Copilot Drops Opus 5 and Terra: Narrower Model Access, API Still Alive

Copilot Drops Opus 5 and Terra on Nov 16

GitHub Copilot's model picker is about to get narrower. According to a GitHub changelog published on October 15, starting November 16, 2026, three models—Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra—will be removed from Copilot's Chat, inline editing, ask and agent modes, as well as code completion. Interestingly, all three models remain available in their respective vendors' APIs—the models aren't dead, it's just the Copilot path that's closing.

Vendors Keep Them Alive; Only Copilot Closes the Window

Anthropic's model deprecation page shows claude-opus-5 with status Active and a tentative retirement date of "no earlier than July 24, 2027"; claude-opus-4-8 is likewise Active, with tentative retirement no earlier than May 28, 2027. Anthropic commits to giving at least 60 days' notice for publicly released models.

OpenAI's deprecation page doesn't list gpt-5.6-terra as deprecated at all—it only appears as a recommended replacement for older models (such as gpt-5-mini-2025-08-07, scheduled for shutdown on December 11, 2026). In other words, after November 16, you can still call all three models through the three vendors' native APIs.

Model "deprecation" and "retirement" are two different stages. During deprecation, vendors typically mark a model as legacy while offering recommended replacements, but API calls still work; once retirement kicks in, requests either return errors or are automatically routed to replacements. Anthropic's deprecation page distinguishes statuses as "Active," "Deprecated," and "Retired," and the phrase "no earlier than a certain date" means that date is a floor, not a committed deadline—vendors may still move it earlier or later. As a downstream integrator, GitHub's model picker availability is governed by its own engineering maintenance, cost structure, and vendor agreements, and doesn't stay in sync with upstream API lifecycles. So "Copilot drops it but the API still works" isn't contradictory—it's just a difference in cadence between two product lines.

What Is GitHub Replacing Them With? One Table Says It All

GitHub's replacement plan is as follows: Claude Opus 4.8 and Claude Opus 5 are both succeeded by Claude Opus 5.5; GPT-5.6 Terra is replaced by either GPT-6.1 Sol or GPT-6 Luna.

There's one exception: Claude Opus 4.8 will remain available to individual Copilot users on annual subscription plans. For annual Pro and Pro+ subscribers still on request-based billing, GitHub's annual plan multiplier table shows Opus 4.8 has a multiplier of 27, meaning each request consumes 27 premium requests.

For Copilot Business and Enterprise users, the changelog states that recommended replacement models are enabled automatically unless an administrator has turned off global defaults or disabled those models. If an organization has done so, an administrator needs to manually enable the replacement models in the Copilot model policy, or they won't appear in the model picker. GitHub also states: "No action is required to remove a model after deprecation."

This is already the second wave of removals within a month. The September 18 changelog set October 19 as the removal date for six other models, including GPT-5.5, GPT-5.4, and Grok 4.5. From October 19 to November 16, the gap is only 28 days.

How Do Prices Change After Switching Models?

Since June 1, 2026, most Copilot plans bill in AI Credits based on "each model's published API rates." GitHub's models and pricing table shows (OpenAI rows are for the ≤272K input tier):

Model	Input	Cached Input	Output
Claude Opus 4.8	$5.00	$0.50	$25.00
Claude Opus 5	$5.00	$0.50	$25.00
Claude Opus 5.5	$4.00	$0.20	$20.00
GPT-5.6 Terra	$2.00	$0.20	$12.00
GPT-6.1 Sol	$2.00	$0.10	$10.00
GPT-6 Luna	$0.10	$0.01	$0.50

Opus 5.5's input and output prices are 20% cheaper than both Opus models it replaces. Terra to Sol keeps the input price the same while dropping output from $12 to $10. Terra to Luna is an entirely different order of magnitude: the same table groups Sol under the top category, Terra under "Versatile," and Luna under "Lightweight," and Luna's output price is just one twenty-fourth of Terra's.

But cheaper tokens don't equal cheaper tasks. A model that produces longer outputs, reads more files, or runs more turns can make each completed change more expensive even at a lower unit price. GitHub's announcement makes no claims about "cost per task" and hasn't published comparative evaluation data for these three models. It's worth watching your AI credit consumption closely during the first week after the switch.

AI Credits is a billing unit GitHub introduced in June 2026, replacing the earlier fixed-multiplier model based on "premium requests." It converts the number of tokens consumed per interaction into credits at each model's published API rates, making costs directly comparable across models. Cached input refers to tokens that hit the prompt cache, typically priced at about one-tenth of standard input—this matters a lot for agent workflows that repeatedly send the same system prompt or long context. Input tiers refer to tiered pricing based on the total input tokens in a single request; for example, OpenAI's ≤272K tier and higher tiers have different rates. Understanding these three concepts is essential to accurately estimating your actual bill after switching models.

Four Things You Must Check Before November 16

1. Find hardcoded model names. Model names hardcoded in prompt files, scripts, agent configs, and team docs won't follow GitHub's automatic switch. The changelog asks you to update "workflows and integrations" before the deadline. These two greps cover the common spellings:

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

The second regex will match opus-5 and Opus 5, but skip opus-5-5 and Opus 5.5.

2. Check your organization's model policy. If an administrator has turned off global defaults, replacement models will stay off until someone enables them in Copilot settings.

3. Rerun your evals on the replacement models right away. Prompts tuned for Opus 5's outputs may behave differently on Opus 5.5. Before November 16, you can still run both old and new models on the same tasks in Copilot—this is your last window for side-by-side comparison within the tools your team already uses. Note that both runs will consume credits.

4. Choose Sol or Luna deliberately. GitHub offers two replacement options for Terra, and its own table groups them into different categories. Luna's price is tempting. Try both on a few real tasks before it becomes the default out of inertia.

What If You Still Need Opus 5 After November 16?

Claude Code accepts Anthropic API's full model names. Its model configuration docs state that the model setting, the --model argument, and the ANTHROPIC_MODEL environment variable all accept full model names. The official pinned example is claude-opus-5-5; to pin Opus 5, use claude-opus-5 (and claude-opus-4-8 likewise):

claude --model claude-opus-5

The docs also warn that a misspelled model name passed via --model or ANTHROPIC_MODEL only errors out on the first request, so be sure to run a short prompt first to confirm it takes effect. As for Terra, according to OpenAI's pricing page, gpt-5.6-terra remains at $2 input / $12 output per million tokens on the OpenAI API (for inputs under 272K tokens).

Claude Code is Anthropic's official command-line coding agent tool that calls the Anthropic API directly, bypassing GitHub Copilot's model routing layer. Its --model argument, the model field in the config file, and the ANTHROPIC_MODEL environment variable have different priorities—the environment variable is usually lowest priority, and the command-line argument is highest. Using full model names (like claude-opus-5) rather than aliases avoids unexpected switching when a vendor points an alias at a newer model. The design where typos only error on the first request means a single misspelling can cause an entire batch job to fail partway through—so verifying with a short prompt after switching is a necessary step.

A Pragmatic Reminder

A model shift is an opportunity to reexamine your cost structure. When you're forced to switch models without choosing to, that's exactly when tracing and evals earn their keep: what you need is cost-per-task and pass-rate data before and after the switch—not gut feel.

If your team already splits work between Copilot and Claude Code, assign the same ticket to both paths and compare the outputs from the two branches. Only by looking at eval results alongside cost data can you determine whether "cheaper tokens" really means "cheaper tasks."

One-line Summary

GitHub Copilot will drop Opus 4.8, Opus 5, and GPT-5.6 Terra on Nov 16, but vendor APIs remain available.

Paragraph Summary

GitHub announced that starting November 16, 2026, it will remove Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra from all Copilot modes, while these three models remain Active in Anthropic's and OpenAI's native APIs with tentative retirement dates after 2027. Replacements are Opus 5.5 and GPT-6.1 Sol/Luna—Opus 5.5 is 20% cheaper, and Luna's output price is just 1/24 of Terra's—though no cost-per-task data was provided. This is the second removal wave in a month, only 28 days apart. The article recommends checking hardcoded model names, org policies, rerunning evals, and deliberately choosing Sol or Luna before the deadline, and notes you can keep using Opus 5 via Claude Code by specifying the full model name.