Switching Models in GitHub Copilot Can Make the Same Job Cost 45x More — The Full Per-Token Rate Table, and the Sol 50% Promo That Ends 9/3 (2026 Verified)

Since GitHub Copilot moved to “subscription + GitHub AI Credits usage-based billing” on 2026-06-01, there is a new thing to think about: changing the model in the dropdown changes your bill.

Most people switch models because one “feels smarter.” Almost nobody checks what that model charges per million tokens first. We read GitHub’s official rate page in full today (2026-08-26), and the spread is wider than it looks: on the same job, the wrong model can cost 45x more.

(All rates below are taken directly from GitHub’s official Models and pricing documentation, not from third-party reporting. Conversions and scenario math are ours and are labeled as such.)

1. The urgent part: inside Copilot, the Sol 50% promo ends 9/3

The official rate page carries a footnote on the GPT-5.6 Sol rows:

“GPT-5.6 Sol is available at promotional pricing, 50% off standard rates” — effective through September 3, 2026.

That is 8 days from our verification date.

This deserves its own section because the same 50% discount, on the same model, has a different deadline depending on where you use it:

In other words: if you take the Sol unit price you see inside Copilot and build a September budget on it, you will be off by half.

Sol’s promotional rate as shown at verification time, plus the derived standard rate:

Being explicit about the derivation (single-source disclosure): GitHub’s page says “50% off” but never prints the standard number. The US$4／US$20 figure is ours, obtained by doubling the promotional price. It happens to match OpenAI’s own list price — when we read OpenAI’s official model page on 8/22, Sol was listed at exactly US$4／US$20. The two line up, but GitHub has not stated in writing what Sol reverts to after 9/3. We treat this as a derivation, not an official commitment.

2. The full rate table: how far apart the cheapest and dearest really are

Below is the complete rate table as shown on the official page at verification time (per 1M tokens, USD). This table is the whole game — your bill is simply “tokens used × this table.”

OpenAI

Anthropic

Google／Microsoft／xAI／others

Output spread: cheapest US$1.20 (GPT-5.6 Luna, MAI-Code-1.1-Flash) against dearest US$50.00 (Claude Fable 5, Opus 4.8 fast mode) = 41.7x.

Input spread: US$0.20 against US$10.00 = 50x.

3. What that means in money: one job, eight choices

A concrete scenario (our math, not official numbers): a fairly heavy refactor that takes in 200K tokens of context and produces 50K tokens of output.

(Method: 0.2M × input rate + 0.05M × output rate. GitHub states 1 AI credit = US$0.01, so the dollar figure × 100 gives the credit count.)

On the same job, the most expensive model burns 45x the allowance of the cheapest. That is not a “save a little” difference. That is the difference between your monthly allowance lasting 30 days and lasting 16 hours.

4. Three details the page does not flag, but that should change your pick

Detail one: Grok’s cache discount is half of everyone else’s

Look closely at the Grok rows. Input US$2.00, cached input US$0.50 — the cache rate is 25% of the input rate.

Every other model on the table (GPT, Claude, Gemini, MAI-Code, Raptor) prices cached input at exactly 10% of input: US$2.00 to US$0.20, US$5.00 to US$0.50, US$0.75 to US$0.075. One tenth, all the way down.

Why it matters: agentic workflows live on cache — the same repository, the same long prompt, read again and again. Grok 4.6’s official positioning is precisely “agentic coding and complex multi-step workflows,” yet its cache discount is the weakest on the table.

Running the same 200K-token context ten times, assuming nine cache hits (our math):

• A model with a 10% cache rate (e.g. Claude Sonnet 5, input US$2.00／cache US$0.20): 0.2×2 + 9×0.2×0.2 = US$0.76 of input cost
• Grok 4.6 (input US$2.00／cache US$0.50): 0.2×2 + 9×0.2×0.5 = US$1.30

Identical headline input price; 71% more input cost after ten rounds. Same sticker, different bill — the difference hides in the cache column.

Detail two: the newer MAI-Code is 3.75x cheaper, not dearer

Intuition says version 1.1 should cost more than version 1. Here it is the reverse:

Both input and output are exactly 3.75x cheaper. And on 8/11 GitHub’s changelog announced the deprecation of MAI-Code-1-Flash alongside the launch of MAI-Code-1.1-Flash — so the old model is both going away and roughly four times more expensive.

If your config (or your team policy) is still pinned to MAI-Code-1-Flash, this is the easiest, most trade-off-free action in this article: switch, get the newer model, and cut that line of the bill to about a quarter.

Detail three: code completions are not billed, and are unlimited

The documentation is explicit:

“Code completions and next edit suggestions are not billed in AI credits. They remain unlimited for all paid Copilot plans”

This matters because it settles the right direction for saving credits. The grey inline suggestions you get while typing cost nothing. What consumes credits is chat, agent tasks, and anything that pulls in a large context and writes back a long answer.

So “use Copilot less to save money” is the wrong move — completions were already free. The right move is picking the right model for the actions that are billed, which brings you back to the table above.

5. Our editorial view (opinion, not official fact)

Usage-based billing pushes one fact into the open: the model dropdown is now a price list.

The old premium-request multiplier system at least compressed complexity into a single number (1×, 10×), so you roughly knew what you were spending. Per-token pricing means your bill is decided by three quantities nobody estimates in their head — how much context you send, how long the answer runs, and how much of it hits cache. The menu shows model names, not prices.

Our read: this is good for heavy users (the cheap models really are cheap — a US$1.20 output rate on Luna and MAI-Code-1.1 is close to free) and bad for casual model-hoppers (click Fable 5, run ten agent tasks, and the allowance is gone).

And the 9/3 deadline makes the same point: any price inside Copilot that looks cheap should first be checked for whether it is promotional. Two promotions are live on the table right now — Sol through 9/3, Gemini 3.6／3.7 Flash through 12/31. Neither of those is the model’s long-term price.

6. Three things you can do today

1. If you use GPT-5.6 Sol: put 9/3 in your calendar. From September, the same usage bills at roughly double (derived figure — see section 1).
2. If you are still pinned to MAI-Code-1-Flash: switch to 1.1. GitHub has announced the old one’s deprecation, and the new one is 3.75x cheaper.
3. If you run a lot of agent tasks: read the cache column before you pick, not just the input price. Two models with the same headline price can diverge 71% over ten rounds.

How much allowance you actually have — and how much the 9/1 transitional top-up expiry takes away — is covered separately in the GitHub Copilot 9/1 credit cliff. That one is about how many credits you get; this one is about how fast you spend them. You need both.

Sources

• GitHub official documentation, “Models and pricing for GitHub Copilot” (full per-1M-token rate table, 1 AI credit = US$0.01, Sol at 50% off through 2026-09-03, Gemini 3.6／3.7 Flash promo through 2026-12-31, code completions unbilled and unlimited): https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing | verified 2026-08-26
• GitHub changelog, “Grok 4.6 is now available in GitHub Copilot” (2026-08-14; Pro／Pro+／Max／Business／Enterprise; billed at provider list pricing under usage-based billing; Business／Enterprise admins must enable the policy, off by default): https://github.blog/changelog/ | verified 2026-08-26
• GitHub changelog (2026-08-11) MAI-Code-1-Flash deprecation and MAI-Code-1.1-Flash launch; (2026-08-13) Gemini 3.7 Flash added to Copilot | verified 2026-08-26
• Our own prior verifications, used to cross-check Sol’s standard price and the other platforms’ promo deadlines: see the links below.

What we did not verify, and are not stating as fact: GitHub’s page does not say what Sol reverts to after 9/3. The US$4／US$20 figure is derived by doubling the promotional rate and cross-checked against OpenAI’s official list price — it is not a commitment from GitHub. Whether the promotion is extended, or ends early, is not addressed by either party.

Related verification on this site

• Tool page: GitHub Copilot plans and free tier
• The GitHub Copilot 9/1 credit cliff: 7,000 credits per Enterprise seat drops back to 3,900
• GPT-5.6 Sol at 50% off on AI gateways is real — but three things go unsaid
• GPT-5.6 Sol’s 20% cut is promotional pricing, guaranteed only through 11/21
• Gemini 3.7 Flash at “half price”? It doubles on 2027-01-01
• GitHub Copilot free vs paid
• The annual-plan trap behind AI price cuts

All rates quoted from GitHub’s official documentation, verified 2026-08-26. Promotional terms can change at any time — check the official page before committing. Scenario math is ours; your actual bill depends on your actual token usage.