GitHub Universe 2026: Copilot Fusion GA and What Changed

At GitHub Universe 2026 (Fort Mason Center, San Francisco, 28–29 October), GitHub made Copilot Fusion generally available: you pick “Copilot Fusion” and an effort level in the model picker, and Copilot routes each prompt to one or more models instead of you choosing one. The same day, GitHub Enterprise customers working with an account team got every Copilot feature without a separate Copilot seat, paying only for the AI credits they use.

I read GitHub’s changelog, the Fusion and billing docs, the Octoverse report and Artificial Analysis’ benchmark pages on 2026-11-03. I have not run Fusion myself, so every number below is either GitHub’s claim or a third party’s measurement, and I label which is which.

1. Context: Two Months From Research Preview to GA

Fusion is not a surprise launch. It merges two things Copilot already had: automatic model selection (the Auto option I described in my agent-mode post) and Project HydraFusion, a multi-model orchestration experiment that GitHub’s research post put into Copilot CLI’s /experimental menu on September 4, with a candid caveat: offline results, tuned on the same benchmarks it reported.

How Fusion got here (2026)

Jun 1Copilot moves to usage-based billing: every plan consumes GitHub AI Credits, with monthly included usage per plan.

Sep 4Project HydraFusion research preview in Copilot CLI, billed per token at each underlying model’s rate.

Sep 30HydraFusion reaches VS Code and the GitHub Copilot app.

Oct 14HydraFusion Max added; the pool includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna.

Oct 28Universe day one: Fusion GA, Copilot included with GitHub Enterprise, Copilot code review Max and Copilot projects (both public preview), Octoverse 2026.

Nov 30Last day of the 30% Fusion discount on paid plans, according to the Fusion docs.

Fusion is a router, not a model

The docs say it plainly: choosing Fusion means choosing “a system that routes and orchestrates models at runtime.” It has no context window of its own (the picker shows the smallest limit among the models it may use), it doesn’t start subagents, and it only draws from models your plan and admin policies allow. If none of those are available, Fusion doesn’t appear in the picker at all.

Earlier on this blog: my Copilot agent mode workflow · Claude Opus 5.5 benchmarks

What changed since my agent-mode post: the Auto option I recommended as a cheap default is now folded into Fusion (in JetBrains, Fusion explicitly replaces Auto), the Business ($19) and Enterprise ($39) seat prices still stand, and Enterprise now has a second way to pay with no seat charge at all.

2. What Fusion Actually Does, and How It Bills

Per the Fusion docs, Fusion scores each task on reasoning, code generation, debugging and tool use, checks system health, then picks an execution pattern and the models to run it. Routing happens at cache boundaries or when the conversation’s complexity shifts, so two prompts in the same session can take different paths. You only see the final answer; intermediate drafts can be discarded.

Fusion effort levels (docs, read 2026-11-03)

What GitHub documents for each level. Fast can’t be picked by hand.

Level	What it prioritizes	Plans	Notes
Standard	Balance of cost, quality and latency; usually one model per task	All plans, including Free, Student and Pro	Only level that can use local models (MAI Code 1.1 Flash, Qwen3.8 Flash Next) in the experimental preview
High	Balanced orchestration with moderate token savings; complex bugs, multi-file changes	Pro+, Max, Business, Enterprise	Can call several models per task, so it may use more credits than Standard
Max	Frontier capability while still weighing token efficiency	Pro+, Max, Business, Enterprise	The only level Artificial Analysis benchmarked
Fast	Not user-selectable; used where fast models are required	All plans	VS Code inline chat only

The three execution patterns

Single

One model solves the task. The cheapest path, and the one Standard leans on.

Cascade

An efficient model drafts, then a judge accepts it or escalates to a stronger model. You pay for the draft even when it gets escalated.

Critique

One model drafts, several critique it in parallel, and the original revises. More calls, more latency, more credits.

Billing in one paragraph

There is no Fusion surcharge. You’re billed for each model Fusion calls, at that model’s standard rate, in AI credits; GitHub’s billing docs price one credit at $0.01 and meter input, output and cached tokens per model. Fusion tries to keep the main conversation on one model to preserve cached tokens, and helpers such as reviewers only get the context they need. According to the docs, paid plans get 30% off model costs when using Fusion through November 30, 2026; the GA changelog only says “limited time,” and the JetBrains entry says “the first 30 days.”

The documented ways to select it

Copied from GitHub’s Copilot CLI and enterprise managed-settings docs. In the VS Code, JetBrains and Copilot app pickers you choose “Copilot Fusion” and then a level; there’s nothing to type. fusionTier needs at least Copilot CLI 1.0.104-3, VS Code 1.144.0 or Copilot app 1.1.33-4.

# Copilot CLI: one-shot prompt with Fusion at an effort level (standard | high | max)
copilot --model fusion --fusion-tier standard -p "YOUR PROMPT"

# Interactive session
/model          # pick Copilot Fusion and an effort level for this session
/config model   # same choice, saved as the default for future sessions

# Enterprise managed-settings.json: Fusion by default, Standard as an overridable default
{
  "model": "fusion",
  "fusionTier": { "overridable": "standard" }
}

Discarded drafts can still touch your files

The docs list one limitation worth repeating: if Fusion discards a draft that already edited files in your workspace, those edits are not rolled back automatically. With Cascade and Critique in play, review the diff before you commit, every time.

3. GitHub’s Claims vs Independent Numbers

GitHub’s headline is that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average cost per task. That’s a vendor citing a third party, so I went to the third party. Artificial Analysis’ comparison page (Index v2.0: FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA, weighted equally) shows these numbers for the two agent setups:

Artificial Analysis Coding Agent Index v2.0 (read 2026-11-03)

AA tested Fusion through Copilot CLI at the max setting. It did not test Standard or High, and its write-up doesn’t say whether the 30% launch discount is reflected in the cost.

Metric	Copilot Fusion (max)	Claude Code, Opus 5.5 (max)
Index	60	62
FrontierCode v1.1	51%	56%
Terminal-Bench 4.0	71%	63%
SWE-Atlas-QnA-AA	57%	66%
Cost per task	$5.59	$11.42
Time per task	36.1 min	1.0 h
Tokens per task	7.3M	14.5M

How to read it

Holds up: The cost and speed claims match AA’s own pages: $5.59 vs $11.42 is about 51% cheaper, and 36.1 minutes vs an hour is roughly 40% less time. Half the tokens, too.
Hides a trade-off: The two-point gap is an average. Fusion leads Terminal-Bench by 8 points but trails by 5 on FrontierCode and by 9 on SWE-Atlas Q&A. If your work is reading and explaining large codebases, the gap is wider than the headline.
Not the top of the table: AA’s article on Fusion also lists Devin Fusion (Opus 5.5 max + GPT-6 Luna high) at 62 for $8.00 per task. Multi-model routing is a category now, not a GitHub exclusive.

AA’s write-up: GitHub Copilot Fusion orchestrates Anthropic and OpenAI models

Vendor claims I could not reproduce

GitHub’s Kyle Daigle posted on X that Fusion gives “frontier intelligence at up to 77% lower cost and up to 40% faster,” and The Next Web reports Satya Nadella making the same 40% and 77% claim on LinkedIn. The 40% matches AA’s time per task; the 77% does not match the one comparison GitHub itself cites (about 51%). Treat “up to 77%” as marketing until someone shows the baseline.

4. Five Implications for Teams

Fusion is the headline, but the changes that move budgets and review queues are spread across four other announcements. Ordered by how soon they’ll hit you:

🧭

Model choice becomes an admin policy

AI controls · Configure Copilot Fusion models

Fusion only routes to models your policies enable, skipping those excluded by data-residency or FedRAMP rules. Enterprise admins get a separate per-model policy for Fusion, and a model enabled there still needs your other model policies to allow it. Whoever owns model policy now shapes Fusion’s quality and cost.

💸

Cost per task becomes variable

High and Max can call several models

With Fusion, High and Max may draft, critique and escalate, so the same prompt can cost differently on different days. Budgets, not habits, are the control, and with the 30% discount ending November 30, December is your real baseline.

🏢

Enterprise gets a seatless option

Copilot access · GHEC · via account team

Enterprise Cloud users with a license get Copilot with no seat charge; usage is metered from the first request, with no included credits. Users without a GHEC license pay $10 a month for Copilot Standalone. Business and Enterprise seats remain, and you can’t mix them with the new model in one enterprise.

🤖

Agents now open a big share of PRs

Octoverse 2026 · September 2026

More than 40% of September’s pull requests started from a detectable agent’s first draft, up from 17% in March: 110 million that month, and GitHub calls it a floor. Agent PRs touched multiple files more often (87% vs 79%) and 90% merged within two weeks, against about 86%. No defect or rework data.

🛡️

Copilot can now approve PRs

Code review approvals · GA, off by default

Copilot code review approvals are GA and can count toward required approvals if an admin allows it. Separately, a ruleset option (public preview, on by default) requires one extra human approval for Copilot-opened PRs not attributed to a person. Those two settings pull in opposite directions; decide on purpose.

Agent figures: the Octoverse 2026 report (Oct 1, 2025 to Sep 30, 2026, detectable agents only). Approval settings: the code review changelog.

Three ways to pay for Copilot in an enterprise

From GitHub’s transition docs; 1 AI credit = $0.01.

Option	Seat charge	Included AI credits
Copilot Business	$19 per user/month	1,900 per user/month, pooled
Copilot Enterprise	$39 per user/month	3,900 per user/month, pooled
Copilot access (new)	$0 with a GHEC license; $10 per user/month as Standalone without one	None: metered from the first request

The details are in GitHub’s transition docs. Transitions are voluntary, go through a GitHub or Microsoft account team and start on the first day of a month; self-serve comes “in stages,” and GitHub Free, Team and self-service GHEC trials aren’t eligible. With no budget left, completions and next edit suggestions keep working; Chat, code review and Copilot in the cloud stop.

On the platform side, Jay Parikh’s keynote post says GitHub added 4 million CPU cores, 160 PB of high-speed storage and 50 Tb/s of network capacity this year, and expects to finish a multi-region Azure migration early next year. A separate engineering post describes a rebuilt Git storage layer with up to 35 times higher write throughput in internal benchmarks; it gives no rollout date and says you won’t need to change your workflow, with branch protections and required reviews preserved.

5. What to Do This Week

Five days, no keynote-driven migrations. Steps marked for admins need enterprise or organization owner rights.

☀️ Days 1–2: Measure Fusion against your current model

1. Pick one real, well-scoped task you’ve already solved. Run it once with your usual fixed model and once with Copilot Fusion at Standard, from the same commit, back to back. On Pro, Standard is all you get; High and Max need Pro+, Max, Business or Enterprise.
2. In Copilot CLI that’s copilot --model fusion --fusion-tier standard -p "…" against the fixed model you pick in /model. Compare the diff, the test result and the wall time, not the vibe.
3. Check the cost at github.com/settings/billing → AI usage, which breaks credits down by model (on Business and Enterprise, Copilot settings → Usage). Credits can’t be traced to a single prompt, so compare run totals; in the CLI, /chronicle cost-tips suggests token savings.

📅 Days 3–4: Admin checks before anyone gets surprised

Fusion models: in your enterprise, open AI controls → Copilot → Configure Copilot Fusion models and confirm which models Fusion may use. The effective set is the overlap with your regular model policies.
Budgets: set user-level or cost-center budgets in US dollars and decide whether “Stop usage when budget limit is reached” should be on. Confirm the “AI credits paid usage” policy in AI controls stays enabled if you want credit-consuming features to work.
Defaults: if you want Fusion everywhere, set "model": "fusion" and an overridable fusionTier in managed-settings.json. JetBrains IDEs support neither key and cloud sessions ignore fusionTier, so don’t assume one file covers every client.

📆 Day 5: Agent PR policy and the calendar

Approvals: in Settings → Copilot → Code review, decide whether “Allow Copilot approvals to count toward merge requirements” should ever be on. If yes, limit it with File paths globs (up to 15), for example to docs or generated files. Org owners can also cap “Maximum review effort.”
Rulesets: check that “Require an additional approval for unattributed Copilot pull requests” is still on in your rulesets. It’s enabled by default, in public preview, and has no effect on rulesets that require zero approvals.
Calendar: November 30 is the last day of the 30% Fusion discount. Pull your AI usage in the first week of December and compare it with this week’s numbers before anyone signs a bigger budget.

Where these settings are documented: Copilot CLI model choice · model availability · managed settings · code review settings · ruleset rules

6. Shipped, Preview, or Only Claimed

The ledger I’d hand a tech lead. Status labels are GitHub’s own; the red box lists what I couldn’t pin down.

✅

Generally available

Copilot Fusion (Standard, High, Max) in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs; Copilot code review approvals (off by default); and Copilot access included with GitHub Enterprise Cloud, through an account team.

🧪

Public preview

Copilot code review Max (Pro, Pro+, Max, Business, Enterprise; uses more credits than Lite or Balanced), Copilot projects in the Copilot app for multi-repo agent work, computer use in the app, and the ruleset rule for unattributed Copilot PRs.

🔬

Experimental

Local models under Fusion Standard (MAI Code 1.1 Flash managed on Windows; Qwen3.8 Flash Next through an OpenAI-compatible local server). Routing still happens in the cloud, so a local model doesn’t make the session offline.

📣

Announced, no date

Self-serve signup for the Enterprise-included Copilot access, the rebuilt Git infrastructure, and the end of the multi-region Azure migration (“early next year”).

What I could not confirm

Plan availability: the GA changelog says Fusion is on “all Copilot plans,” while the docs limit High and Max to Pro+ and up; I follow the docs. The discount ends November 30 in the docs, is “limited time” in the changelog and “the first 30 days” in the JetBrains entry. Impress Cloud Watch’s Japanese recap compares Fusion with “Claude Opus 4.4 Max” at about 40% lower cost, matching neither GitHub nor Artificial Analysis (Opus 5.5 Max, about half), and cites 7,200 pushes per second in testing; GitHub’s engineering post only says “up to 35 times higher write throughput.” The 77% claim has no published baseline.

Golden rule for November

Measure Fusion on your own tasks while the discount lasts, but budget at full price. And settle who may approve agent pull requests before the volume forces the question.

Conclusion

Universe 2026 moved Copilot from “pick a model” to “pick an effort level and let the router decide,” and moved Enterprise from seats to metered credits for anyone who opts in. The independent numbers back GitHub’s cost and speed story for Fusion Max, with a real quality trade-off on code comprehension.

The parts that will bite aren’t in the model picker. They’re budgets that assume last month’s prices and review rules written before agents opened four in ten pull requests. Fix those this week and Fusion becomes just another setting you can measure.

This week in one screen

Measure

• Fusion vs fixed model, same task
• AI usage per model
• December without the discount

Configure

• Fusion model policy
• Budgets and stop-at-limit
• managed-settings defaults

Decide

• Copilot approvals: on or off
• Extra approval for agent PRs
• Seats vs metered access