← All posts

FinOpsAid blog

December's Copilot model swaps aren't all price cuts. Two of them double.

The question that lands on your desk in January won't be "which models did GitHub retire?" It'll be "why did our Gemini spend double when nobody changed anything?"

Something did change. GitHub retires six Copilot models on December 14, and its suggested replacements aren't priced like the originals. In November's round every replacement was cheaper. This round is mixed, and the two most expensive surprises come from models that look like routine upgrades.

What's being retired, and what replaces it

On November 13 GitHub announced that six models will be deprecated across every Copilot experience on December 14, 2026 (GitHub Changelog, November 2026). The announcement has no pricing in it. The prices below come from GitHub's models and pricing page, per million tokens, retrieved November 14.

Retiring Suggested replacement Input Output Change
Claude Fable 5 Claude Fable 5.2 $10.00 → $8.00 $50.00 → $40.00 −20%
Claude Haiku 4.5 Claude Haiku 5.5 $1.00 → $0.10 $5.00 → $0.50 −90% (≤100K context)
Grok 4.6 Grok 4.7 $2.00 → $2.00 $6.00 → $6.00 none
Gemini 3.8 Flash Gemini 4 Hydrogen $0.75 → $1.50 $3.75 → $7.50 +100%
Gemini 4 Argon Gemini 4 Carbon $2.00 → $4.00 $10.00 → $20.00 +100% (after Dec 1)
Kimi K3 to be announced $3.00 $15.00 unknown

One AI credit is $0.01, so a million Fable 5 output tokens is 5,000 credits today and 4,000 on Fable 5.2. Haiku 5.5 has a long-context tier above 100K tokens ($0.50 in, $2.50 out). That's still half Haiku 4.5's price, not a tenth of it.

Change in output-token price from each retiring Copilot model to its suggested replacement A diverging bar chart with a zero line. Bars to the left are price cuts, bars to the right are increases. Claude Fable 5 to Fable 5.2 is 20 percent cheaper. Claude Haiku 4.5 to Haiku 5.5 is 90 percent cheaper for contexts up to 100 thousand tokens. Grok 4.6 to Grok 4.7 is unchanged. Gemini 3.8 Flash to Gemini 4 Hydrogen is 100 percent more expensive. Gemini 4 Argon to Gemini 4 Carbon is 100 percent more expensive once Carbon's promotional rate ends on December 1, compared with the promotional Argon rate paid today. Kimi K3 has no announced replacement, so its change is unknown. OUTPUT PRICE CHANGE, RETIRING MODEL → SUGGESTED REPLACEMENT CHEAPER DEARER Fable 5 → Fable 5.2 −20% Haiku 4.5 → Haiku 5.5 −90% Grok 4.6 → Grok 4.7 0% Gemini 3.8 Flash → Gemini 4 Hydrogen +100% Gemini 4 Argon → Gemini 4 Carbon +100%* Kimi K3 → to be announced ? * Carbon's post-December 1 rate, against the promotional Argon rate you pay today.
The two Gemini swaps double the price per token; the two Claude swaps cut it. Output rates from GitHub's models and pricing page, retrieved 2026-11-14. Input rates move by the same ratio in every row.

The Gemini swaps are the ones to watch

Gemini 3.8 Flash is on promotional pricing that runs through December 31. It retires on December 14, so that promo never actually ends for anyone. Its replacement, Gemini 4 Hydrogen, has no promotion and costs exactly twice as much for input and output. Cached input is the same $0.075 on both, so heavy caching narrows the gap a little. It doesn't close it.

Here's an illustrative month: 400M input tokens, 75% of them cached, plus 40M output. On Flash that's about $248. On Hydrogen it's about $473. That's 1.9x for the same work.

Gemini 4 Argon to Carbon has a timing trap. Both are on a promotional $2 / $10 rate that ends on December 1. From that date Carbon goes to $4 input and $20 output. GitHub's page doesn't list Argon's post-promotion rate, so for the two weeks between December 1 and December 14 we can't tell you what Argon costs. What we can say is that a team on Argon today will pay twice as much per token on Carbon in late December, and the bill will show it in January.

The Claude swap may not happen on its own

Under default model enablement, GitHub says Business and Enterprise customers get the suggested replacements turned on automatically. That holds for Haiku 5.5, Grok 4.7 and the two Gemini models.

Claude Fable 5.2 is different. Its own changelog says administrators must enable the Fable 5.2 policy, which is off by default, and enabling it means accepting that Anthropic keeps prompts and outputs by default to run safety classifiers (GitHub Changelog, November 2026). Eligible enterprises can get zero data retention through the end of the year, but only through their account team. Even after approval, someone still has to switch the policy on.

The two announcements don't spell out how they interact. Don't assume Fable 5 users will land on Fable 5.2 on December 15. If the policy stays off, they'll pick something else, and whatever they pick sets the price. That might be cheaper, or it might be a pricier frontier model.

Two smaller details. Claude Haiku 4.5 stays available to individual subscribers on annual plans, so the deprecation doesn't reach them. Kimi K3 retires with no named replacement. Anyone who standardized on it needs to choose a replacement before December 14.

Our finding: FinOpsAid's per-user AI credit figures come from GitHub's billing data, which reports credits per user per day with no model breakdown. So we can show you that a developer's daily credits doubled after December 14, but billing alone can't prove it was the Gemini swap. To link the change to a model, compare the per-model token counts in GitHub's usage metrics before and after the deprecation date. Read those as an explanation, not as the bill. We cover the two-ledger problem in our November usage metrics post.

What to do before December 14

  1. Find your Gemini 3.8 Flash and Argon users now. Pull totals_by_model from the usage metrics API for the last 28 days. Those teams carry the doubling.
  2. Decide on Fable 5.2 deliberately. Read the data-retention terms, ask your account team about zero retention if it matters, then enable the policy or pick an alternative. Don't let the default decide.
  3. Consider Haiku 5.5 where developers used Flash. For light tasks under 100K context it's far cheaper than Hydrogen. It isn't GitHub's suggested swap, so it won't happen unless someone chooses it.
  4. Mark December 1 on the budget calendar. Carbon's rate doubles two weeks before the deprecation, so December's run-rate jumps mid-month.
  5. Choose a Kimi K3 replacement. Nothing will be picked for you.

Where FinOpsAid fits

FinOpsAid reads your org read-only and tracks AI credits per user, team and cost center from GitHub's billing data, with budgets and alerts that flag a mid-month jump like the December 1 repricing. It won't tell you which model caused a jump, because billing doesn't record that. Related: which Copilot model should we default to, temporary model discounts and your Copilot bill, and how to read the Copilot page.

Frequently asked questions

Which Copilot models are deprecated on December 14, 2026?

Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3. GitHub's suggested replacements are Claude Fable 5.2, Claude Haiku 5.5, Gemini 4 Hydrogen, Gemini 4 Carbon and Grok 4.7. Kimi K3's replacement hasn't been announced. Haiku 4.5 stays available to individual subscribers on annual plans.

Will the December model replacements raise my Copilot bill?

It depends on your mix. Moving from Gemini 3.8 Flash to Gemini 4 Hydrogen doubles input and output prices. Gemini 4 Carbon costs twice today's promotional Argon rate from December 1. Fable 5.2 is 20% cheaper than Fable 5, and Haiku 5.5 is 90% cheaper than Haiku 4.5 for contexts up to 100K tokens.

Is Claude Fable 5.2 enabled automatically for Copilot Business and Enterprise?

No. Its policy is off by default, and an administrator has to enable it. Turning it on accepts Anthropic's default data retention for safety classifiers. Eligible enterprises can request zero data retention through the end of 2026 from their GitHub account team, but approval doesn't enable the model by itself.

Sources

Want to see which teams' credits jump after December 14 before the invoice tells you? Connect your GitHub org — read-only and free during beta — or explore the demo dashboard first.

copilotmodel-mixai-creditsgovernancefinops

See it in your own org

FinOpsAid meters the GitHub spend no invoice itemizes — read-only, and free during beta.