Two months ago the argument was that Copilot's default model deserved a decision, because the default drives most credit consumption. Copilot Fusion changes that. When a developer picks Fusion, nobody picks the model for that task. A router does, once per prompt, and sometimes it uses more than one model.
That's a reasonable design, and GitHub's own numbers say it can be cheaper. It also means the cost of one task now depends on a routing decision you can't see in your billing data. Worth understanding before November 16, when three models retire on the same timeline.
What GitHub shipped on October 28
Copilot Fusion is generally available on all Copilot plans in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs (GitHub Changelog, October 2026). It merges Project HydraFusion's multi-model orchestration with Copilot's auto model selection. Developers pick an effort level instead of a model:
- Fusion Standard prioritizes versatile models for everyday coding.
- Fusion High prioritizes capable models for harder tasks and deeper reasoning.
- Fusion Max prioritizes the most capable models for the most demanding work.
For each prompt, Fusion chooses one of three workflows. Single: one model does the task. Cascade: an efficient model drafts, a quality check accepts the draft or escalates to a stronger model. Critique: one model drafts, a second reviews, and the first revises.
Anyone in the HydraFusion or auto research preview gets moved across automatically. Their earlier selections map onto the matching Fusion effort levels.
How Fusion is billed
The billing line is one sentence: "When you use Copilot Fusion you're billed for each model it uses, at each model's standard rate." There's no Fusion rate and no flat price per task. There's also a limited-time 30% discount on model costs for paid plans, and the announcement doesn't say when it ends.
GitHub's headline claim is that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max. It says Fusion got there at about half the average API cost per task, with about half the tokens and roughly 40% less time. That's GitHub's benchmark, reported by GitHub. We haven't reproduced it, and your own task mix will decide whether it holds for you.
The structural point holds whatever the benchmark says. Under per-model billing, the workflow Fusion picks matters about as much as which models it picks. A Cascade that escalates pays for both models. A Critique pays for at least three passes.
None of that makes Fusion a bad deal. If most everyday prompts end as accepted cascades, Fusion beats a strong-model default by a wide margin. The point is that the bill now depends on a mix of routing outcomes, and that mix shifts with the effort level. Fusion Max leans toward expensive models by design. A team defaulted to Max because it "sounds safest" will pay for that choice on every prompt.
The cost control is your model policy
This is the most useful part of the release for whoever owns the bill. Fusion only routes to models that are enabled in your enterprise and organization model policies. Enterprise admins can see which models Fusion draws from and can disable a model for Fusion without disabling it for everyone. That second part is new.
So your model policy now acts as the price ceiling for routing. If Fusion can't reach the most expensive model in the catalogue, it can't escalate to it either. An admin can keep a frontier model available for developers who pick it on purpose and still keep it out of automatic escalations.
Admins can also set Fusion as the default model, at any effort level, through enterprise managed settings (fusionTier). It works the same way as the default-model control we covered in the enterprise default model cost math, with one difference. You're no longer setting a model. You're setting a router's appetite.
Every November 16 swap is a price cut
On October 15 GitHub announced three models retiring across all Copilot experiences on November 16, 2026: chat, inline edits, ask and agent modes, and completions (GitHub Changelog, October 2026). The suggested replacements are enabled automatically for Business and Enterprise, unless an admin turned off the global default or disabled the model explicitly.
Last time the forced migration hid a rate increase: Kimi K2.7 Code moved to a model that cost 3.75x more per output token. This time the swaps go the other way.
| Retiring November 16 | In / out per 1M | Suggested replacement | In / out per 1M | Rate change |
|---|---|---|---|---|
| Claude Opus 4.8 | $5.00 / $25.00 | Claude Opus 5.5 | $4.00 / $20.00 | 20% cheaper both ways |
| Claude Opus 5 | $5.00 / $25.00 | Claude Opus 5.5 | $4.00 / $20.00 | 20% cheaper both ways |
| GPT-5.6 Terra | $2.00 / $12.00 | GPT-6.1 Sol | $2.00 / $10.00 | same input, 17% cheaper output |
| GPT-5.6 Terra | $2.00 / $12.00 | GPT-6 Luna | $0.10 / $0.50 | 20x cheaper input, 24x cheaper output |
One exception: Claude Opus 4.8 stays available to individual Copilot subscribers on annual plans, so they keep an Opus option. For Business and Enterprise it goes on November 16.
The Opus swap is the one that moves real money. Teams that defaulted to an Opus model for agent work keep that tier and pay a fifth less per token. Nothing about it needs a decision. It's the rare case where accepting the default is fine.
Our finding: the cheapest like-for-like option for GPT-5.6 Terra users isn't on GitHub's suggested list. GPT-6.1 Terra, which shipped on October 26, is priced at $1.00 input and $5.00 output per million tokens. That's half of GPT-6.1 Sol's $2.00 / $10.00 on both sides, and output is 58% below the retiring GPT-5.6 Terra's $12.00. The deprecation notice points Terra users at Sol or Luna. It doesn't mention the model that carries Terra's name. We aren't claiming 6.1 Terra matches Sol on quality; we haven't benchmarked it, and GitHub doesn't say. But anyone who chose Terra as a middle-of-the-range option should evaluate 6.1 Terra before accepting the auto-enabled Sol. On the published rates, that one choice decides whether the migration cuts output cost by 17% or by 58%.
Don't budget on the 30% discount
The Fusion discount is real money, and it's labelled "limited time" with no end date. We made the same point when Sol launched with a promotional rate: temporary model discounts are the worst baseline for a forecast, because the first full month at standard rates looks like a cost spike with no change behind it.
The model rates carry their own timers too. Gemini 4 Argon's published price is promotional until December 1, 2026, and Gemini 3.8 Flash's through December 31. If Fusion is routing to either of them, December's numbers will move twice. Fusion drops its discount at some point, and the underlying rates expire on dates you can already see.
Budget Fusion at standard rates. Treat the discount as a bonus that shows up in the actuals, not a line in the plan. This is the same lesson as falling model prices that don't shrink the bill: cheaper routing tends to turn into more usage, not a smaller invoice.
What to do before November 16
- Search your configuration for the three retiring models. Check team mappings, CLI configs, Actions workflows and any automation that names a model explicitly. After November 16 those calls stop working.
- Price the Terra replacement rather than accepting Sol by default. Compare GPT-6.1 Terra, GPT-6.1 Sol and GPT-6 Luna on a sample of your own tasks. On output, the cheapest of the three is 20x below the dearest.
- If you turned off the global default, enable the replacements yourself. In that case GitHub won't auto-enable them, and November 16 becomes an outage you scheduled.
- Decide which models Fusion may route to. Disable for Fusion anything you wouldn't want reached by automatic escalation. Keep it enabled for developers who choose it on purpose.
- Pick the effort level a default would use before anyone sets one. Standard for most teams, High or Max by exception. Same logic as a default model: size it for the common case, not for your heaviest team.
- Forecast at standard rates. Note the Argon and Flash promotional end dates, and assume the 30% Fusion discount ends without notice.
Where FinOpsAid fits
FinOpsAid reads your GitHub org read-only and shows AI credit consumption per user, per team and per cost center against your allowance, next to the rest of the GitHub bill. Here's a limit Fusion makes worse, and it's better to say it up front. The per-user AI credit figure GitHub exposes is a single daily number with no model or feature breakdown. We can show that someone's consumption went up after Fusion became the default. We can't show which models Fusion picked, which workflow ran, or how often a cascade escalated, because that data isn't in what GitHub reports. GitHub's Copilot metrics also lag roughly two to three days, so none of this is real-time. To check whether Fusion saved money, compare a team's credits per active developer before and after the switch. Related: what actually consumes GitHub AI credits, which Copilot model to default to, how to read the Copilot page, and the companion post on GitHub Enterprise now including Copilot. That model meters usage from the first request, so routing efficiency matters even more there.
Frequently asked questions
How is GitHub Copilot Fusion billed?
Copilot Fusion has no rate of its own. GitHub bills each model Fusion uses on a task at that model's standard per-token rate, so a Cascade that escalates pays for two models and a Critique pays for at least three passes. Paid plans get a limited-time 30% discount on Fusion model costs, and GitHub hasn't announced an end date.
Can administrators limit which models Copilot Fusion uses?
Yes. Copilot Fusion only routes to models enabled in your enterprise and organization model policies. Enterprise admins can see which models Fusion draws from and disable a model for Fusion alone, without removing it for developers who select it directly. Admins can also set Fusion, at any effort level, as the default through enterprise managed settings.
Which Copilot models are deprecated on November 16, 2026?
Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra retire across all Copilot experiences. GitHub suggests Claude Opus 5.5 for both Opus models, a 20% rate cut, and GPT-6.1 Sol or GPT-6 Luna for Terra. GPT-6.1 Terra, priced at $1 / $5 per million tokens, isn't on the suggested list but costs half as much as Sol.
Sources
- GitHub, "Route and orchestrate models with GitHub Copilot Fusion", retrieved 2026-11-03, https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion
- GitHub, "Upcoming deprecation of selected GitHub Copilot models in mid-November", retrieved 2026-11-03, https://github.blog/changelog/2026-10-15-upcoming-deprecation-of-selected-github-copilot-models-in-mid-november
- GitHub, "GPT-6.1 Terra in GitHub Copilot", retrieved 2026-11-03, https://github.blog/changelog/2026-10-26-gpt-6-1-terra-in-github-copilot
- GitHub Docs, "Models and pricing for GitHub Copilot", retrieved 2026-11-03, https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
- GitHub Docs, "About GitHub Copilot Fusion", retrieved 2026-11-03, https://docs.github.com/copilot/concepts/models/copilot-fusion
If you want to see credits per active developer before and after you switch a team to Fusion, connect your GitHub org — read-only and free during beta — or explore the demo dashboard first.