github-copilot: session cost ignores cache writes and long-context pricing #52089

Activity

1. github-actions

   assigned

   kitlangton

   on Sep 29, 2026
2. github-actions commented on Sep 29, 2026

   

   github-actionsBot

   on Sep 29, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • #51845: Reports GitHub Copilot cache-write tokens omitted from session cost estimates, including Claude Opus 5.5.
   • #42910: Reports session cost estimates that fail to apply higher context-tier pricing.
3. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   Contributor

   More actions

   Thanks for the detailed report. I've confirmed both parts on 2.0.19 and on the latest dev (86ce19d), using a local mock of Copilot's /models and /responses endpoints built from the payloads in your report:

   • Prices: for a /models entry with cache_write_price: 500 and a long_context block, the model ends up with cost.cache.write: 0 and no long-context tier.
   • Usage: for a /responses result reporting cache_write_tokens: 8270, both the streaming and non-streaming paths return cacheWrite as undefined and noCache: 8273, so cache writes are billed at the plain input price.

   The script below reproduces it without credentials. The Chat Completions adapter has the same cacheWrite: undefined code, though I didn't run it. Related: #35765, which would prefer Copilot's own reported billed amount when it is available.

   Reproduction script (repro.ts)

   // Reproduces #52089: Copilot cache-write / long-context prices and cache_write_tokens are dropped.
   // Run (after `bun install` in the checkout):
   //   cd <checkout>/packages/core && REPO=<checkout> bun /path/to/repro.ts
   // Uses a local mock of Copilot's /models and /responses endpoints; no credentials needed.
   const REPO = process.env.REPO ?? "/workspace/repo"
   const { CopilotModels } = await import(`${REPO}/packages/core/src/github-copilot/models.ts`)
   const { createOpenaiCompatible } = await import(`${REPO}/packages/core/src/github-copilot/copilot-provider.ts`)
   
   const remoteModel = {
     model_picker_enabled: true,
     id: "gpt-5.6-sol",
     name: "GPT-5.6 Sol",
     version: "gpt-5.6-sol-2026-08-01",
     supported_endpoints: ["/responses"],
     billing: {
       token_prices: {
         batch_size: 1000000,
         default: { input_price: 400, cache_read_price: 20, cache_write_price: 500, output_price: 2000, max_prompt_tokens: 272000 },
         long_context: { input_price: 800, cache_read_price: 80, cache_write_price: 1000, output_price: 3000, max_prompt_tokens: 922000 },
       },
     },
     capabilities: {
       family: "gpt-5.6",
       limits: { max_context_window_tokens: 922000, max_output_tokens: 128000, max_prompt_tokens: 922000 },
       supports: { tool_calls: true, streaming: true },
     },
   }
   
   const usage = {
     input_tokens: 8273,
     input_tokens_details: { cache_write_tokens: 8270, cached_tokens: 0 },
     output_tokens: 10,
     output_tokens_details: { reasoning_tokens: 0 },
     total_tokens: 8283,
   }
   const response = {
     id: "resp_1",
     object: "response",
     created_at: 1760000000,
     model: "gpt-5.6-sol",
     status: "completed",
     output: [{ type: "message", id: "msg_1", role: "assistant", status: "completed", content: [{ type: "output_text", text: "hi", annotations: [] }] }],
     usage,
   }
   
   const server = Bun.serve({
     port: 0,
     async fetch(req) {
       const url = new URL(req.url)
       if (url.pathname === "/models") return Response.json({ data: [remoteModel] })
       if (url.pathname === "/responses") {
         const body = await req.json()
         if (!body.stream) return Response.json(response)
         const events = [
           { type: "response.created", response: { ...response, status: "in_progress", output: [], usage: null } },
           { type: "response.output_item.added", output_index: 0, item: { type: "message", id: "msg_1", role: "assistant", status: "in_progress", content: [] } },
           { type: "response.output_text.delta", item_id: "msg_1", output_index: 0, content_index: 0, delta: "hi" },
           { type: "response.output_item.done", output_index: 0, item: response.output[0] },
           { type: "response.completed", response },
         ]
         const text = events.map((e) => `data: ${JSON.stringify(e)}\n\n`).join("") + "data: [DONE]\n\n"
         return new Response(text, { headers: { "content-type": "text/event-stream" } })
       }
       return new Response("not found", { status: 404 })
     },
   })
   const base = `http://localhost:${server.port}`
   
   const snapshot = await CopilotModels.load(base, {})
   const models = CopilotModels.derive(base, snapshot, [])
   const model = [...models.values()][0]
   console.log("model.cost (remote: write=$5/M default, long_context tier $8/$0.80/$10/$30 above 272k):")
   console.log(JSON.stringify(model.cost, null, 2))
   
   const lm = createOpenaiCompatible({ baseURL: base, apiKey: "x", name: "copilot" }).responses("gpt-5.6-sol")
   const prompt = [{ role: "user", content: [{ type: "text", text: "hello" }] }]
   const gen = await lm.doGenerate({ prompt } as any)
   console.log("doGenerate inputTokens (remote: cache_write_tokens=8270):", JSON.stringify(gen.usage.inputTokens))
   const stream = await lm.doStream({ prompt } as any)
   const reader = stream.stream.getReader()
   for (;;) {
     const { done, value } = await reader.read()
     if (done) break
     if ((value as any).type === "finish") console.log("doStream inputTokens:", JSON.stringify((value as any).usage.inputTokens))
   }
   server.stop(true)

   Run it: bun repro.ts
4. opencodejr

   added

   reproduced

   on Oct 28, 2026
5. golfromeo-fr commented on Oct 28, 2026

   

   golfromeo-fr

   on Oct 28, 2026

   More actions

   Opened #59803 covering both halves of this issue: the builder now parses cache_write_price and the optional long_context tier (its own prices and max_prompt_tokens) from the /models payload and emits a context-tier cost entry — SessionUsage.calculateCost already selected tiers per request, so no calculator change was needed. Both usage adapters map cache_write_tokens into cacheWrite instead of dropping it, and the non-cached input stays consistent with the total. Tested against the Claude Haiku 5.5 rate card, including the ≤100K boundary.
6. rekram1-node commented on Nov 10, 2026

   

   rekram1-node

   on Nov 10, 2026

   Collaborator

   More actions

   Fixed in v2.2.7 by #62034. Session cost for GitHub Copilot now uses the amount Copilot reports it charged for each request (copilot_usage) instead of the catalog estimate, so cache writes and long-context pricing are included. If you still see a cost that doesn't match Copilot's billing on 2.2.7, let me know and I'll reopen.

   ❤️React with ❤️1Reacted by Golf R0meo
7. rekram1-node

   closed this as completedon Nov 10, 2026