Cursor Maps 50 Repos at Once While Copilot Cuts 3 Models
 |
Cursor Maps 50 Repos at Once While Copilot Cuts 3 Models
Cursor shipped cross-repo dependency mapping across 50 linked repositories the same week GitHub announced it is deleting three models from Copilot. The IDE war just had its busiest week — here's what's changing, and what to do before November 16.
There is a rhythm to how AI coding tools used to evolve: a lab released a model, GitHub added it to Copilot, and developers went back to work. That rhythm broke this week. On October 15, GitHub's changelog announced that three models — Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra — will be deleted from Copilot on November 16, 2026. Days later, Cursor shipped version 0.45 with what is reportedly cross-repository dependency mapping across as many as 50 linked repos in a single workspace.
One tool is subtracting models. The other is adding the kind of capability that used to define Copilot's moat. If you write code for a living, this is the week the AI coding assistant market stopped being a Copilot monopoly and started being a real fight.
What Cursor 0.45 actually adds
The headline feature of Cursor 0.45, according to reports from its October 17 release, is the ability to trace function calls and type definitions across as many as 50 linked repositories in a single workspace. For anyone working in a microservices architecture — frontend in one repo, backend in another, shared libraries in a third — this is the feature that matters most.
Here is why: most AI coding tools are brilliant inside a single repository and blind outside it. They can index your whole repo, but they cannot tell you that changing a function signature in the payments service will break three downstream consumers in repos they have never seen. Cross-repo dependency mapping attacks exactly that problem. Instead of retrieving similar code by vector similarity, the tool parses the actual dependency edges between repositories.
Version 0.45 also ships with a new model for codebase understanding rolling out to users, a substantially improved Tab model for code completion, repository-level rules in .cursor/rules that the agent picks automatically, and support for DeepSeek R1 and DeepSeek V3 models, which Cursor self-hosts in the US. Put together, it reads like a release aimed squarely at large, messy, real-world codebases — the kind enterprise teams actually maintain.
What GitHub is cutting — and when
GitHub's October 15 changelog entry is precise about the scope. On November 16, 2026, the following models leave every GitHub Copilot experience — Copilot Chat, inline edits, ask and agent modes, and code completions:
- Claude Opus 4.8 → suggested replacement: Claude Opus 5.5
- Claude Opus 5 → suggested replacement: Claude Opus 5.5
- GPT-5.6 Terra → suggested replacement: GPT-6.1 Sol or GPT-6 Luna
There is one carve-out, and it is unusual for GitHub: Claude Opus 4.8 will remain available to individual GitHub Copilot subscribers on annual plans, so they retain an Opus offering. On annual Pro and Pro+ request-based billing, Opus 4.8 carries a premium-request multiplier of 27 — meaning each request counts as 27 premium requests. Everyone on monthly plans, team seats, or enterprise licenses loses Opus 4.8 along with the other two models.
This is also not the first cut of the season. A September 18 changelog entry set October 19 as the removal date for six other models, including GPT-5.5, GPT-5.4, and Grok 4.5. Two waves, 28 days apart, nine models out. The era of Copilot accumulating every new model indefinitely is over.
Don't assume only the Chat picker is affected — the deprecation covers completions and agent modes too. Before November 16:
- Grep for pinned model names. Any workflow, CI job, extension config, or agent config that calls
claude-opus-4-8,claude-opus-5, orgpt-5.6-terraby ID will fail or silently reroute after the cutoff. GitHub's own advice: update your workflows and integrations before the date. - Check your org's model policy. The replacement models switch on automatically for Business and Enterprise unless an admin turned off the global default. If yours did, the replacements stay off until someone enables them in Copilot's model policies.
- Re-run your evals on the replacements now. Until November 16 you can run the old model and the new one side by side inside Copilot — this is the only window where a fair comparison is free.
- Pick Sol or Luna on purpose. GitHub offers two Terra replacements in very different classes: GPT-6 Luna costs roughly $0.10 per million input tokens versus Terra's $2.00, but it is a lightweight model. Test both on real tasks before the cheaper one becomes the default by habit.
Why GitHub is trimming now
GitHub has not published a detailed rationale beyond the changelog itself, but the pattern is consistent with how Copilot has handled every previous model generation: newer, cheaper, or more efficient models arrive, and the older ones get a fixed retirement date instead of an indefinite legacy tail. Running a long model list is expensive in inference cost and in support overhead — every model needs its own routing, rate limits, and fallback behavior tested across every IDE integration Copilot ships.
The replacement economics are hard to argue with. Claude Opus 5.5 costs 20% less per token than the Opus models it replaces — $4.00 per million input tokens versus $5.00. GPT-6.1 Sol holds Terra's $2.00 input price while cutting output from $12 to $10 per million. GitHub's long-term bet is task-based automation over manual model selection. The teams that will feel November 16 the hardest are the ones that built prompts or evaluation pipelines tuned to a specific model's behavior.
The rest of the week: Codex, OpenHands, and the open-weight squeeze
The Cursor-versus-Copilot story did not happen in isolation. OpenAI's Codex CLI moved through three point releases in under two weeks (0.161 to 0.163), making GPT-6.1 Sol the default model, adding Git worktree creation, and introducing instant steering. The open-source OpenHands agent framework reported a 78% score on SWE-bench Verified — a 12-point quarterly improvement — driven by a new verification-loop architecture that tests every proposed change against the project's existing test suite before a human sees it.
Meanwhile, two open-weight coding models went public within days of each other: Meta's Muse Spark Open (816B total parameters, 42B active, 1M-token context) on October 15, and Reflection AI's Beam (501B total, 23B active, Apache 2.0 license) on October 16. For a platform like Copilot, every open-weight release that closes the capability gap makes it cheaper for a large enterprise to self-host a coding model instead of paying GitHub's markup. That dynamic adds a second layer of pressure behind the November retirement.
The honest summary of the week: builders are shipping features and open weights weekly while the incumbent trims its lineup. That is the reversal. GitHub used to set the pace of change in AI coding; this month it is reacting to one.
Frequently asked questions
Which Copilot experiences lose the three models on November 16?
All of them. GitHub's changelog says the deprecation covers Copilot Chat, inline edits, ask and agent modes, and code completions. If you only audit the Chat model picker, you will miss agent-mode pins, Copilot CLI scripts, and IDE defaults still routing inline edits through a retiring model.
Do Claude Opus 5 and GPT-5.6 Terra disappear entirely on November 16?
No — they disappear from Copilot, not from existence. Anthropic's own deprecations page still lists claude-opus-5 as Active with a tentative retirement date "not sooner than July 24, 2027" (and Opus 4.8 "not sooner than May 28, 2027"). OpenAI's deprecations page does not list gpt-5.6-terra as deprecated at all. After November 16 you can still call all three from the vendors' own APIs.
Why does Claude Opus 4.8 survive for individual annual subscribers?
GitHub carved out that exception so individual subscribers "retain an Opus offering." It is a narrow reprieve: it does not apply to Opus 5 or GPT-5.6 Terra, and it does not cover Business or Enterprise seats. On annual Pro/Pro+ request-based billing, Opus 4.8 carries a multiplier of 27, so each request counts as 27 premium requests.
Which Terra replacement should I pick — GPT-6.1 Sol or GPT-6 Luna?
GitHub files them in different classes. Sol is the closest successor ($2.00 input / $10 output per million tokens versus Terra's $2.00 / $12). Luna is lightweight and dramatically cheaper ($0.10 input / $0.50 output) but at roughly a twenty-fourth of Terra's output price, it is a different model class. Try both on a few real tasks before the cutoff — a cheaper token is not a cheaper finished task if the model takes more turns.
Is Cursor 0.45's cross-repo dependency mapping available to everyone?
The feature was reported with Cursor's October 17 release as able to trace function calls and type definitions across up to 50 linked repositories in one workspace. As with any fast-moving IDE release, confirm availability in your own install via the Check for Updates command before planning a migration around it.
The bottom line
This week told us exactly where the AI coding market is heading. GitHub is consolidating Copilot around fewer, cheaper models and betting that task-based automation matters more than a long model menu. Cursor is betting the opposite: that developers want more of everything — more repos in one view, more models to choose from, more of their real codebase understood at once.
Both bets can be right at the same time, but only one of them has a hard deadline attached. Whatever editor you use, search your configs for the three retiring model names before November 16. It takes five minutes, and it is the cheapest insurance you will buy all quarter.
Sources and further reading
- GitHub Changelog — Upcoming deprecation of selected GitHub Copilot models in mid-November (Oct 15, 2026)
- TrendinTech — GitHub Copilot Models: 3 Cut by Nov. 16 (Oct 2026)
- DEV Community — Copilot Drops Opus 5 on Nov 16; Claude's API Keeps It Until at Least July 2027 (Oct 16, 2026)
- Cursor — 0.45.x changelog (rules, DeepSeek models, codebase understanding)
