Political monitor · Germany · Article 5 Sep 2026 (before ST count) · Länder briefing after 6 Sep →

Analysis

Germany's AfD Is First in the Polls. It Still Does Not Govern.

Why the next two weeks of state elections matter more for business than a Dexit scare

5 September 2026 — before the Sachsen-Anhalt count

This piece was closed before the Magdeburg count. The Länder briefing after 6 Sep is on the dossier page.

Germany’s Alternative for Germany (AfD) is this week the country’s most popular party in two named national polls. Yet it does not govern. That gap — between a survey lead and the power to write law — is what should govern how investors read the next two Sundays.

Infratest dimap, the Berlin firm that runs the monthly Deutschlandtrend for ARD, Germany’s main public broadcaster, and Forsa, a commercial pollster for the private channels RTL and n-tv, both put the AfD first this week, at 27% and 28%. The governing center-right alliance — the Christian Democratic Union (CDU) and its Bavarian sister party, the Christian Social Union (CSU), together called the Union — is at 21% in each survey.1, 2 In the 23 February 2025 election to the Bundestag, the AfD took 20.8% of the party-list vote and 152 of 630 seats — second, behind the Union at 28.6% and 208 seats.3 Sachsen-Anhalt votes for its Landtag (state parliament) on 6 September; Mecklenburg-Vorpommern and the city-state of Berlin follow on 20 September.

The crucial variable is the Brandmauer, the refusal of other parties to form a coalition with the AfD. It is why Chancellor Friedrich Merz’s Christian Democrats govern with the center-left Social Democrats (SPD), not with the party that now outpolls them. Until that firewall cracks, the AfD functions as a shadow agenda-setter: able to force shifts on migration and industrial power, unable to sign them into law. Over the next six to eighteen months those shifts will outrun any move into the federal cabinet. Dexit, a Russia reset, and an AfD government stay blocked while other parties refuse coalition, ministerial office, or formal toleration.

Sunday’s count will show how popular the party is. The talks after will show whether anyone will let it govern.

ContestKindAfDNextSource
National, Infratest 31 Aug–2 Sep 2026Poll27%CDU/CSU 21%dawum / ARD1
National, Forsa 25–31 Aug 2026Poll28%CDU/CSU 21%wahlrecht.de2
Baden-Württemberg (southwest), 8 Mar 2026Result18.8%3rd; 35 of 157 seatsOfficial final4
Rhineland-Palatinate (west), 22 Mar 2026Result19.5%3rd, statewide voteOfficial statistics5
Sachsen-Anhalt (east), 6 Sep 2026Poll42–43%Election SundayInfratest / INSA6, 7
Mecklenburg-Vorpommern (northeast), 20 Sep 2026Poll35%SPD 32%NDR / dawum8
Berlin, 20 Sep 2026Poll18%CDU 21%, The Left 19%dawum9

From Second Place to Policy Catalyst

The February 2025 election ended the Ampel, the traffic-light coalition of Social Democrats, Greens, and Free Democrats. Merz formed a Union–SPD cabinet, leaving the AfD as the largest opposition party, 10.4 percentage points above its 2021 share.3 On 5 September the aggregator dawum showed a population-weighted average of state-parliament polls at 19.7% for the AfD in the former West and 34.8% in the former East, including Berlin.10 Eastern strength scrambles coalition arithmetic in the states long before it produces a federal majority.

March in the West showed the other side of the map. In Baden-Württemberg on 8 March 2026 the AfD won 18.8%, finished third, and took 35 of 157 seats.4 In Rhineland-Palatinate on 22 March it won 19.5% and finished third again.5 The party can be exceptionally strong in the East and still appoint nobody to a federal ministry.

Three grievances drive the rise: asylum inflows felt in housing and schools; the post-2022 power-price spike and a law that pushed households to replace gas and oil boilers; and distrust of the parties that have rotated through office since reunification in 1990. The AfD’s offer is halt irregular entry, cheap conventional power, and an attack on the firewall itself. The Union has tried to compete on the first two without touching the third.

First asylum applications at the Federal Office for Migration and Refugees fell from 229,751 in 2024 to 113,236 in 2025, a drop of 50.7%.11 Routes, other EU decisions, and the cycle of earlier inflows sit in the same year; the coalition still has every political reason to show control, and none to adopt Remigration, the AfD’s word for large-scale deportation of non-citizens, sometimes framed more broadly. The overlap is the wish for lower energy costs and tighter inflows. The divergence is the generation mix, Russia, and whether policy stays inside EU law. The Union will adopt stricter borders and talk industrial power caps. It will not cross the Brandmauer for the AfD’s agenda on EU membership or Russia. What firms already face is AfD-lite.

Rhetoric Versus Delivery

The platform is expansive. Asylum claims would be lodged outside the EU; returns would accelerate; Remigration would become policy. Nuclear would reopen, coal would run longer, Paris targets would go, long-term Russian gas would return. On Europe the party prefers a loose “bund of nations,” keeps a German EU exit (Dexit, on the Brexit model) as a fallback, and would end Ukraine arms and lift Russia sanctions. On tax it would abolish the solidarity surcharge — the extra levy on income tax first charged to help rebuild the East after 1990 — and inheritance tax, while still keeping the Schuldenbremse, the constitutional debt brake.

The feasibility of this agenda is near zero without executive power. Remigration collides with the Basic Law and EU asylum rules. A generation-mix reversal hits closed plants, EU climate law, and years of build time. Dexit would put at risk the single market German exporters live on. Article 115 caps the federal structural deficit at 0.35% of GDP.12 The solidarity surcharge yielded about €12.6bn in cash in 2024; assessed inheritance tax in 2025 was €13.3bn.13, 14 Cutting both while keeping the brake means spending offsets or a constitutional amendment, neither of which an opposition party can pass.

Migration enforcement and industrial-price politics already move through the Union–SPD cabinet. BDEW, the German utilities association, put new-contract industrial electricity for small and mid-size firms at 17.2 cents per kilowatt-hour for 2026, against 17.64 cents in 2025. The renewable-energy surcharge (EEG) on that bill has been zero since 2023.15 There is no unused levy left to repeal. Manufacturing was 19.7% of German gross value added in 2024 — the slice most sensitive to those bills.16 Tax abolition, Dexit, and a wholesale Russia reset wait on a finance ministry the party does not hold.

The legal cloud is narrower than a ban and broader than a nuisance. On 2 May 2025 the Federal Office for the Protection of the Constitution raised the federal party to a “confirmed right-wing extremist endeavour.” On 26 February 2026 the administrative court in Cologne barred the service, pending the main case, from classifying or publicly treating the federal party as that confirmed endeavour.17 As of 17 June 2026, Correctiv reported five state branches that state intelligence offices may still call confirmed extremist: Brandenburg, Lower Saxony, Saxony, Sachsen-Anhalt, and Thuringia.18 Only the Bundestag, the Bundesrat, or the federal government may petition the Federal Constitutional Court in Karlsruhe; no such petition is treated here as filed. For firms the immediate consequence is counterparties and reputation, especially in the East.

What Firms Are Pricing

Base 70 percent: Brandmauer holdsIndustrial relief and stricter migration
Partial 25 percent: state executive roleRegional disruption and EU friction
Tail 5 percent: federal executive pathGeneration-mix reversal and alliance shock

The probable path — by some distance, perhaps seven chances in ten — is that the Brandmauer holds. The AfD remains locked out of the federal cabinet and still forces the coalition toward border enforcement and factory energy subsidies. For bulk power users that near-term climate is mixed-to-favourable: factory-price relief without a change of government.

A state-level breach, nearer one in four than a shock, would raise regional friction and noise in Brussels. A minister-president, a ministry, or a tolerated minority would not itself write a Dexit, restart Nord Stream — the Baltic pipelines that once brought Russian gas to Germany — or abolish the solidarity surcharge. It would show the firewall is a declining convention.

The tail, perhaps one in twenty, is a material route toward federal executive participation: generation-mix reversal, a fight over Ukraine aid, deep friction with the EU and the United States. Exporters would carry that exposure, not the domestic power bill. The shares in the diagram are a judgement. A national poll that prints 26 instead of 28 leaves them where they are.

After Magdeburg

Two polls put the AfD at 42–43% in Sachsen-Anhalt ahead of 6 September: Infratest dimap at 42% and INSA at 43%.6, 7 The question that matters is what happens after such a result. A first-place finish followed by a united refusal to govern with the AfD would leave the firewall as law and raise the pressure on the Union to compete on migration and energy. An AfD executive role would do the reverse: Dexit would remain distant, but the assumption that has kept the high-delivery path remote would be broken.

The 20 September elections are the second test. In Mecklenburg-Vorpommern the last named Infratest poll has the AfD at 35% and the SPD at 32%.8 In Berlin, Infratest has the AfD at 18%, behind the CDU at 21% and The Left at 19%.9 A strong eastern-rural result beside a weaker metropolitan showing would be geographic polarisation and harder state arithmetic — the March western thirds already point that way. Only Sunday’s count in Magdeburg, and the 20 September counts in Schwerin and Berlin, are results.

Over the next year the federal Brandmauer should hold. Apocalyptic scenarios — Dexit, a sudden restart of Nord Stream, the abolition of key levies — remain structurally blocked. The Union–SPD coalition will keep executing AfD-lite on borders and factory bills. A two-point miss in tomorrow’s poll leaves that reading intact; a durable breach of the firewall rewrites it.

Watch the coalitions. If the AfD finishes first and still faces a united opposition bloc, the firewall holds as law; if any local branch of an established party forms a working majority with it, the tail moves closer.

Watch the courts. Cologne’s merits ruling can change compliance judgements without deciding who governs.

The definitive takeaway for this month's balance sheet is not how high the AfD climbs in the polls, but whether any established party is finally willing to share power with them.

References

  1. dawum, Infratest dimap Bundestag poll published 3 September 2026 (field 31 Aug–2 Sep 2026), https://dawum.de/Bundestag/Infratest_dimap/2026-09-03/
  2. wahlrecht.de, Forsa time series, row 1 September 2026, https://www.wahlrecht.de/umfragen/forsa.htm
  3. Bundeswahlleiterin, “Bundestagswahl 2025: Endgültiges Ergebnis”, 14 March 2025, https://www.bundeswahlleiterin.de/info/presse/mitteilungen/bundestagswahl-2025/29_25_endgueltiges-ergebnis.html
  4. Landeswahlleiterin Baden-Württemberg, final result, 27 March 2026, https://www.baden-wuerttemberg.de/de/service/presse/pressemitteilung/pid/endgueltiges-ergebnis-der-landtagswahl-2026
  5. Statistisches Landesamt Rheinland-Pfalz, Landtagswahl 2026 Datenblick, https://www.datenblick.statistik.rlp.de/landtagswahl-2026
  6. dawum, Sachsen-Anhalt Infratest dimap 26 August 2026, https://dawum.de/Sachsen-Anhalt/Infratest_dimap/2026-08-26/
  7. dawum, Sachsen-Anhalt INSA, published 2 September 2026, https://dawum.de/Sachsen-Anhalt/INSA/2026-09-02/
  8. NDR, “MV-Trend”, 3 September 2026, https://www.ndr.de/nachrichten/mecklenburg-vorpommern/landtagswahl/mv-trend-spd-verkuerzt-rueckstand-zur-afd-cdu-rutscht-weiter-ab,mvtrend-168.html ; dawum https://dawum.de/Mecklenburg-Vorpommern/Infratest_dimap/2026-09-03/
  9. dawum, Berlin Infratest dimap 2 September 2026, https://dawum.de/Berlin/Infratest_dimap/2026-09-02/
  10. dawum, AfD page, retrieved 5 September 2026, https://dawum.de/AfD/
  11. BAMF, asylum figures 2025, 12 January 2026, https://www.bamf.de/SharedDocs/Meldungen/DE/2026/260112-asylzahlen-dezember-und-gesamtjahr-2025.html?nn=284830
  12. Grundgesetz Art. 115, https://www.gesetze-im-internet.de/gg/art_115.html
  13. BMF monthly report June 2025, Solidaritätszuschlag, https://www.bundesfinanzministerium.de/Monatsberichte/Ausgabe/2025/06/Inhalte/Kapitel-3-Analysen/3-2-solidaritaetszuschlag-vor-bverfge.html
  14. Destatis, PM 26 August 2026, inheritance and gift tax 2025, https://www.destatis.de/DE/Presse/Pressemitteilungen/2026/08/PD26_306_736.html
  15. BDEW, Strompreisanalyse August 2026, https://www.bdew.de/media/documents/BDEW_Strompreisanalyse_082026.pdf
  16. deutschland.de, “Deutschlands Industrie”, https://www.deutschland.de/de/topic/wirtschaft/deutschlands-industrie-die-wichtigsten-zahlen-und-fakten
  17. VG Köln, press release 26 February 2026, https://www.vg-koeln.nrw.de/behoerde/presse/Pressemitteilungen/05_26022026/index.php
  18. Correctiv, “Wie stuft der Verfassungsschutz die AfD in den Bundesländern ein?”, 17 June 2026, https://correctiv.org/aktuelles/afd/2026/06/17/wie-stuft-der-verfassungsschutz-die-afd-in-den-bundeslaendern-ein/
Source: afd_article_reworded_tone.md. Self-contained HTML; no other files required.

Method

Harvest, Warehouse, View

A seven-step way to turn model dumps into a piece a senior reader can be wrong about

A one-day method for turning several AI drafts into a piece that takes a falsifiable position — with every figure traceable to a public source. Below: the general method first, then a worked run on one topic (Germany's AfD, 5 September 2026) so you can copy the pattern onto a different question.

Core idea: Harvest is hypothesis, not fact. A number without publisher, date, and URL is dropped. A human, not a model, judges whether a draft is description or a view. Conviction and voice are separate passes — never do both at once.

The arc

1. Decision question    →  4 lines on paper, before any model
2. Harvest               →  3–4 models, same brief, same day
3. Warehouse              →  every number sourced or dropped; Fact/Inference/Hypothesis
4. Prune                  →  cut to the argument; still just description
5. Human gate              →  "pupil" or "publish"? only a person can say
6. Elevate                →  steal method from a model that won't bet; rewrite for conviction
7. Voice pass              →  same bet, same figures, different clothes

Quality here didn't come from more facts or a better model — it came from a numbers rule, a fact/inference/hypothesis taxonomy, a human refusal to call it finished, and two separate final rewrites instead of one.


Step 1 — Write four lines before touching a model

Decision question:   What should [audience] do / believe / watch, because of [event], by [date]?
Audience:            [who this is actually for — be specific]
Forbidden base case: [the dramatic scenario you will NOT price unless a named condition hits]
Kill-condition:      [the single observation that would make you recant]

If you can't fill these in, you don't yet know what the piece is for. You may still harvest, but don't write the article — you'll pay for it later in painful pruning.

Worked example:

Decision question:   What should firms do with an AfD poll lead before two state elections?
Audience:            International corporate strategy / risk
Forbidden base case: Dexit, a Russia reset, an AfD-led federal cabinet
Kill-condition:      A durable breach of the "firewall" — a minister-president post,
                     ministries, or an explicit mainstream-party retreat from refusal to ally

(In the real run these were written after the harvest, not before — and it showed. Do it first.)


Step 2 — Harvest: several engines, same question

Send an identical brief to 3–4 different models the same day. Save every reply as its own file; never edit a draft in place. Ask for named figures with sources, institutional constraints vs. stated platform, and what would change the picture — and instruct models explicitly not to average numbers or treat each other's output as a source.

Drafts will agree on shape and disagree on numbers. That's expected: treat all of it as a hypothesis list, not as anything publishable.

Worked example: four engines, one shared brief covering five policy pillars (immigration, energy, sovereignty, fiscal, culture), each asked to go deeper on a different angle — legal/electoral status, international/fiscal implications, a feasibility table, sector-specific spillover. Eight raw files saved; none cited for a number afterward.


Step 3 — Warehouse: one pass, two rules

Feed every draft into one model with two non-negotiable rules:

  1. A draft is narrative, never a source for a figure.
  2. Every number needs publisher, date, URL — or it's dropped and logged. Where drafts conflict, log the conflict rather than silently picking a side.

Output three files:

  • a warehouse memo — encyclopedic, comprehensive, allowed to be too long
  • a source map — claim ID → which draft/engine → the public citation that actually backs it
  • a claim ledger — same IDs tagged Fact (an election result, court order, enacted text, a named poll) / Inference (follows from facts under a stated assumption) / Hypothesis (testable, not yet shown), with a note on what would overturn each

Worked example: the four harvested + companion drafts fed into one pass with a public-source check (statistics office, pollster sites, ministry releases, court press releases) against every number, producing the warehouse memo plus the two ledger files.


Step 4 — Prune like a journalist, not an analyst

Cut the warehouse to the essential argument for your actual audience. Every remaining number keeps its citation. Gloss unfamiliar names, dates, and jargon on first use, as if for a reader who doesn't know the country or the topic. Don't reach for a forecast yet — the job here is an honest, tight description, nothing more.

This produces a separate *_article.md. Never overwrite the warehouse file to make it — keep both.

Worked example: the encyclopedic memo cut to a Wall-Street-Journal-style piece; then a second light pass added glosses for an international reader (what a Landtag is, when reunification happened, who runs which poll).


Step 5 — The human gate

No model can tell you whether a piece is description or a view. Only a person willing to put their name on it — or refuse to — can. This is the hinge of the whole method: everything before it is coverage, everything after is an attempt to place a bet without inventing facts.

Send only the pruned article (not the warehouse, not the ledgers) to a senior reader. Ask exactly three questions, nothing more:

Is this description, or a view?
If I'm wrong in six months, can I say what I expected?
Would you put your name on this outside the building?
You hearVerdictDo next
"Excellent summary / strong / useful internally"PupilDon't add facts — go write the elevation prompt (Step 6)
"I disagree with the view but I see the bet"Publishable, with editsLight edit or voice pass only
"I'd put my name on this"PublishCopy-edit and ship

The dangerous answer is "excellent, and unpublished." That means facts and description are already good enough — what's missing is conviction with a receipt, not more research.

Worked example: the reviewer called the article detailed, accurate, comprehensive, and "no worse than a leading news weekly" — but with no clear direction and no actionable conclusion. Verdict: pupil work. Nothing was added; the next step was the elevation prompt.


Step 6 — Steal the method, then take the bet

A model that refuses to take a position can still describe how a confident analyst would structure the argument — pyramid principle, scenario weighting, where a thesis statement belongs. Extract that technique (even from a model that declined the actual bet) and fold it into a prompt sent to a model that will commit to a view.

Elevation prompt skeleton:

You are a senior analyst for [audience]. Facts are solid; a senior reader has
called this unpublished description. Your job is not more background — it is
to take a view.

Do all three:
1. Critique — why is this strong pupil work, not master work?
2. Editorial principles a writer can apply tomorrow.
3. A complete rewrite — keep 100% of sourced figures; restructure only.

Rules: no hedging that undoes the thesis; no invented numbers; no deleting
sourced facts to sound bold; open with the view, drivers, and kill-condition
in ~100 words; three ranked paths with a clear base case; every retained
figure keeps its citation.

Worked example: one model refused to opine on the political outcome twice, but under a "method only" reframe produced a solid write-up of BLUF structure and scenario framing. That got folded into the elevation prompt, which was then sent — with the pruned article, the claim ledger, the source map, and the warehouse memo attached — to a model willing to commit. Result: a clear governing view (agenda capture with the firewall intact, not a change of government, roughly 70/25/5 across three scenarios).


Step 7 — Two rewrites, never mixed

Pass A — conviction (may still sound like an internal memo):

  • One falsifiable sentence in the opening ~100 words
  • Named drivers, hold-condition, kill-condition
  • Three ranked paths; state which one you'd bet on
  • Facts placed under the argument they support
  • No new numbers

Pass B — voice (same bet, same figures, different register):

Rewrite for outside publication. Same bet, same figures — different clothes.
Register: magazine analysis, not an internal risk memo. Let the reader infer
the call; don't label it. Title: two short sentences that collide. After a
fact, don't announce what it proves — advance the argument. No takeaway box,
no action-column table. Close on one quotable sentence, not a recap.
Test: if it reads like a manager being told what to do, rewrite until it
reads like a reader being shown why.

Do these separately. Mixing them in one pass tends to produce either a timid explainer or an unreadable memo.

Worked example: Pass A rewrote the piece around the elevation report's view, in desk diction (hold/reprice-style language). Pass B then rewrote that, in magazine diction, closing on a single quotable line rather than a recommendation — this is the version you're reading in the companion article.


Standing rules

  1. Model prose is a witness, not evidence. Official URLs are evidence.
  2. Named institute + fieldwork date, or the poll doesn't get used.
  3. Two honest numbers beat one blended, unattributed range.
  4. A classification (e.g. a legal/regulatory label) is not the same as a policy prohibition — say so explicitly so the reader doesn't conflate them.
  5. News-quality description is the floor. A view is the actual product.
  6. If a sourced fact disappears during the "elevation" pass, that's cheating, not editing.
  7. A closing "of course, anything could happen" is a quiet recantation of the whole piece — don't do it.
  8. Keep the method notes separate from the published piece; the reader doesn't need the scaffolding.

Time budget

If you haveDoSkip
~90 minutesFour lines, one harvest, the numbers rule, a pruneMulti-engine harvest, layout
Half a day+ a second engine, the claim taxonomy, a foreign-reader glossHuman gate and elevation
A full day+ human gate, elevation, Pass A, Pass B, layoutNothing that deletes a sourced figure

The 90-minute version is an internal note — useful, but don't mistake it for the finished piece.

Before calling it done

  • [ ] Four lines written down (and, ideally, written first)
  • [ ] Every remaining figure has a live URL, or sits in a visible "dropped" list
  • [ ] No hypothesis dressed up as a fact
  • [ ] Warehouse and pruned article still exist as separate files
  • [ ] A human has said "pupil" or "publish" — a model never gets to make that call
  • [ ] If "pupil": Pass A contains a sentence that can be proven wrong
  • [ ] Pass B no longer sounds like an internal memo
  • [ ] A reader unfamiliar with the topic survives the first page
  • [ ] The published piece doesn't explain its own method
Source: method_compact.md. Self-contained HTML; no other files required.

Usage

Token usage — AfD political memo run

Reconstructed spend for 5 September 2026 (Cursor + browser engines)

6 September 2026

How many tokens this one-day AfD project used, and by which model. Cursor’s local chat records do not store billed input or output — every bubble has inputTokens = 0. The reconstruction below is what you can attribute to this folder without the account-wide dashboard.

The number you can trust for Cursor spend is the CSV at cursor.com/dashboard/usage. Filter to 5–6 September 2026 and the chat windows in the table. The dashboard has no project column; time-filtering is how you attach spend to this repo.

Method

  1. Cursor billed truth — export the usage CSV; keep rows whose timestamps fall in the five windows below; sum input, output, and cache-read by model.
  2. Local reconstruction — read each chat’s composer metadata: model name, last-turn context, completed generations. Uncached input band ≈ generations × (fixed overhead + half of final conversation). Unique-token floor ≈ sum of last-turn context meters.
  3. Engines outside Cursor — tokenize saved markdown with chars ÷ 4 (±25%). That is an output floor. Chrome and Deep Research prompts, tool traces, and hidden reasoning are not in the repo unless the product’s usage page still has them.

Do not sum last-turn context across chats and call it billed. Last-turn is a snapshot; billed input is snapshot × generations, minus cache.

Cursor agent — 5 September 2026

Five project chats. Almost all IDE work ran on grok-4.6 (high effort). One short evening chat used Cursor Auto. The two long sessions dominate: each sat near a 200k-token window and replayed it ~300 times.

ModelChatsGenerationsLast-turn contextUncached input band
grok-4.6 (high)4727~658k~81M
Cursor Auto133~66k~1.4M
Project total5760~724k~82M

Treat ~82M as the no-cache ceiling and ~724k as the unique-tokens floor. Billed usage sits between them once cache hits are counted.

ChatWhen (UTC+2)ModelGenerationsLast contextUncached band
HTML kit + memo + article07:58–08:52grok-4.6306213,558~35.8M
Gemini prompt + HTML17:17–19:43grok-4.6302227,909~36.9M
Style comparison19:21–19:37grok-4.668134,679~5.2M
Git commit19:51–19:53grok-4.65182,233~2.5M
Command fix20:48–21:28Auto3365,909~1.4M

On the Gemini-prompt chat (227,909 of 256k; 89% full), 92% of last-turn context was conversation, not files.

Outside Cursor — saved output only

Chrome drafts, Gemini Deep Research, and OpenAI Deep Research ran outside the IDE. Figures below are file tokens only (chars ÷ 4).

EngineSaved tokensCaveat
Gemini Deep Research~30kStreams overlap; billed job is larger
OpenAI Deep Research~9.5kReport only
Grok harvest drafts~8.0kUnique files
Claude Chrome~7.8kOutput draft
Mistral~5.3kOne file
Gemini Chrome~3.0kOutput draft
Saved non-Cursor output~64kFloor, not billed total

Deep Research jobs are often 10–100× the saved report once browsing and hidden reasoning are included.

How to tighten next time

  1. Join the Cursor usage CSV to the five time windows above.
  2. For Gemini / Claude / Grok / Mistral / OpenAI, check the original product’s usage page if the account still has the jobs.
  3. Keep a short research/usage.md on each run: model, chat window, link to the CSV row set.