GitHub Copilot Fusion and auto model routing: what it costs and how to control it (Nov 2026)
Shared by Pavel · Updated 2026-11-16
As of 2026-11-16. Every source below was read on 2026-11-16. Labels: vendor = GitHub/Microsoft docs, changelog or staff statement; measured: <username> = a user's own numbers, not verified; opinion = a user's or commenter's view; community = a non-staff commenter's explanation (checked against vendor docs where stated).
Terms: since 2026-06-01 Copilot bills in "GitHub AI Credits" (1 credit = $0.01), not premium requests with multipliers. Multipliers survive only for legacy annual Pro/Pro+ plans. "Premium-request credits" means AI credits today.
Questions people are asking (2026-11-02 to 2026-11-16)
Staff = GitHub/Microsoft employee or bot with authority. All community threads below have only the automatic "feedback submitted" bot reply from GitHub plus non-staff users; no thread is marked Answered. Staff answered one thread (11).
- Why does Fusion use Opus for a commit/push, and where is my 30% discount shown in usage? (https://github.com/orgs/community/discussions/211890, rivie13, 11-02, active to 11-16; also wiredrabbit666, Tejareddy94, martintain3409-byte, drdrummie). Answer: GitHub has not said. No routing explanation; the discount is not visible in the usage view per the asker, and the docs do not show it as a line. drdrummie (11-16) reports about 50 PRs for about 2,000 credits and one Fusion High review over 1,700 credits (their numbers).
- Can I allow/block models (never Fable) and get a warning before a costlier route? (https://github.com/orgs/community/discussions/211933, wiredrabbit666, 11-03). Answer: no documented per-user block list or pre-switch warning. Community answer (aungzayyanaing, 11-12): model policy is org/enterprise admin only; an individual can pick a named model instead of Fusion. Feature request still open.
- Why was "Kimi K3 the only option in the model pool" with 13 models visible, and why 75% of the monthly quota in one day? (https://github.com/orgs/community/discussions/212119, Caramel-Andrew, 11-05; also jpvignolo). Answer: GitHub has not said. Per-task eligibility criteria are not published; the cause of the burn is unproven. Community answer 11-12.
- Fusion wasted about 80% of my credits; why does it pick expensive models (Fable 5) for frontend work Sol/Luna did fine? (https://github.com/orgs/community/discussions/212238, NeoSparkTechnology, 11-06; markusswedpal). Answer: GitHub has not said. No answer in the thread.
- My Pro+ picker shows only "HydraFusion"; where are my models? (https://github.com/orgs/community/discussions/212391, sunniewang, 11-08). Answer: HydraFusion is the old name of the Fusion preview; users who had auto/HydraFusion selected were moved to matching Fusion levels (vendor, changelog 10-28). Community answers 11-08 and 11-12; asker never confirmed.
- Pro: 1,500 of 1,500 AI credits gone almost at once; is that expected for agent mode? (https://github.com/orgs/community/discussions/213140, GeeckLance, 11-14, adjacent). Answer: GitHub has not said; support ticket unanswered 6 days per the asker. Whether any usage predates activation is for Support only.
- How do I reclaim credits burned by a failed cloud-agent session? (https://github.com/orgs/community/discussions/211874, fkpama, 11-02, adjacent). Answer: GitHub has not said in the thread; the thread points to a support ticket.
- Let me control which models Fusion uses (it favours GPT). (https://github.com/microsoft/vscode/issues/345596, ronmichael, 11-02). Answer: individuals have no documented setting; enterprise admins can set status per model for Fusion (see Budgets and admin controls). Open feature request.
- Let me pick a model family, or per complexity, for Fusion. (https://github.com/microsoft/vscode/issues/346470, l-kratochvil, 11-06; erikweisz29). Answer: GitHub has not said. Open feature request.
- Let me turn off Fusion / go back to Auto, or limit which models it routes to. (https://github.com/microsoft/vscode/issues/345808, Tejareddy94, 11-03). Answer: GitHub has not said in the thread (only a version-check bot). A community commenter says picking a named model avoids Fusion (unverified, see issue 345570); in VS Code the opt-in setting
chat.copilot.dynamicAutoTiers.enabledgates Fusion (see How to see and pin the model). - Fusion High/Max does not say which model was picked. (https://github.com/microsoft/vscode/issues/346353, adrianlop, 11-05). Answer, staff (alexdima, GitHub, 11-06): "works as designed"; Standard names one model, High/Max may orchestrate several. Kept open as a feature request (needs 20 upvotes in 60 days).
- I selected Claude Opus 5.5 but Fusion replied "selected GPT-6 Luna"; how do I make an explicit selection win? (https://github.com/microsoft/vscode/issues/345570, ivolopeseuronext, 11-02). Answer: GitHub has not said. The asker cites a local log showing a change to claude-opus-5.5, then Auto resolving to gpt-6-luna.
- Cache-read tokens drop on the first call of a new user message in Fusion (cache write 49k tokens). (https://github.com/microsoft/vscode/issues/346031, ralphgee, 11-04). Answer: GitHub has not said. Open, measured, bug label "new release".
Other credit threads not about Fusion: 212014/212015 (refund ticket), 213113 (cannot edit budget).
What auto routing and Fusion do
- Fusion is "a runtime model router and orchestrator": you choose it in the picker; it picks models and an execution pattern per request and returns one response. It has no context window of its own (the displayed window is a conservative minimum across the models used). [vendor] https://docs.github.com/en/copilot/concepts/models/copilot-fusion
- Effort levels: Standard (cost, quality, latency; usually one model), High (balanced orchestration, moderate token savings), Max (frontier capability, still considers efficiency). Fast exists only in VS Code inline chat and is not user-selectable. [vendor, same page]
- Patterns: Single (one model), Cascade (a cheap model drafts, a judge accepts or escalates), Critique (one model drafts, several critique in parallel, the original revises). Chosen per prompt, so one session can mix them. "Review calls are added only when a task warrants them." [vendor, same page]
- Selection inputs: capability signals for reasoning, code generation, debugging and tool use, plus real-time health and availability. Routing happens at natural cache boundaries and large complexity shifts. [vendor, same page]
- The pool excludes models not in your plan, excluded by admin policy, by data residency/FedRAMP, or by evaluation-model restrictions. [vendor, same page] The per-task eligibility rules are not published (a community commenter says the same, https://github.com/orgs/community/discussions/212119).
- Fusion replaced Auto in some surfaces. Changelog 2026-10-28: Fusion generally available in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains; it "combines Project HydraFusion's multi-model orchestration with Copilot auto model selection"; users with auto/HydraFusion selected moved to matching Fusion levels; "Support for additional clients is coming soon." [vendor] https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion (A community commenter says GA was Oct 22; the changelog says Oct 28.)
- Classic Auto (still documented): picks per task, tiers Efficiency / Balance / Intelligence, "all tiers use the same models"; changelog 2026-09-14. [vendor] https://docs.github.com/en/copilot/concepts/models/auto-model-selection and https://github.blog/changelog/2026-09-14-configure-cost-and-quality-in-copilot-auto-model-selection
- Limitation: Fusion does not start subagents; if a discarded draft already edited files, those edits are not undone automatically. [vendor, Fusion page]
- In VS Code, Fusion shows in the Copilot harness when
chat.copilot.dynamicAutoTiers.enabledis on and the account qualifies; otherwise the picker keeps showing Auto. [vendor] https://code.visualstudio.com/docs/agent-customization/language-models
Prices and what it costs
- Current model: per-token billing in AI credits, 1 credit = $0.01. Each plan includes credits; beyond that you pay at per-token rates. [vendor] https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
- Fusion is "billed for each model it uses, at each model's standard rate"; there is no separate Fusion fee. High/Max can use several models, so a task "may cost more AI credits than on Fusion Standard." [vendor, Fusion page and changelog]
- Discount: paid plans get 30% off model costs on Fusion "through November 30, 2026"; 10% off on classic Auto "in Copilot Chat, or Copilot cloud agent". The changelog calls Fusion's discount "for a limited time"; the docs give 2026-11-30. [vendor] https://docs.github.com/en/copilot/concepts/billing/usage-based-billing-for-individuals . Arithmetic (not a vendor figure): when a 30% discount ends, the same usage costs about 43% more on December 1 (1/0.7). No extension was found as of 2026-11-16.
- Included credits per month: Pro ($10) = 1,500 (1,000 base + 500 flex); Pro+ ($39) = 7,000 (3,900 + 3,100); Max ($100) = 20,000 (10,000 + 10,000). Free and Student have an allowance and auto model selection only (amount not stated on that page). Credits reset at 00:00 UTC on the 1st, no carry-over. Business 1,900 and Enterprise 3,900 per license, in a shared pool. [vendor] usage-based-billing-for-individuals page above; Business/Enterprise figures from https://docs.github.com/en/copilot/concepts/billing/budgets-for-usage-based-billing
- Per-1M-token list prices (input / cached input / output, USD) of the models named in the complaints [vendor, models-and-pricing page; GPT rows are the default tier up to 272K input tokens, higher long-context rates exist]:
  * Claude Fable 5: $10 / $1 / $50   * Claude Fable 5.2: $8 / $0.20 / $40   * Claude Opus 5.5: $4 / $0.20 / $20   * GPT-6 Astra: $10 / $1 / $50   * GPT-6 Luna: $0.10 / $0.01 / $0.50   * GPT-6.1 Sol: $2 / $0.10 / $10   * Kimi K3: $3 / $0.30 / $15   * Claude Sonnet 5.5: $2 / $0.10 / $10   * Spread from Luna to Fable 5: 100x on input, 100x on output.
- Code completions and next edit suggestions do not consume credits. [vendor, models-and-pricing]
- Legacy premium-request multipliers apply only to Pro/Pro+ annual subscribers who stayed on request billing after 2026-06-01. The page lists Opus 4.8 = 27, Sonnet 4.6 = 9, GPT-5.3-Codex = 6, Haiku 4.5 = 0.33, MAI-Code-1.1-Flash = 0.25; no Fable or Fusion entries; auto gets 0.9x on a 1x model. [vendor] https://docs.github.com/en/copilot/reference/copilot-billing/request-based-billing-legacy/model-multipliers-for-annual-plans
- Copilot code review uses an undisclosed auto-selected model; estimated averages Lite $0.25, Balanced $1.50, Max $10 per review. [vendor, models-and-pricing]
How to see and pin the model
- See the model: hover the effort-level label under a Fusion response to see the models used; intermediate drafts are not shown. [vendor, Fusion page; VS Code 1.144 notes https://code.visualstudio.com/updates/v1\_144] In Standard the response text names the model ("Copilot Fusion selected GPT-6 Luna since the task has low complexity", reproduced by alexdima, [vendor-staff] https://github.com/microsoft/vscode/issues/346353). In High and Max the UI shows only the effort level (staff: works as designed; feature request open).
- Classic Auto / any model: hover the response in VS Code to see the model; on-device responses are labelled On-device with 0 credits. [vendor] https://code.visualstudio.com/docs/agent-customization/language-models
- Per-model spend after the fact: github.com/settings/billing, "AI usage": included vs additional credits, a chart of models, per-model credits and cost. The IDE status bar Copilot icon shows quota and reset date. Limit: you "generally can't trace credits to specific prompts, sessions, or features." [vendor] https://docs.github.com/en/copilot/how-tos/manage-and-track-spending/monitor-ai-usage
- Not documented: the Chat Debug view docs do not say they show the model per request (https://code.visualstudio.com/docs/copilot/chat/chat-debug-view). One user cites a local Copilot log line recording the model switch (https://github.com/microsoft/vscode/issues/345570, [measured: ivolopeseuronext], unverified).
- Pin a model: pick a named model in the picker instead of Fusion. "A model selector you choose in the model picker, not a mode you are forced into" is the community reading (aungzayyanaing, discussion 211933); vendor support: the changelog says you choose Fusion in the picker. Counter-evidence: issue 345570 reports a named Opus 5.5 selection that still returned a Fusion-routed answer; unanswered by staff. "A named model is always honoured" is unverified.
- "Pinned" in the VS Code picker only means the favorites list (hover, pin icon); hide models via the eye icon in the Language Models editor. These do not restrict routing. [vendor, language-models page]
- Fusion in VS Code is gated by
chat.copilot.dynamicAutoTiers.enabled; a user on the Copilot harness can turn it off to stop being offered Fusion (an inference from the doc, not stated as a cost control; the VS Code page says enabling the setting "doesn't grant access" to Fusion tiers). [vendor, language-models page] - No documented setting for "never use model X" for an individual. The community says the same; vendor docs show no such setting.
Budgets and admin controls
Personal accounts [vendor, https://docs.github.com/billing/tutorials/set-up-budgets]: github.com/settings/billing, Budgets and alerts, New budget; types Product, SKU (e.g. "Copilot AI credits"), or "Bundled AI credits budget" (all AI-credit SKUs); scope a repo or the whole account; optional "Stop usage when budget limit is reached" (if cleared, email only, usage continues); alerts at 75%, 90% and 100% by email and banner; optional included-usage alerts at 90%/100%.
The documented recipe to avoid charges beyond your plan (personal account): create or edit a Bundled AI credits budget, set it to $0 USD, and enable "Stop usage when budget limit is reached". [vendor] https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing . A $0 budget alone, without Stop usage, is not documented as a stop.
- The docs conflict, as quoted on 2026-11-16:
  * (a) Budgets concept page (written for Business/Enterprise admins): "'Stop usage when budget limit is reached' applies to enterprise spending limits, cost center budgets, and organization budgets only, and is off by default... User-level budgets always enforce a hard stop and do not have this setting." and "A $0 USD ULB immediately blocks features that consume AI credits." [vendor] https://docs.github.com/en/copilot/concepts/billing/budgets-for-usage-based-billing   * (b) Personal tutorial: personal budget types are Product, SKU or Bundled AI credits; "select Stop usage when budget limit is reached, if available. This option is not available for user-level budgets, which always enforce a hard stop." and "If you do not select Stop usage when budget limit is reached, you will be notified by email if you exceed your budget, but usage will not be stopped." [vendor] https://docs.github.com/billing/tutorials/set-up-budgets   * (c) Models-and-pricing page: "To avoid charges beyond your plan, create or edit a Bundled AI credits budget, set it to $0 USD, and enable Stop usage when budget limit is reached." [vendor]   * Reading: the "always hard-stops" text is about enterprise per-user budgets, a scope personal accounts do not have. For personal accounts the docs treat Stop usage as a checkbox, and (c) says to tick it. Whether a personal account with no budget can incur overage is not stated (the individuals page says you can continue "by setting a budget for additional usage").
- A commenter on discussion 213140 mentions "other discussions about an existing $0 AI credits budget" failing to save; pramaanz (discussion 213113) cannot edit a budget. [community / measured, unverified]
- Exhausted credits: upgrade (pay the difference), pay for additional usage under a budget, or wait for the reset. The docs do not state whether additional usage is on by default. [vendor] usage-based-billing-for-individuals page
- Session limits to cap agent task spend are mentioned in the docs, with the how-to linked ("Optimizing your AI usage"); not covered here. [vendor mention]
- Org/enterprise budgets: user-level budgets (individual, cost center, universal) cap one user's use including included credits; cost-center, org and enterprise budgets cap metered charges; included-usage control caps the pool draw (1,900 per Business, 3,900 per Enterprise license). [vendor] budgets page above
- Admin model policy: Enterprise owners: AI controls, Copilot, Configure models: Enabled/Disabled/Delegate per model; targeted model rules; "Configure Copilot Fusion models" lets an admin set status per model for Fusion; "Disabling a model required by Fusion Max makes Fusion Max unavailable." Org owners control only delegated models. Enterprise managed settings can set Fusion at any effort level as the default (changelog 10-28). Admin-page wording: "Frontier models are often used by Fusion Max. If you disable any of the models required for Fusion Max, then Fusion Max will not be available for use." [vendor] https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/manage-availability-of-default-models ; https://docs.github.com/en/copilot/how-tos/copilot-on-github/set-up-copilot/configure-access-to-ai-models
What users measured
All user numbers are unverified; none reproduced by GitHub.
- [measured: wiredrabbit666, 11-03] One day used about as many tokens as a month before, with an unchanged prompt; "I think it decided everything needed fable" is their guess. https://github.com/orgs/community/discussions/211890 (repeated in their own thread 211933)
- [measured: Caramel-Andrew, 11-05] 0% to about 75% of the monthly allowance in a day on Pro, VS Code 1.145.0, Chat 0.73.0; the agent re-read the same file repeatedly. https://github.com/orgs/community/discussions/212119
- [measured: drdrummie, 11-16] About 50 PRs for about 2,000 credits; one review in Fusion High with the 30% discount over 1,700 credits; a Fusion Standard grub analysis picked Astra for about 150 credits; "Luna would take less then 2" is their estimate, not a measurement. https://github.com/orgs/community/discussions/211890
- [measured: markusswedpal, 11-06] Credits gone in two days after Fusion picked Fable 5; Sol and Luna handled the same tasks before. https://github.com/orgs/community/discussions/212238
- [measured: ralphgee, 11-04] In Fusion, the first call after a new user message shows cache read 30,079 (vs 76,850 on the previous call) and cache write 49,475 tokens, before the logged 300-second expiry; a second sample for GPT-6.1 Terra. https://github.com/microsoft/vscode/issues/346031 . Relevant because cache write costs more than cache read on the vendor table (Opus 5.5 write $5 vs read $0.20 per 1M).
- No independent cost benchmark found. GitHub's own claim, measured and reported by GitHub (not independently reproduced, no methodology link in the post): in the Artificial Analysis Coding Agent Index, Fusion Max scored 60, "within two points of Claude Code with Opus 5.5 Max", at about half the average API cost per task, about half as many tokens and about 40% less time. [vendor claim] changelog 2026-10-28.
- [opinion] "pure gambling", "intent grading is bad", "feels like a way to increase usage-based billing" (drdrummie, markusswedpal). GitHub (alexdima) offered no usage figures.
Open questions (no primary source answers)
- The per-task eligibility rules for the "available model pool" (why Kimi K3 only). Not published.
- Does selecting a named model always bypass Fusion? Issue 345570 says no for one user; unanswered.
- Whether the 30% Fusion discount appears as a line in the usage view (asked by rivie13 and martintain3409-byte); the docs do not show it.
- Which models handled High/Max turns (staff: feature request, needs 20 upvotes).
- A per-user "never use model X" or pre-escalation warning: no documented setting; individual users have no admin policy.
- Whether a personal $0 bundled AI credits budget plus Stop usage is the exact hard cap (the docs conflict above).
- What happens to Fusion costs after the 30% discount ends on 2026-11-30 (no announced extension found).
- Cache-write spikes in Fusion (issue 346031): no staff response.
- Why the burn: Fusion routing, extra review/critique calls, or agent loops? No GitHub breakdown; community commenters say all three are plausible.
Sources (all read 2026-11-16)
Vendor:
- https://docs.github.com/en/copilot/concepts/models/copilot-fusion
- https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion
- https://docs.github.com/en/copilot/concepts/models/auto-model-selection
- https://github.blog/changelog/2026-09-14-configure-cost-and-quality-in-copilot-auto-model-selection
- https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
- https://docs.github.com/en/copilot/concepts/billing/usage-based-billing-for-individuals
- https://docs.github.com/en/copilot/concepts/billing/budgets-for-usage-based-billing
- https://docs.github.com/billing/tutorials/set-up-budgets
- https://docs.github.com/en/copilot/how-tos/manage-and-track-spending/monitor-ai-usage
- https://docs.github.com/en/copilot/reference/copilot-billing/request-based-billing-legacy/model-multipliers-for-annual-plans
- https://docs.github.com/en/copilot/how-tos/administer-copilot/manage-for-enterprise/manage-availability-of-default-models
- https://docs.github.com/en/copilot/how-tos/copilot-on-github/set-up-copilot/configure-access-to-ai-models
- https://code.visualstudio.com/docs/agent-customization/language-models
- https://code.visualstudio.com/updates/v1\_144
- https://code.visualstudio.com/docs/copilot/chat/chat-debug-view
Threads:
- https://github.com/orgs/community/discussions/211874
- https://github.com/orgs/community/discussions/211890
- https://github.com/orgs/community/discussions/211933
- https://github.com/orgs/community/discussions/212119
- https://github.com/orgs/community/discussions/212238
- https://github.com/orgs/community/discussions/212391
- https://github.com/orgs/community/discussions/213140
- https://github.com/orgs/community/discussions/212411
- https://github.com/microsoft/vscode/issues/345596
- https://github.com/microsoft/vscode/issues/345570
- https://github.com/microsoft/vscode/issues/345808
- https://github.com/microsoft/vscode/issues/346031
- https://github.com/microsoft/vscode/issues/346353
- https://github.com/microsoft/vscode/issues/346470