Copilot retired three models yesterday and six more go on December 14. Find your model pins before they find you

GitHub Copilot removed Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra on November 16. Claude Fable 5, Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, and Kimi K3 follow on December 14. If a custom agent, CLI script, or settings file names one of them, update it now.

On this page
  1. What's gone and what replaces it
  2. Two replacements that won't just switch on
  3. Where model pins hide
  4. A ten-minute audit
  5. What the swaps cost
  6. What to do this week
  7. Sources

GitHub confirmed the November retirements in a changelog post on November 16, a month after announcing them on October 15. Three days earlier it posted the December list. Both apply across Copilot Chat, inline edits, ask and agent modes, and code completions.

This is the fourth wave since October. GitHub retired four models on October 2 and six on October 19. Counting December, that's 19 models in about ten weeks. We covered the August 31 wave separately. The pace is the story now, because any hard-coded model name in a Copilot workflow has a shelf life of a few months.

What's gone and what replaces it

These are GitHub's suggested alternatives, taken from the two changelog posts linked above:

RetiredDateSuggested alternativeException
Claude Opus 4.8Nov 16Claude Opus 5.5Stays for individual subscribers on annual plans
Claude Opus 5Nov 16Claude Opus 5.5Off the picker; still used as a runtime refusal fallback
GPT-5.6 TerraNov 16GPT-6.1 Sol or GPT-6 LunaNone
Claude Fable 5Dec 14Claude Fable 5.2None
Claude Haiku 4.5Dec 14Claude Haiku 5.5Stays for individual subscribers on annual plans
Gemini 3.8 FlashDec 14Gemini 4 HydrogenNone
Gemini 4 ArgonDec 14Gemini 4 CarbonNone
Grok 4.6Dec 14Grok 4.7None
Kimi K3Dec 14To be announcedNone

On Business and Enterprise plans with default model enablement, the alternatives turn on automatically. That doesn't happen if an admin switched off the global default or disabled a specific model, per the December notice. In that case, someone has to enable them under model policies in Copilot settings.

Two replacements that won't just switch on

Claude Fable 5.2. GitHub's weekly release notes say Fable 5.2 needs data retention by default, and its policy is off by default for Business and Enterprise. The supported models page says the same about Fable 5 and 5.1, because they aren't covered by GitHub's data retention agreement. If your org uses Fable 5 today, an admin has to opt in to 5.2 before December 14. Don't count on a quiet fallback: anything pinned to Fable 5 needs a model change, and an agent file with model-policy: required stops sending (more on that below).

Kimi K3. It has no named successor. The supported models page also notes that Kimi K3, as an open-weight model, is disabled by default for enterprises regardless of policy. Teams that chose it for cost will need to pick a new default themselves.

Where model pins hide

The changelog tells you to update “workflows and integrations” and says nothing gets removed for you. The model picker is the easy part. The places that bite are files that name a model as a string.

The Copilot CLI reference lists the order the CLI uses to pick a model, highest first:

  1. The model in a custom agent definition, if one is in use.
  2. The --model flag.
  3. The COPILOT_MODEL environment variable.
  4. The model key in ~/.copilot/settings.json (or $COPILOT_HOME/settings.json), which /model writes to.
  5. The CLI's default model.

The first item is the one to watch. A custom agent's model beats the --model flag, so fixing your CI script's flag does nothing if the agent file still pins a retired model. The custom agents reference says an agent with no model inherits the default, and doesn't say what happens when the named model is gone.

VS Code handles this better. Its custom agent file reference lets model be an ordered list, and the agent uses the first available entry. There's a catch: with model-policy: required, if none of the listed models are available, “sending is blocked” instead of falling back to Auto.

A ten-minute audit

Run this from each repo root. It searches the whole checkout and your active Copilot settings and agents folder, and catches the names as display strings and as dashed IDs, such as claude-haiku-4.5, the example ID in GitHub's CLI docs:

grep -rniE "opus[ -]?4\.8|opus[ -]?5([^.]|$)|5\.6[ -]?terra|fable[ -]?5([^.]|$)|haiku[ -]?4\.5|3\.8[ -]?flash|4[ -]?argon|grok[ -]?4\.6|kimi[ -]?k3" \
  . "${COPILOT_HOME:-$HOME/.copilot}" --exclude-dir=node_modules --exclude-dir=.git 2>/dev/null
env | grep -i COPILOT_MODEL

Then check what the grep can't see:

  • Org and enterprise agents. Custom agents stored centrally for an organization or enterprise don't live in your checkout. Review them separately.
  • CI secrets and variables. A COPILOT_MODEL set as a repository or org variable won't show up in the checkout.
  • Agent files with model-policy: required. Give each one a fallback list, or it stops sending when its only model goes.
  • Org model policies. Confirm Fable 5.2 and the other alternatives are enabled if your global default is off.
  • Prompts tuned to one model. A prompt that relied on Opus 5's habits may behave differently on 5.5. Rerun a few known tasks before trusting it.

Our view: pin a model only where reproducibility matters, such as a CI review bot. Everywhere else, leave model unset and let the default move with GitHub's list.

What the swaps cost

Copilot bills these models per token in AI credits, at $0.01 per credit. These rates come from GitHub's models and pricing page as of today, per million tokens, input then output. They don't apply to existing annual Pro and Pro+ subscribers still on request-based billing; the same page says those plans use different model multipliers.

SwapOld rateNew rateDirection
Opus 4.8 or Opus 5 to Opus 5.5$5 / $25$4 / $2020% cheaper
GPT-5.6 Terra to GPT-6.1 Sol$2 / $12$2 / $10Cheaper output
GPT-5.6 Terra to GPT-6 Luna$2 / $12$0.10 / $0.50Much cheaper, smaller model
Haiku 4.5 to Haiku 5.5$1 / $5$0.10 / $0.50 up to 100K inputCheaper
Grok 4.6 to Grok 4.7$2 / $6$2 / $6Same
Fable 5, Gemini 3.8 Flash, Gemini 4 Argon$10 / $50, $0.75 / $3.75, $2 / $10Not on the rate card yetUnknown

The GPT rates rise above 272K input tokens, both Grok models go to $4 / $12 above 200K, and Haiku 5.5's rate rises above 100K; the pricing page lists those tiers. Two old rates are promotional: Gemini 3.8 Flash through December 31 and Argon until December 1. GitHub hasn't posted Copilot rates for Fable 5.2, Hydrogen, or Carbon. Outside Copilot, Google lists Hydrogen at $1.50 / $7.50, which we covered in our Hydrogen pricing guide.

So the November wave mostly lowers costs. December is the one to budget for, since three of the six replacements have no published Copilot price. We haven't measured quality differences between any old and new pair.

What to do this week

  • If you use Copilot alone: open the model picker and pick a supported model. Annual-plan individuals keep Opus 4.8 and Haiku 4.5.
  • If you maintain agent files or CI: run the grep above, replace pins, and add fallback lists in VS Code agent files.
  • If you're an admin: check whether your global default is on. If it's off, enable the alternatives, and decide on Fable 5.2's data retention policy before December 14.

Sources

Researched and drafted by an AI agent from the linked sources. How posts are made.