ b/teammapper-backend/src/map/services/llm-usage-counter.service.spec.ts
@@ -65,13 +65,4 @@ describe('LlmUsageCounterService', () => {
       -50,
     ])
   })
-
-  it('release decrements both tokens and request count', async () => {
-    queryMock.mockResolvedValueOnce([])
-    await service.release('2026-05-08', 200)
-    expect(queryMock).toHaveBeenCalledWith(
-      expect.stringContaining('"requestsCount" - 1'),
-      ['2026-05-08', 200]
-    )
-  })
 })
diff --git a/teammapper-backend/src/map/services/llm-usage-counter.service.ts b/teammapper-backend/src/map/services/llm-usage-counter.service.ts
index 1ea7e51e..aed89c67 100644
--- a/teammapper-backend/src/map/services/llm-usage-counter.service.ts
+++ b/teammapper-backend/src/map/services/llm-usage-counter.service.ts
@@ -20,7 +20,6 @@ export interface LlmUsageCounting {
     cap?: number
   ): Promise<{ tokensUsed: number; requestsCount: number } | null>
   adjustTokens(dateUsage: string, delta: number): Promise<void>
-  release(dateUsage: string, tokens: number): Promise<void>
 }
 
 /**
@@ -82,18 +81,4 @@ export class LlmUsageCounterService implements LlmUsageCounting {
       [dateUsage, delta]
     )
   }
-
-  /**
-   * Roll back a reservation entirely (token-wise and request-wise),
-   * e.g. after a failed `generateText` call.
-   */
-  async release(dateUsage: string, tokens: number): Promise<void> {
-    await this.repo.query(
-      `UPDATE llm_usage_counter
-         SET "tokensUsed" = GREATEST(0, "tokensUsed" - $2),
-             "requestsCount" = GREATEST(0, "requestsCount" - 1)
-       WHERE "dateUsage" = $1`,
-      [dateUsage, tokens]
-    )
-  }
 }
diff --git a/teammapper-backend/src/map/utils/prompts.ts b/teammapper-backend/src/map/utils/prompts.ts
index 0bbdc406..0b4b8346 100644
--- a/teammapper-backend/src/map/utils/prompts.ts
+++ b/teammapper-backend/src/map/utils/prompts.ts
@@ -50,5 +50,9 @@ export const systemPrompt = (shape: AiMapShape) =>
    Do not explain the refusal.
    </content_policy>`
 
+// Escapes the markup characters so a description cannot close the <topic> tag.
+const escapeMarkup = (text: string) =>
+  text.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;')
+
 export const userPrompt = (description: string, language: SupportedLanguage) =>
-  `<topic lang="${language}">${description}</topic>`
+  `<topic lang="${language}">${escapeMarkup(description)}</topic>`
diff --git a/teammapper-backend/test/llm-usage-counter.e2e-spec.ts b/teammapper-backend/test/llm-usage-counter.e2e-spec.ts
index 2cee3606..7e6514cc 100644
--- a/teammapper-backend/test/llm-usage-counter.e2e-spec.ts
+++ b/teammapper-backend/test/llm-usage-counter.e2e-spec.ts
@@ -73,12 +73,4 @@ describe('LlmUsageCounterService (e2e)', () => {
     await service.adjustTokens(date, -150)
     expect(await readRow(date)).toEqual({ tokensUsed: 50, requestsCount: 1 })
   })
-
-  it('release rolls a reservation back fully', async () => {
-    const date = '2026-05-13'
-    await service.reserve(date, 200, 1000)
-    await service.reserve(date, 300, 1000)
-    await service.release(date, 300)
-    expect(await readRow(date)).toEqual({ tokensUsed: 200, requestsCount: 1 })
-  })
 })
diff --git a/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.html b/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.html
index 57edc6aa..214ba790 100644
--- a/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.html
+++ b/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.html
@@ -15,7 +15,7 @@
       </p>
       <p class="deleted-at">
         {{ 'PAGES.SETTINGS.DELETED_AT' | translate }}
-        {{ attached.cachedMap.deletedAt | date: 'yyyy-MM-dd' }}
+        {{ attached.cachedMap.deletedAt | localizedDate }}
       </p>
     </mat-card-content>
     @if (mapAdminId | async) {
diff --git a/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.spec.ts b/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.spec.ts
index 0e6478d1..a3d8eae3 100644
--- a/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.spec.ts
+++ b/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.spec.ts
@@ -1,6 +1,6 @@
 import { ComponentFixture, TestBed } from '@angular/core/testing';
 import { Router } from '@angular/router';
-import { TranslateModule } from '@ngx-translate/core';
+import { TranslateModule, TranslateService } from '@ngx-translate/core';
 import { BehaviorSubject } from 'rxjs';
 import { ToastrService } from 'ngx-toastr';
 import { CachedMapEntry } from '@teammapper/shared';
@@ -52,6 +52,7 @@ describe('MapDeletionComponent', () => {
         { provide: ToastrService, useValue: toastrService },
       ],
     }).compileComponents();
+    TestBed.inject(TranslateService).use('de');
 
     fixture = TestBed.createComponent(MapDeletionComponent);
     fixture.detectChanges();
@@ -73,7 +74,7 @@ describe('MapDeletionComponent', () => {
 
     expect(
       fixture.nativeElement.querySelector('.deleted-at').textContent
-    ).toContain('2026-10-28');
+    ).toContain('28.10.2026');
   });
 
   it('renders nothing while no map is attached', async () => {
diff --git a/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.ts b/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.ts
index f39bbb42..fbe683d2 100644
--- a/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.ts
+++ b/teammapper-frontend/src/app/modules/application/components/map-deletion/map-deletion.component.ts
@@ -12,7 +12,8 @@ import {
   MatCardActions,
 } from '@angular/material/card';
 import { MatButton } from '@angular/material/button';
-import { AsyncPipe, DatePipe } from '@angular/common';
+import { AsyncPipe } from '@angular/common';
+import { LocalizedDatePipe } from 'src/app/shared/pipes/localized-date.pipe';
 import { StorageService } from 'src/app/core/services/storage/storage.service';
 import { MapSyncService } from 'src/app/core/services/map-sync/map-sync.service';
 
@@ -28,7 +29,7 @@ import { MapSyncService } from 'src/app/core/services/map-sync/map-sync.service'
     MatCardActions,
     MatButton,
     AsyncPipe,
-    DatePipe,
+    LocalizedDatePipe,
     TranslatePipe,
   ],
 })
diff --git a/teammapper-frontend/src/app/shared/components/mindmaps-overview/mindmaps-overview.component.html b/teammapper-frontend/src/app/shared/components/mindmaps-overview/mindmaps-overview.component.html
index 311c901f..9d5c3229 100644
--- a/teammapper-frontend/src/app/shared/components/mindmaps-overview/mindmaps-overview.component.html
+++ b/teammapper-frontend/src/app/shared/components/mindmaps-overview/mindmaps-overview.component.html
@@ -18,7 +18,7 @@
               </span>
               <span matLine>
                 {{ 'PAGES.SETTINGS.TTL' | translate }}:
-                {{ ownedEntry.cachedAdminMapValue.ttl | date: 'yyyy-MM-dd' }}
+                {{ ownedEntry.cachedAdminMapValue.ttl | localizedDate }}
               </span>
               @if (ownedEntry.cachedAdminMapValue.adminId) {
                 <button
@@ -62,7 +62,7 @@
             </span>
             <span matLine>
               {{ 'PAGES.SETTINGS.TTL' | translate }}:
-              {{ cachedMapEntry.cachedAdminMapValue.ttl | date: 'yyyy-MM-dd' }}
+              {{ cachedMapEntry.cachedAdminMapValue.ttl | localizedDate }}
             </span>
             @if (cachedMapEntry.cachedAdminMapValue.adminId) {
               <button
diff --git a/teammapper-frontend/src/app/shared/components/mindmaps-overview/mindmaps-overview.component.ts b/teammapper-frontend/src/app/shared/components/mindmaps-overview/mindmaps-overview.component.ts
index a7d29176..5c7fb18e 100644
--- a/teammapper-frontend/src/app/shared/components/mindmaps-overview/mindmaps-overview.component.ts
+++ b/teammapper-frontend/src/app/shared/components/mindmaps-overview/mindmaps-overview.component.ts
@@ -16,8 +16,9 @@ import { MatList, MatListItem, MatListItemMeta } from '@angular/material/list';
 import { MatLine } from '@angular/material/core';
 import { MatIconButton } from '@angular/material/button';
 import { MatIcon } from '@angular/material/icon';
-import { CommonModule, DatePipe } from '@angular/common';
+import { CommonModule } from '@angular/common';
 import { TranslatePipe, TranslateService } from '@ngx-translate/core';
+import { LocalizedDatePipe } from 'src/app/shared/pipes/localized-date.pipe';
 
 @Component({
   selector: 'teammapper-mindmaps-overview',
@@ -34,7 +35,7 @@ import { TranslatePipe, TranslateService } from '@ngx-translate/core';
     MatLine,
     MatIconButton,
     MatIcon,
-    DatePipe,
+    LocalizedDatePipe,
     TranslatePipe,
     CommonModule,
   ],
diff --git a/teammapper-frontend/src/app/shared/pipes/localized-date.pipe.spec.ts b/teammapper-frontend/src/app/shared/pipes/localized-date.pipe.spec.ts
new file mode 100644
index 00000000..94d35305
--- /dev/null
+++ b/teammapper-frontend/src/app/shared/pipes/localized-date.pipe.spec.ts
@@ -0,0 +1,35 @@
+import { TestBed } from '@angular/core/testing';
+import { TranslateModule, TranslateService } from '@ngx-translate/core';
+import { LocalizedDatePipe } from './localized-date.pipe';
+
+describe('LocalizedDatePipe', () => {
+  const date = new Date(2026, 9, 28);
+
+  function format(language: string): string {
+    TestBed.inject(TranslateService).use(language);
+    return TestBed.runInInjectionContext(() =>
+      new LocalizedDatePipe().transform(date)
+    );
+  }
+
+  beforeEach(() => {
+    TestBed.configureTestingModule({ imports: [TranslateModule.forRoot()] });
+  });
+
+  it('orders day before month in German', () => {
+    expect(format('de')).toBe('28.10.2026');
+  });
+
+  it('orders month before day in English', () => {
+    expect(format('en')).toBe('10/28/2026');
+  });
+
+  it('orders year first in Japanese', () => {
+    expect(format('ja')).toBe('2026/10/28');
+  });
+
+  it('returns an empty string without a date', () => {
+    const pipe = TestBed.runInInjectionContext(() => new LocalizedDatePipe());
+    expect(pipe.transform(null)).toBe('');
+  });
+});
diff --git a/teammapper-frontend/src/app/shared/pipes/localized-date.pipe.ts b/teammapper-frontend/src/app/shared/pipes/localized-date.pipe.ts
new file mode 100644
index 00000000..bbc9ed8a
--- /dev/null
+++ b/teammapper-frontend/src/app/shared/pipes/localized-date.pipe.ts
@@ -0,0 +1,22 @@
+import { Pipe, PipeTransform, inject } from '@angular/core';
+import { TranslateService } from '@ngx-translate/core';
+
+// Formats a date in the order and separators of the current UI language,
+// e.g. 28.10.2026 in German and 10/28/2026 in English. Impure, so the output
+// follows a language switch.
+@Pipe({
+  name: 'localizedDate',
+  pure: false,
+})
+export class LocalizedDatePipe implements PipeTransform {
+  private translateService = inject(TranslateService);
+
+  transform(value: Date | number | string | null | undefined): string {
+    if (value === null || value === undefined) return '';
+
+    return new Date(value).toLocaleDateString(
+      this.translateService.getCurrentLang(),
+      { year: 'numeric', month: '2-digit', day: '2-digit' }
+    );
+  }
+}
diff --git a/teammapper-frontend/src/assets/i18n/de.json b/teammapper-frontend/src/assets/i18n/de.json
index fdfd69fc..03fedace 100644
--- a/teammapper-frontend/src/assets/i18n/de.json
+++ b/teammapper-frontend/src/assets/i18n/de.json
@@ -198,7 +198,7 @@
       "SHOW_LINKTEXT": "Linktext anzeigen",
       "DARK_MODE": "Dunkelmodus",
       "DELETION": "Löschung",
-      "DELETED_AT": "Diese Mindmap wird gelöscht am:",
+      "DELETED_AT": "Löschdatum:",
       "DELETE_AFTER_DAYS": "Mindmaps werden auf diesem Server nach {{deleteAfterDays}} Tagen gelöscht.",
       "DELETE": "Mindmap löschen",
       "CONFIRM_DELETE": "Wirklich löschen?"
diff --git a/teammapper-frontend/src/assets/i18n/en.json b/teammapper-frontend/src/assets/i18n/en.json
index 256e2944..158160f1 100644
--- a/teammapper-frontend/src/assets/i18n/en.json
+++ b/teammapper-frontend/src/assets/i18n/en.json
@@ -198,7 +198,7 @@
       "SHOW_LINKTEXT": "Show linktext",
       "DARK_MODE": "Dark mode",
       "DELETION": "Deletion",
-      "DELETED_AT": "Mindmap is going to be deleted at:",
+      "DELETED_AT": "Deletion date:",
       "DELETE_AFTER_DAYS": "Mindmaps will be deleted on this server after {{deleteAfterDays}} days.",
       "DELETE": "Delete mindmap",
       "CONFIRM_DELETE": "Delete?"
diff --git a/teammapper-frontend/src/assets/i18n/es.json b/teammapper-frontend/src/assets/i18n/es.json
index 1614a2f8..904f0cae 100644
--- a/teammapper-frontend/src/assets/i18n/es.json
+++ b/teammapper-frontend/src/assets/i18n/es.json
@@ -196,7 +196,7 @@
       "TTL": "Última fecha de eliminación conocida",
       "DARK_MODE": "Modo oscuro",
       "DELETION": "Eliminación",
-      "DELETED_AT": "Este mapa se eliminará el:",
+      "DELETED_AT": "Fecha de eliminación:",
       "DELETE_AFTER_DAYS": "Este servidor elimina los mapas después de {{deleteAfterDays}} días.",
       "DELETE": "Eliminar mapa",
       "CONFIRM_DELETE": "¿Eliminar?"
diff --git a/teammapper-frontend/src/assets/i18n/fr.json b/teammapper-frontend/src/assets/i18n/fr.json
index b1b920e1..760b513d 100644
--- a/teammapper-frontend/src/assets/i18n/fr.json
+++ b/teammapper-frontend/src/assets/i18n/fr.json
@@ -196,7 +196,7 @@
       "TTL": "Dernière date de suppression connue",
       "DARK_MODE": "Mode sombre",
       "DELETION": "Suppression",
-      "DELETED_AT": "Cette carte sera supprimée le :",
+      "DELETED_AT": "Date de suppression :",
       "DELETE_AFTER_DAYS": "Ce serveur supprime les cartes après {{deleteAfterDays}} jours.",
       "DELETE": "Supprimer la carte",
       "CONFIRM_DELETE": "Supprimer ?"
diff --git a/teammapper-frontend/src/assets/i18n/it.json b/teammapper-frontend/src/assets/i18n/it.json
index 780bda44..222a4ede 100644
--- a/teammapper-frontend/src/assets/i18n/it.json
+++ b/teammapper-frontend/src/assets/i18n/it.json
@@ -196,7 +196,7 @@
       "TTL": "Ultima data di eliminazione conosciuta",
       "DARK_MODE": "Modalità scura",
       "DELETION": "Eliminazione",
-      "DELETED_AT": "La mappa mentale verrà eliminata a:",
+      "DELETED_AT": "Data di eliminazione:",
       "DELETE_AFTER_DAYS": "Le mappe mentali verranno eliminate da questo server dopo {{deleteAfterDays}} giorni.",
       "DELETE": "Elimina la mappa mentale",
       "CONFIRM_DELETE": "Eliminare?"
diff --git a/teammapper-frontend/src/assets/i18n/ja.json b/teammapper-frontend/src/assets/i18n/ja.json
index 707ed136..fff0e421 100644
--- a/teammapper-frontend/src/assets/i18n/ja.json
+++ b/teammapper-frontend/src/assets/i18n/ja.json
@@ -198,7 +198,7 @@
       "SHOW_LINKTEXT": "リンクテキストを表示",
       "DARK_MODE": "ダークモード",
       "DELETION": "削除",
-      "DELETED_AT": "マインドマップの削除予定日：",
+      "DELETED_AT": "削除日：",
       "DELETE_AFTER_DAYS": "このサーバーのマインドマップは {{deleteAfterDays}} 日後に削除されます。",
       "DELETE": "マインドマップを削除",
       "CONFIRM_DELETE": "削除しますか？"
diff --git a/teammapper-frontend/src/assets/i18n/pt-br.json b/teammapper-frontend/src/assets/i18n/pt-br.json
index f90873c6..a9118c6a 100644
--- a/teammapper-frontend/src/assets/i18n/pt-br.json
+++ b/teammapper-frontend/src/assets/i18n/pt-br.json
@@ -196,7 +196,7 @@
       "TTL": "Última data de exclusão conhecida",
       "DARK_MODE": "Modo escuro",
       "DELETION": "Exclusão",
-      "DELETED_AT": "Este mapa será excluído em:",
+      "DELETED_AT": "Data de exclusão:",
       "DELETE_AFTER_DAYS": "Este servidor exclui os mapas após {{deleteAfterDays}} dias.",
       "DELETE": "Excluir mapa",
       "CONFIRM_DELETE": "Excluir?"
diff --git a/teammapper-frontend/src/assets/i18n/zh-cn.json b/teammapper-frontend/src/assets/i18n/zh-cn.json
index 8a112ed3..8187a425 100644
--- a/teammapper-frontend/src/assets/i18n/zh-cn.json
+++ b/teammapper-frontend/src/assets/i18n/zh-cn.json
@@ -196,7 +196,7 @@
       "TTL": "最后已知删除日期",
       "DARK_MODE": "深色模式",
       "DELETION": "删除",
-      "DELETED_AT": "思维导图将在以下时间删除：",
+      "DELETED_AT": "删除日期：",
       "DELETE_AFTER_DAYS": "思维导图将在此服务器上 {{deleteAfterDays}} 天后删除。",
       "DELETE": "删除思维导图",
       "CONFIRM_DELETE": "删除？"
diff --git a/teammapper-frontend/src/assets/i18n/zh-tw.json b/teammapper-frontend/src/assets/i18n/zh-tw.json
index 98673bbc..de38a8a3 100644
--- a/teammapper-frontend/src/assets/i18n/zh-tw.json
+++ b/teammapper-frontend/src/assets/i18n/zh-tw.json
@@ -196,7 +196,7 @@
       "TTL": "最後已知的刪除日期",
       "DARK_MODE": "深色模式",
       "DELETION": "刪除",
-      "DELETED_AT": "此心智圖的刪除日期：",
+      "DELETED_AT": "刪除日期：",
       "DELETE_AFTER_DAYS": "此伺服器會在 {{deleteAfterDays}} 天後刪除心智圖。",
       "DELETE": "刪除心智圖",
       "CONFIRM_DELETE": "確定要刪除嗎？"