Switching the GitHub Copilot model can save credits. It can also turn a straightforward fix into repeated attempts, context reads and manual review. A useful team measure is the cost of reaching a correct result, including human effort.
Luna and Haiku are candidates for bounded, verifiable work. Terra 6.1 and Sonnet can handle execution across more dependencies. Sol and Opus deserve consideration for difficult diagnosis, architecture decisions or work with significant consequences. This is a working approach to calibrate on your repository, rather than a proven ranking for every task.
Checked on October 27, 2026. This guide uses GPT-6 Luna, GPT-6.1 Terra, GPT-6.1 Sol and Claude Haiku, Sonnet and Opus 5.5. Availability depends on the client, plan and company policies. The lab above uses these versions; recheck prices and models before making budget decisions.
What the evidence supports
Among the sources reviewed, we found no public controlled test comparing these six models inside the same Copilot environment with identical context, tools, reasoning effort and tasks spanning code, tests, automation and documents. Available evidence combines agent research, vendor evaluations and GitHub operational guidance.
Several concrete signals can inform a choice:
- Luna and Haiku: GitHub's comparison places them in the fast, small-task category. That supports an economical starting point, without establishing equivalence on extended investigations.
- Haiku and Sonnet: Anthropic reports 39.2% and 70.6%, respectively, on Terminal-Bench 4.0. These are vendor-published evaluations with details in the system card, rather than tests of your Copilot installation.
- Terra: OpenAI reports 50.3% on Terminal-Bench 4.0 at maximum reasoning effort. Its launch still positions Sol for more complex coding. Numbers from different vendors do not constitute a controlled cross-model leaderboard.
- Sol: OpenAI's published results include DeepSWE 1.1 for engineering in real codebases and AutomationBench 1.0.6 for tool-based workflows. They provide relevant signals for investigation and automation, subject to each evaluation's conditions.
- Sonnet and Opus: Anthropic describes Sonnet as an option for everyday work and documents, and Opus for open-ended problems requiring sustained judgment. Strong performance on one evaluation does not make the expensive variant necessary for every edit.
The system around the model matters too. SWE-Bench Mobile evaluated 22 agent-model configurations on an iOS codebase and observed performance gaps of up to six times for the same model across agents. It is a specific mobile-development setting with earlier configurations, illustrating why another agent's result should not be treated as a Copilot prediction.
Credits, tokens and attempts are different measures
For Business and Enterprise, GitHub charges model usage through tokens converted into credits: one credit equals US$0.01. Included credits are pooled at the billing entity. Code completions and next edit suggestions remain outside AI-credit billing on paid plans.
Compare exact versions: “Haiku” and “Luna” are families. An earlier generation can have different pricing from the versions shown here.
The calculator uses Copilot's own pricing table, rather than external application subscriptions. It separates uncached input, cache reads, cache writes and billable output. Long context can change the rate, with different thresholds across models.
In the initial example, a hypothetical request with 20,000 uncached input tokens and 4,000 output tokens costs US$0.004 on Luna or Haiku, US$0.04 on Terra, US$0.08 on Sol or Sonnet and US$0.16 on Opus. Equal volumes illustrate the arithmetic only: models can use different token counts.
Twenty requests of this size on Luna would total the price of one on Sol. That does not mean Sol always finishes in one attempt or Luna needs twenty. Adjust repetitions and observe where the unit-price advantage disappears. Then account for review time, rework and the consequences of an error.
Switching without wasting previous work
In VS Code, open Chat and use the model picker below the input. The official instructions allow switching during a conversation and explain that this does not change the inline-suggestion model. If an option is absent, check the client, plan and policies before assuming an account problem.
A switch makes sense when the task's needs change. For example, finish a diagnosis with Sol and hand Haiku a mechanical edit with defined files and acceptance criteria. Repeatedly switching just to try names can add cost and make comparisons harder.
GitHub's efficiency guide explains that changing models invalidates the previous cache. Prefer transitions between stages and focused context. A useful handoff contains:
- The objective and expected behavior.
- Relevant files and confirmed facts, separated from hypotheses.
- Tests run, commands used and outstanding failures.
- Change boundaries and a stopping condition.
Keeping the model during ongoing work can preserve cache value. When changing problems, a new conversation avoids irrelevant history. Preserve evidence that contradicts the initial hypothesis too; an optimistic summary can hide the actual cause.
What about Auto?
Auto with task optimization offers Efficiency, Balance and Intelligence. These priorities guide routing without fixing a model or maximum spend. For paid plans, the documentation lists a 10% discount on model costs in eligible experiences.
For varied work, Auto/Balance is an option to evaluate. For repeatable model comparisons, select models manually and record the version and configuration. Use budget controls when a spending cap is the objective: the Efficiency label alone does not provide one.
Where to start cheap and where to escalate
The choices below are editorial recommendations for a pilot. Requirements, risk and verification matter more than line count.
Local code changes. Updating a message, generating an adapter with a known contract or applying a mechanical transformation are candidates for Luna/Haiku. Inspect the diff and run the relevant test. When the change requires understanding effects across services, try Terra/Sonnet; for an unknown cause, consider Sol/Opus for diagnosis.
Test generation. Fixtures and repetitive cases can start lightweight. Defining expected behavior, covering concurrency or authorization and finding a regression require analysis. Do not let an agent “fix” a test simply to accept the implementation. Check whether a deliberately introduced defect makes the test fail.
TestGenEval, published at ICLR 2025, evaluated test generation and completion across 11 Python repositories. It identified difficulties with coverage and assertions along complex paths. It does not compare this guide's current models; it supports evaluating test quality separately from code generation.
Automation. A read-and-transform script with a stable schema can start with Luna/Haiku. Workflows with multiple tools, partial failures, retries or shared state justify trying Terra/Sonnet or Sol. If deletion, payments, credentials or production data are involved, begin with permissions, dry runs, idempotency and recovery. A more capable model is not authorization to execute.
Documents. Summarizing a diff or updating instructions from supplied facts can be lightweight work. Reconciling sources, writing an architecture decision or producing a runbook requires checking contradictions, constraints and commands. Terra/Sonnet are intermediate candidates; Sol/Opus may help with open-ended analysis. Professional appearance and fluent prose do not validate content. Reopen generated files and test examples.
Refactoring and architecture. Separate investigation, planning, implementation and review. Use greater capability where uncertainty is concentrated and hand bounded parts to economical models. Avoid imposing this division when a task is so small that preparing the handoff costs more than doing the work.
When to stop retrying the cheaper model
SWE-Router, presented at an ICML 2026 workshop, found benefits from observing an agent's early steps before deciding to escalate. Experiments used mini-SWE-agent with GPT-5-mini or DeepSeek-V3.2 and Gemini-3-Pro-Preview, rather than these six models. Results also varied across test sets.
The transferable lesson is to observe progress. If an agent repeats the same command, changes files without a testable hypothesis, invents an API or cannot explain a failure, stop and review the context. Then choose whether to clarify requirements, repair the environment, increase model capability or investigate manually. An unavailable dependency does not necessarily need a different model.
A team could initially allow one attempt and one error-informed correction before reassessing. That limit is a suggested policy, not a research-validated number. Adjust it using your own work.
Turning the guide into a team practice
Build a small set of representative tasks covering fixes, tests, automation and documents. Keep a reproducible initial state and compare configurations from that same state, without leaking one run's solution into the next. Repeat evaluations to distinguish a good attempt from consistency.
Record model and effort, Copilot client, acceptance results, available usage measurements, elapsed time, human review and defects found later. Set criteria before seeing which model wins. A passing test with an incorrect business rule should count as a failure.
Track two separate measures: AI cost per accepted delivery and total human time per accepted delivery. Include usage from discarded attempts. Prompts, tokens and generated lines are not productivity measures.
Enterprise users can check consumption in Copilot settings → Usage. The monitoring documentation distinguishes this view from individual-plan billing. Not every view can attribute usage to a specific prompt; record the limitations of your measurement.
For CLI automation, session credit limits are in public preview, with a minimum of 30 credits. The interactive command is /limits set max-ai-credits NUMBER; programmatic runs use --max-ai-credits NUMBER. It is a soft limit: an ongoing response can exceed it. Combine it with monthly budgets and administrator policies.
For more on context and validation, see prompt caching and harness engineering.
If you use Copilot daily, start this comparison with one recurring task. The goal is to identify the configuration that delivers the work with less cost and review, and to recognize early when a problem needs more capability.
Original sources and documentation are in English, checked on October 27, 2026. Cited benchmarks are results from their authors or vendors; MnzAI Labs did not run a controlled six-model comparison. Examples, the selector and operational limits are editorial guidance with explicit assumptions.