GitHub Copilot 將於 2026 年 11 月 16 日從模型選單移除 Claude Opus 5，Anthropic 列明退役不早於 2027 年 7 月 24 日

• Book: LLM Observability Pocket Guide
• Also by me: Thinking in Go (2-book series) — Complete Guide to Go Programming + Hexagonal Architecture in Go
• My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
• Me: xgabriel.com | GitHub

Three models disappear from the GitHub Copilot model picker on November 16, 2026: Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra. GitHub's October 15 changelog entry says the removal covers Chat, inline edits, ask and agent modes, and code completions.

None of the three is going away at the company that makes it, so the model stays and one route to it closes.

The same models stay at their makers

Anthropic's model deprecations page lists claude-opus-5 as Active, with a tentative retirement date of "Not sooner than July 24, 2027". claude-opus-4-8 is Active too, with "Not sooner than May 28, 2027". The same page says Anthropic gives at least 60 days' notice before it retires a publicly released model.

OpenAI's deprecations page does not list gpt-5.6-terra as deprecated. It appears there only as the recommended replacement for older models, such as gpt-5-mini-2025-08-07, which leaves on December 11, 2026.

So after November 16 you can still call all three from the vendors' own APIs.

What leaves Copilot, and what GitHub swaps in

Here is GitHub's list, from the October 15 entry:

There is one exception. In GitHub's words, "Claude Opus 4.8 will remain available to individual GitHub Copilot subscribers on annual plans". For annual Pro and Pro+ subscribers still on request-based billing, GitHub's annual-plan multiplier table lists Opus 4.8 at a multiplier of 27, so each request counts as 27 premium requests.

For Copilot Business and Enterprise, the changelog says the suggested alternatives are switched on automatically, unless an administrator has turned off the global default or disabled those models. If your org did that, an admin has to enable the replacements in Copilot's model policies, or they won't appear in the model picker. GitHub also says "No action is required to remove the models after they have been deprecated."

This is the second removal wave in a month. A September 18 entry set October 19 as the removal date for six other models, including GPT-5.5, GPT-5.4 and Grok 4.5. October 19 to November 16 is 28 days.

The model menu inside your agent is the vendor's shelf, and the vendor restocks it on its own schedule.

The swap mostly lowers the per-token price

Since June 1, 2026, Copilot bills usage on most plans in AI Credits consumed "according to the published API rates for each model" (GitHub blog). GitHub's models and pricing table lists these rates per million tokens (OpenAI rows are the ≤272K-input tier):

Opus 5.5 costs 20% less per input and output token than both Opus models it replaces. Terra to Sol keeps the input price and lowers output from $12 to $10. Terra to Luna is a change of class: the same table files Sol in its top category, Terra under "Versatile" and Luna under "Lightweight", and Luna's output price is a twenty-fourth of Terra's.

A cheaper token is not a cheaper task. A model that writes longer answers, reads more files or takes more turns can cost more per finished change at a lower rate. GitHub's notice makes no claim about cost per task, and I haven't measured these models against each other. Watch your AI credit spend for the first week after the switch.

Four things to check before November 16

1. Find model names written by hand. Prompt files, scripts, agent configs and team docs that name a model will not follow GitHub's automatic switch. The changelog asks you to update "workflows and integrations" before the date. Two searches cover the common spellings:

git grep -niE 'opus[ -]4[.-]8|5\.6[ -]terra'
git grep -niE 'opus[ -]5([^.0-9-]|$)'

The second pattern matches opus-5 and Opus 5 but skips opus-5-5 and Opus 5.5.

2. Check your org's model policy. If an admin turned off the global default, the replacements stay off until someone enables them in Copilot settings.

3. Re-run your evals on the replacement now. Prompts tuned against Opus 5's output may behave differently on Opus 5.5. Until November 16 you can still run the old model and the new one in Copilot on the same tasks, so this is the window for a side-by-side comparison inside the tool your team already uses. Expect to pay credits for both runs.

4. Pick Sol or Luna on purpose. GitHub offers two replacements for Terra, and its own table puts them in different categories. Luna's price makes it tempting. Try both on a few real tasks before the cheaper one becomes the default by habit.

Price per token on one side, the eval sheet on the other.

If you still need Opus 5 after November 16

Claude Code accepts full model names on the Anthropic API: its model configuration docs say the model setting, --model and ANTHROPIC_MODEL all take a full model name. Their pinning example is claude-opus-5-5; for Opus 5, use claude-opus-5 (claude-opus-4-8 works the same way):

claude --model claude-opus-5

The same docs warn that a mistyped name passed through --model or ANTHROPIC_MODEL only fails at the first request, so run one short prompt to confirm it took. For Terra, gpt-5.6-terra stays on OpenAI's API at $2 input and $12 output per million tokens for inputs up to 272K tokens, per OpenAI's pricing page.

I build Hermes IDE, so weigh this paragraph with that in mind. If your team already splits work between Copilot and Claude Code, Hermes runs GitHub Copilot CLI and Claude Code side by side, each in its own terminal with its own sign-in, and each task can get its own branch and worktree. Give the same ticket to both and compare the two branches. Sessions show the cost the agent reports, or an estimate clearly marked as estimated when Hermes knows the model's price. Hermes doesn't change which models Copilot offers. It is free and source-available (BSL 1.1), on macOS, Windows and Linux. It is not affiliated with GitHub, Anthropic or OpenAI.

Hermes 2.3.4 with two agent sessions side by side, each in its own worktree, and per-session cost and branch in the sidebar.

If the git grep above finds a hard-coded opus-5 in a shared prompt file, that is the file to fix before November 16.

If this was useful

A model swap you didn't choose is the moment tracing and evals pay for themselves: you want cost per task and pass rates from before and after, not a hunch. The LLM Observability Pocket Guide covers how to pick the tracing and eval tools that give you those numbers. If you run more than one coding agent, Hermes IDE is free to download.