Georgia · Senate

Who is running?

Jon Ossoff (Democratic), Mike Collins (Republican). The margin below is Jon Ossoff minus the other candidate in the main contest; the rest are on the ballot and not in it.

The race, in brief

Georgia elects a U.S. senator on November 3, 2026, with a runoff on December 1 if no candidate wins a majority. The text calls Georgia a purple or swing state: it backed Joe Biden by 0.24% in 2020 and Donald Trump by 2.2% in 2024, Democrats hold both Senate seats, and Republicans hold all statewide executive offices and the legislature.

Democratic incumbent Jon Ossoff, senator since 2021, is seeking a second term and was unopposed in his primary. The Republican nominee is U.S. Representative Mike Collins, who led the May 19 primary with 40.5% and, after Donald Trump's endorsement, won the June 16 runoff against former football coach Derek Dooley with 55.5%; Governor Brian Kemp had endorsed Dooley. As of June 30, 2026 Ossoff had raised $97,986,263 with $42,587,451 on hand; Collins $7,409,350 with $2,079,981 on hand.

All eleven raters listed call the race D, all in September 2026: CBS News, Cook, Decision Desk HQ, The Economist, FiftyPlusOne, Fox News, Inside Elections, RealClearPolitics, Sabato's Crystal Ball, Silver Bulletin and Split Ticket. The aggregate polling average is Ossoff 50.6%, Collins 42.8%. Recent polls include InsiderAdvantage (R), September 22–23, Ossoff 50% to 42%; YouGov, September 15–18, 51% to 41% among registered voters; and Quantus Insights (R), September 14–16, 48% to 44%.

A paragraph quoting a model forecast from the day this brief was written is omitted; the current forecast is at the top of this page.

The facts, each beside its sentence

Not in the source: The forecast rows in facts carry no basis field, so the dossier cannot say what the model's numbers stand on (for example, whether the statewide polls in the text are in it).; The ratings table gives only the party letter or 'Tossup'/'Flip'; no Lean/Likely/Safe strength is given.; The Republican runoff results table is corrupted in the text (replaced by the ratings table); only Collins's 55.5% share is given, with no vote counts or Dooley's share.; facts.candidates marks Jon Ossoff as incumbent: false, but the text says he is the incumbent.; Allen Buckley appears in one poll but the text does not list him as a candidate.; The text gives no result for Ossoff's 2020–21 election.

Read from the district's section of Wikipedia's state page (CC BY-SA), archived as 9ebbbd8d370b; written by claude-fable-5-1 (batch reader) on 2026-09-30 and checked mechanically: every quote occurs verbatim in the archived text, every number in the brief occurs in the inputs. No probability here is the reader's.

Who holds the seat now

Jon Ossoff (Democrat), in this chamber since 2021, 1 term, class 2 (term ends 2027). Committees: Senate Committee on Appropriations; Senate Committee on Rules and Administration; Senate Select Committee on Intelligence.

Raphael G. Warnock (Democrat), in this chamber since 2021, 2 terms, class 3 (term ends 2029). Committees: Senate Committee on Agriculture, Nutrition, and Forestry; Senate Committee on Banking, Housing, and Urban Affairs; Senate Committee on Finance; Senate Special Committee on Aging.

From the congress-legislators dataset (CC0), archived as ae8cb9477c48. For a Senate race this lists both of the state's senators; the one whose term ends in the election year holds the seat on the ballot. For a House district redrawn for 2026 it is the member elected under the old lines with this number, who may be running elsewhere; the dossier above names who is running here.

The money

FEC bulk candidate summaries through June 30, 2026; $2,445,053 of independent expenditures for or against these candidates this cycle (FEC Schedule E, types 24E and 24A), which is the parties' and outside groups' own revealed view of whether this seat is in play.

How it has voted for president

A shift of 3.1 points toward the Democrats from 2016 to 2024. county returns summed; MEDSL county file (2016), an MIT-licensed county file (2020, checked against certified state results) and MEDSL precinct returns (2024). For a district split across counties this is an estimate, and where the state published the exact 2024 figure for these lines, that exact figure is the one the model uses.

The state in numbers

ACS 5-year state tables; BLS county unemployment and BEA county income, transfers and jobs summed to the state. Every figure is a named column of a named public-domain table, archived before it was read.

Who is ahead, and by how much?

Two numbers because they answer two questions. The nowcast is where opinion stands; the election-day forecast is where it is likely to end up.

How likely is each to win?

Monte Carlo error is ±0.2 points. A change smaller than that is simulation noise rather than news, and this page will not describe it as movement.

How wide is the range?

Nowcast: 90% of simulations fall between -3.1 and +15.3 points. Election day: -4.0 to +18.4. The election-day interval is 4.0 points wider, which is the 33 days of movement still to come.

What is the uncertainty made of?

Total sd 7.11 points, made of race-specific polling error 49%, drift to election day 32%, correlated polling error 12%, latent opinion 7%. The decomposition is stored with the forecast rather than reconstructed, so a probability whose dominant input is a placeholder is visibly that.

What evidence is this built on?

15 polls, most recent 2026-09-18. 24% verified against a primary source.

What changed since the last run?

Shaded band is the 90% interval, line is the central estimate, positive is Jon Ossoff ahead of Mike Collins. Over 18 stored forecasts the central estimate moved -0.2 points. There were 7 specification changes (marked on the axis) and no new polls arrived over this window — so every movement in this chart is a change to the model, not a change in the race.

Since 2026-10-01: margin +0.00 points, win probability +0.0 points, on +0 polls.

Not answered here: why it changed. Splitting a move into new polls, time and specification means re-running the earlier model against the later inputs -- two extra posterior fits -- so it is computed on a cadence (scripts/attribute_forecasts.py) rather than while you wait. This pair of runs has not been decomposed yet. The size of the move is shown; the cause is not guessed at.

How has this model done before?

Measured on 230 forecasts across 230 races in cycles 2016,2018,2020,2022,2024: mean absolute error 4.56 points, Brier score 0.0543. The nominal 90% interval contained the result 94% of the time, the 80% interval 84%, and the 50% interval 47%. Read that with care: a national polling miss moves every race in a cycle at once, so the sample size for a question about the intervals is 5 cycle(s), not 230 forecasts. Per cycle -- bias +1.644 pts; 2016: n=43 bias=+4.68 cov90=0.860; 2018: n=63 bias=-0.52 cov90=1.000; 2020: n=35 bias=+5.44 cov90=0.886; 2022: n=54 bias=-0.60 cov90=0.944; 2024: n=35 bias=+1.48 cov90=1.000.

Where did every number come from?

Nowcast run b414f1ca-5d27-4f9c-924b-702d1c6bec27 and election-day run 8afde914-a7cc-4378-a24b-e2b50c1ca2ce, model single_race_latent/0.8.1 at commit bdf7e83107c17a2399efe61a44ef1bbee48f9284, seeds 20260924 and 20260924.