Anthropic & OpenAI ARR Forecast 2026–2027
On our estimates, Anthropic is running at about $86B of annualized revenue and OpenAI at about $62B. Our base case has them at $96B and $73B at the end of 2026, and $205B and $200B at the end of 2027.
ARR here is the annualized revenue run-rate: about a month’s revenue × 12. Anthropic’s own measure, in its IPO filing, annualizes usage fees over the last 28 days and adds a year of subscriptions at today’s rate. Ranges are 80%. Each company on its own accounting: Anthropic counts sales through cloud marketplaces in full, OpenAI only its share, so Anthropic’s figures are about 7.5% higher than OpenAI’s method would give.
Run-rates since 2024, and where they’re heading
Annualized revenue, log scale. Solid lines are our estimate, fitted to every reported figure; the thinner stretch at the end is since each company’s last report, which for Anthropic also uses spend through OpenRouter. Dashed lines are the base case and shading its 80% range.
Data
| Month end | Anthropic | 80% range | OpenAI | 80% range |
|---|---|---|---|---|
| Nov 19, 2026 (latest) | $86B | $79–94B | $62B | $56–68B |
| Nov 2026 | $89B | $82–98B | $64B | $59–70B |
| Dec 2026 | $96B | $87–110B | $73B | $67–78B |
| Jan 2027 | $105B | $92–125B | $81B | $74–88B |
| Feb 2027 | $110B | $96–140B | $88B | $80–98B |
| Mar 2027 | $120B | $100–160B | $97B | $87–110B |
| Apr 2027 | $130B | $105–180B | $105B | $93–125B |
| May 2027 | $140B | $110–205B | $115B | $99–140B |
| Jun 2027 | $145B | $110–230B | $125B | $105–160B |
| Jul 2027 | $155B | $115–260B | $140B | $110–180B |
| Aug 2027 | $165B | $120–290B | $150B | $120–200B |
| Sep 2027 | $175B | $120–320B | $160B | $125–220B |
| Oct 2027 | $185B | $125–355B | $175B | $130–245B |
| Nov 2027 | $195B | $130–390B | $185B | $135–270B |
| Dec 2027 | $205B | $130–430B | $200B | $140–300B |
Full history and forecast: openai-anthropic-run-rate.csv
Where the forecast was
Our latest forecast, in color, against an earlier one, in gray: pick how far back. Log scale, the same for both companies; shading is the latest forecast’s 80% range. Each forecast is dated by the last day of data it uses, and the last run of each day is kept, as published.
None of the figures the model uses were published in between. Anthropic’s forecast takes in one more day of OpenRouter data; OpenAI’s is unchanged.
Month by month
| Month end | Anthropic, latest | Anthropic, Nov 18 | OpenAI, latest | OpenAI, Nov 18 |
|---|---|---|---|---|
| Nov 2026 | $89B | $89B | $64B | $64B |
| Dec 2026 | $96B | $96B | $73B | $73B |
| Jan 2027 | $105B | $105B | $81B | $81B |
| Feb 2027 | $110B | $110B | $88B | $88B |
| Mar 2027 | $120B | $120B | $97B | $97B |
| Apr 2027 | $130B | $130B | $105B | $105B |
| May 2027 | $140B | $135B | $115B | $115B |
| Jun 2027 | $145B | $145B | $125B | $125B |
| Jul 2027 | $155B | $155B | $140B | $140B |
| Aug 2027 | $165B | $165B | $150B | $150B |
| Sep 2027 | $175B | $175B | $160B | $160B |
| Oct 2027 | $185B | $185B | $175B | $175B |
| Nov 2027 | $195B | $195B | $185B | $185B |
| Dec 2027 | $205B | $205B | $200B | $200B |
None of the figures the model uses were published in between. Anthropic’s forecast takes in 7 more days of OpenRouter data; OpenAI’s is unchanged.
Month by month
| Month end | Anthropic, latest | Anthropic, Nov 12 | OpenAI, latest | OpenAI, Nov 12 |
|---|---|---|---|---|
| Nov 2026 | $89B | $88B | $64B | $64B |
| Dec 2026 | $96B | $94B | $73B | $73B |
| Jan 2027 | $105B | $100B | $81B | $81B |
| Feb 2027 | $110B | $110B | $88B | $88B |
| Mar 2027 | $120B | $115B | $97B | $97B |
| Apr 2027 | $130B | $125B | $105B | $105B |
| May 2027 | $140B | $130B | $115B | $115B |
| Jun 2027 | $145B | $140B | $125B | $125B |
| Jul 2027 | $155B | $145B | $140B | $140B |
| Aug 2027 | $165B | $155B | $150B | $150B |
| Sep 2027 | $175B | $160B | $160B | $160B |
| Oct 2027 | $185B | $170B | $175B | $175B |
| Nov 2027 | $195B | $180B | $185B | $185B |
| Dec 2027 | $205B | $190B | $200B | $200B |
Between these two forecasts the method changed (Nov 10): Anthropic’s amended filing put its run-rate at about $81B on October 31, up $10B from September 30; the tracker had October 31 at $93B. Spend on Anthropic’s models through OpenRouter rose 88% between the four weeks to those two dates, and the tracker had read a launch month on OpenRouter as a level of revenue. From now a move in OpenRouter, up or down, moves the estimate by only part of its size: little just after one of Anthropic’s own figures, more as weeks pass without one (see Anthropic between reports). Anthropic’s forecast growth is the fitted path’s pace over the last three months, with no separate blend of paths, and its range lets the starting growth vary by 50% rather than 25%, because the narrower range held too few past outcomes. Reported figures now count from the day they are published rather than a day later. On data through Nov 9, the two new figures lower Anthropic’s estimate from $98B to $83B and its forecasts from $115B to $94B for December 2026 and from $315B to $185B for December 2027. The change of method takes those two on to $93B and $180B and leaves the estimate at $83B; it matters from here, as OpenRouter data arrives without new figures. It also changed (Nov 6): OpenAI’s forecast growth after its year-end target moves from its pace over the last three months to its pace over the last 12. Its run-rate steps up around its biggest launches, about twice in the last 12 months (from $13B to $21.4B around GPT-5.1 and 5.2 at the end of 2025, and from $29B to $50B around GPT-5.6 and GPT-6 this summer), growing about 5% a month in between, so 12 months spans whole cycles, while the last three months sit on the GPT-6 step. Tested on forecasts made 30, 61 and 91 days before each OpenAI figure since mid-2025, the 12-month pace missed less at all three. December 2026 still follows OpenAI’s own target. If its launches come faster or slower, we’ll change it. On data through Nov 5, the change lowers OpenAI’s forecast for December 2027 from $240B to $200B. It also changed (Nov 6): OpenRouter readings of Anthropic go back from ±15% to ±20%, and Anthropic’s forecast stops treating its latest launch step as monthly growth. Its launches come in waves, and the big ones lift revenue for about a month before fading; the old rule carried October’s step (Opus 5.5) forward at 16% a month, which repeats it every three months. Three of Anthropic’s eight launch waves in the last 12 months lifted its OpenRouter spend by 40% or more, about one every four months, so the forecast now assumes that pace (see Launch pacing). If launches come faster or slower, or their steps get bigger or smaller, we’ll change it. Q2 and Q3 revenue from the amended S-1, $11.56B and $16.96B, also go in, lifting the forecast about 3%. It began as a 14-day phase-in and went in fully the same day. On data through Nov 4, the change lowers Anthropic’s estimate from $101B to $96B and its forecasts from $135B to $120B for December 2026 and from $505B to $315B for December 2027. It also changed (Oct 22): OpenRouter readings of Anthropic now enter the fit at ±15% rather than ±20%, the setting with the smallest median miss in the backtest. On the same data, through Oct 21, this raised Anthropic’s estimate from $83B to $87B and its forecasts from $100B to $110B for December 2026 and from $225B to $285B for December 2027. Figures the model uses that were published in between: Anthropic’s run-rate of $71B on Sep 30, published Nov 10; Anthropic’s run-rate of $81B on Oct 31, published Nov 10. Anthropic’s forecast takes in 31 more days of OpenRouter data.
Month by month
| Month end | Anthropic, latest | Anthropic, Oct 19 | OpenAI, latest | OpenAI, Oct 19 |
|---|---|---|---|---|
| Oct 2026 | $80B | $84B | $57B | $57B |
| Nov 2026 | $89B | $90B | $64B | $64B |
| Dec 2026 | $96B | $98B | $73B | $73B |
| Jan 2027 | $105B | $105B | $81B | $82B |
| Feb 2027 | $110B | $115B | $88B | $91B |
| Mar 2027 | $120B | $120B | $97B | $105B |
| Apr 2027 | $130B | $130B | $105B | $115B |
| May 2027 | $140B | $135B | $115B | $125B |
| Jun 2027 | $145B | $145B | $125B | $140B |
| Jul 2027 | $155B | $155B | $140B | $155B |
| Aug 2027 | $165B | $165B | $150B | $170B |
| Sep 2027 | $175B | $175B | $160B | $185B |
| Oct 2027 | $185B | $185B | $175B | $205B |
| Nov 2027 | $195B | $190B | $185B | $220B |
| Dec 2027 | $205B | $200B | $200B | $240B |
The latest against earlier forecasts
Below each earlier figure, how far the latest is above or below it. Every day’s estimate and year-end forecasts: openai-anthropic-forecast-log.csv
| LatestNov 19 | A day earlierNov 18 | A week earlierNov 12 | A month earlier†Oct 19 | |
|---|---|---|---|---|
| Anthropic | ||||
| Run-rate that day | $86B | $86B+0.4% | $84B+3% | $81B+7% |
| End of 2026 | $96B | $96B+0.2% | $94B+2% | $98B−1% |
| End of 2027 | $205B | $205B+1% | $190B+9% | $200B+2% |
| OpenAI | ||||
| Run-rate that day | $62B | $61B+0.4% | $60B+3% | $54B+14% |
| End of 2026 | $73B | $73Bunchanged | $73Bunchanged | $73Bunchanged |
| End of 2027 | $200B | $200Bunchanged | $200Bunchanged | $240B−17% |
† The method changed between this forecast and the latest, on Oct 22 and on Nov 6 and on Nov 10: see the note under the charts and Changes, under Method.
The record starts with the forecast dated Oct 18, 2026, so comparisons with 3 months ago begin with Jan 18, 2027’s and with 6 months ago with Apr 18, 2027’s.
Where today’s numbers come from
Neither company reports its run-rate on a schedule. Figures arrive through funding announcements, investor updates and press reports. We fit one smooth path through each company’s figures, 42 of them so far; it misses the typical Anthropic figure by 2% and the typical OpenAI figure by 1%. Anthropic’s last figure was a run-rate of $81B, published Nov 10, 2026 (Anthropic S-1/A (Nov 10), via Reuters). OpenAI’s was a run-rate of $50B, published Oct 8, 2026 (Reuters, via The Standard).
Anthropic between reports. Spend on Anthropic’s models through OpenRouter is a small slice of its revenue, 1.9% on average over the last 5.5 months, and a jumpy one: measured over four weeks at a time it ran from 1.3% to 2.3% in that span, up after a launch and down between launches. Between Anthropic’s last two run-rates, for Sep 30 and Oct 31, that spend rose 88% and the run-rate rose 14%. So a move in OpenRouter, up or down, moves the estimate by only part of its size, and by more the longer it has been since Anthropic’s last figure: about 11% of it 30 days after a figure, 31% after 61 and 48% after 91. Today, 19 days after the figure for Oct 31, the latest reading (Oct 23–Nov 19) is 2.2% of our estimate, and a 10% move in it would move the estimate by 0.8%. Anthropic’s reported figures alone give $85B; with OpenRouter, $86B. Tested on the 17 run-rates and quarterly revenue figures reported for Anthropic since May 2025, each estimated the day before it came out, this missed by 6% at the median, against 14% from Anthropic’s reports alone. The settings were chosen on these same figures, so expect less of a gain from here.
OpenAI between reports. Most of OpenAI’s revenue is ChatGPT and direct enterprise sales that OpenRouter doesn’t see, so its estimate moves only when new figures come out. Until then it follows a path toward OpenAI’s own target, a run-rate of at least $70B by December 2026 (Bloomberg, via Bloomberg Línea, Oct 9, 2026). OpenAI has come in about 3% above its near-term targets on average, so the path aims at about $72B.
What OpenRouter says about Anthropic
Spend on Anthropic’s models through OpenRouter over the past year, annualized and divided by 1.90%, its average share of Anthropic’s revenue over the last 5.5 months, against our estimate and reported figures. Log scale. The heavy segments are four-week readings, one every four weeks. Where they sit above our estimate, as after a launch, OpenRouter’s share is running above that average, and the model counts only part of the gap as revenue. By calendar quarter, counting quarters where Anthropic’s revenue is our estimate, the share has ranged from 1.6% to 4.2% since 2025.
OpenRouter against Anthropic’s own figures
| Between | Anthropic’s run-rate | OpenRouter spend |
|---|---|---|
| Dec 31 ’25 to Feb 12 | +56% $9B to $14B | +84% |
| Feb 12 to Feb 28 | +36% $14B to $19B | +31% |
| Feb 28 to Apr 6 | +58% $19B to $30B | +53% |
| Apr 6 to Apr 29 | +33% $30B to $40B | +28% |
| Apr 29 to May 7 | +18% $40B to $47B | +5% |
| May 7 to Jul 31 | +38% $47B to $65B | +105% |
| Jul 31 to Sep 30 | +9% $65B to $71B | −29% |
| Sep 30 to Oct 31 | +14% $71B to $81B | +88% |
Each row sets two consecutive reported run-rates beside spend on Anthropic’s models through OpenRouter over the four weeks to the same two dates, the span Anthropic’s own run-rate annualizes.
Quarter by quarter
| Quarter | Spend through OpenRouter | Anthropic revenue | Ratio |
|---|---|---|---|
| 2025Q1 | $14M | $328M (our estimate) | 4.21% |
| 2025Q2 | $20M | $787M | 2.50% |
| 2025Q3 | $30M | $1.37B (our estimate) | 2.21% |
| 2025Q4 | $32M | $2.04B (our estimate) | 1.56% |
| 2026Q1 | $84M | $4.73B | 1.77% |
| 2026Q2 | $238M | $11.6B | 2.06% |
| 2026Q3 | $285M | $17B | 1.68% |
Daily tokens and estimated spend: openrouter-daily.csv
How the estimate did, figure by figure
| Published | Figure | Reported | Estimate the day before | Reports only |
|---|---|---|---|---|
| Nov 10, 2026 | Run-rate, Oct 31 | $81B | $80.6B | $77.7B |
| Nov 10, 2026 | Run-rate, Sep 30 | $71B | $73.7B | $73.5B |
| Oct 14, 2026 | Q3 2026 revenue | $17B | $16.9B | $17.8B |
| Aug 17, 2026 | Run-rate, Jul 31 | $65B | $75.8B | $82.4B |
| Aug 14, 2026 | Q2 2026 revenue | $11.6B | $12.8B | $13.8B |
| May 28, 2026 | Run-rate, May 7 | $47B | $44.3B | $44.5B |
| May 20, 2026 | Q1 2026 revenue | $4.73B | $3.99B | $3.98B |
| Apr 29, 2026 | Run-rate, Apr 29 | $40B | $41.1B | $41.2B |
| Apr 6, 2026 | Run-rate, Apr 6 | $30B | $27.3B | $26.1B |
| Mar 3, 2026 | Run-rate, Feb 28 | $19B | $16.1B | $15.7B |
| Feb 12, 2026 | Run-rate, Feb 12 | $14B | $11.5B | $10.4B |
| Jan 21, 2026 | Run-rate, Dec 31 | $9B | $8.99B | $9.75B |
| Oct 15, 2025 | Run-rate, Oct 15 | $7B | $6.98B | $9.06B |
| Jul 30, 2025 | Run-rate, Jul 31 | $5B | $4.72B | $5.01B |
| Jul 1, 2025 | Run-rate, Jul 1 | $4B | $3.57B | $3.83B |
| May 30, 2025 | Run-rate, May 31 | $3B | $3.12B | $3.8B |
| May 16, 2025 | Run-rate, Mar 31 | $2B | $1.54B | $1.58B |
Each figure is re-estimated with today’s method, from the figures published and the OpenRouter data available the day before it came out.
Tokens and dollars through OpenRouter
Seven-day averages, log scale. Over the last four weeks Anthropic’s models took 39% of the money spent through OpenRouter on 6% of its tokens, and OpenAI’s took 12% on 8%. OpenAI’s open-weight models, which other companies serve, are under everyone else. Totals, the US–China split and forecasts are in the OpenRouter tracker.
How the forecast works
Each forecast starts from today’s estimate and assumes growth halves every 18 months, the rule in pack 008. For Anthropic, growth starts at the pace of the fitted path over the last three months, 8.6% a month: the path rose from $67B around Aug 20 to $86B. What Anthropic itself is reported to expect is not used: annualized revenue staying below $100B (Financial Times, via Yahoo Finance, Oct 15), and a run-rate of $110B by the end of the year (Reuters, via Business Standard, Oct 15). For OpenAI, the path to December 2026 bends toward its target; after that, growth starts at the path’s pace over the last 12 months, 11% a month, rather than the last three months’ 13%: OpenAI’s run-rate steps up around its biggest launches, about twice in the last 12 months, so 12 months spans whole cycles of steps and slower stretches (see Launches, below). Carrying the last three months’ pace forward instead, as before November 2026, would give $240B for December 2027. The ranges vary today’s level, the starting growth and the half-life, from 6 months to 4½ years. Starting growth varies by 50% either way for Anthropic and 25% for OpenAI (one standard deviation), which puts four of five draws for Anthropic between 4% and 17% a month.
The half-life decides 2027. Rules like this one ran too low on both companies while they sped up; pack 008 tests them against every figure published since 2024. Since mid-2026 they have run too high on Anthropic, whose run-rate grew about 50% a month from January to early May 2026 and about 10% a month from then to the end of October. On the base case each company still adds more run-rate in 2027 than in 2026: about $110B for Anthropic, against $86B in 2026, and $125B for OpenAI, against $52B.
| Growth halves every | Anthropic, Dec 2027 | OpenAI, Dec 2027 |
|---|---|---|
| 6 months | $150B | $145B |
| 12 months | $185B | $180B |
| 18 months (base) | $205B | $200B |
| 3 years | $230B | $225B |
| Never | $260B | $255B |
Launches
Anthropic’s launches come in waves, and the big ones show up first on OpenRouter: in the last 12 months, 3 of its 9 launch waves lifted spend on its models there by 40% or more in their first four weeks, about one every 4 months. Since mid-2026 Anthropic’s revenue has stepped up by much less than that spend, and kept growing while it fell (see What OpenRouter says about Anthropic). So the forecast’s growth is the fitted path’s pace over the last three months, with OpenRouter counted for only part of its own move. Three months is shorter than the gap between big launches, so that pace runs higher just after a step and lower before the next one; the range on starting growth is there to cover the difference. If launches come faster or slower, or their steps get bigger or smaller, the pace will trail them until Anthropic’s own figures show it.
| Released | Models | OpenRouter spend, first 4 weeks against the 4 before | Step |
|---|---|---|---|
| Nov 24, 2025 | Opus 4.5 | +12% | |
| Feb 5 | Opus 4.6, Sonnet 4.6 | +49% | Yes |
| Apr 16 | Opus 4.7 | +32% | |
| May 28 | Opus 4.8, Fable 5 | +59% | Yes |
| Jun 30 | Sonnet 5 | +7% | |
| Jul 23 | Opus 5 | −31% | |
| Aug 31 | Fable 5.1 | −7% | |
| Sep 21 | Opus 5.5, Sonnet 5.5, Haiku 5.5 | +59% | Yes |
| Nov 9 | Fable 5.2 | Under 4 weeks in |
A wave’s four weeks start the day its first model shows up on OpenRouter; launches within 14 days of the one before count as one wave, and share its lift, which is Anthropic’s total OpenRouter spend.
OpenAI. OpenRouter is no guide to OpenAI’s steps: its spend there rose 40% or more after almost every OpenAI launch in early 2026, while OpenAI’s revenue grew about 5% a month, because OpenRouter was still catching on for its models. The steps show in OpenAI’s reported run-rate instead, about twice in the last 12 months: from $13B to $21.4B around GPT-5.1 and 5.2 at the end of 2025, and from $29B to $50B around GPT-5.6 and GPT-6 in the summer of 2026, with about 5% a month between. So its growth after its year-end target is its pace over the 12 months to then, which spans whole cycles; the last three months sit on the GPT-6 step. If its launches come faster or slower, we’ll change that too.
Compared with pack 008. Pack 008, published the day the tracker started, put Anthropic at $265B and OpenAI at $230B in December 2027. For Anthropic, pack 008 applied the same rule to growth of about 10% a month late in 2026 and took the midpoint of that and the OpenRouter view below, with Anthropic’s share of OpenRouter settling at 30%. The tracker applies the rule to the growth it starts from, 8.6% a month, and shows the OpenRouter view separately, because it moves too much from week to week to average in every day.
What OpenRouter’s growth would imply
Spend through OpenRouter grew about 29% a month over the last three months, comparing four-week averages, up from 20% in the three before. If that growth halves every 6 months, as in pack 007, OpenRouter spend is 6.7× today’s by December 2027. Anthropic’s share of it is 39%, down from 60% at the end of January 2026. OpenRouter’s spend has swung much further than Anthropic’s revenue, so read these as what keeping pace with OpenRouter would take, not as forecasts.
| If Anthropic’s share of OpenRouter… | Dec 2026 | Dec 2027 |
|---|---|---|
| holds at 39% | $120B | $580B |
| falls at half its pace since January | $120B | $430B |
| keeps falling at that pace | $115B | $315B |
| Base case (Anthropic’s own trend) | $96B | $205B |
How the forecast has moved
Our base case for each year-end on each day since the tracker started, log scale. The last run for each day of data is kept: forecast log.
Latest reported figures
Most recent first, on each company’s own accounting. Every figure, with its source: openai-anthropic-reported.csv
| Published | Company | Figure | Value | Source |
|---|---|---|---|---|
| Nov 10, 2026 | Anthropic | Run-rate, Oct 31 | $81B | Anthropic S-1/A (Nov 10), via Reuters |
| Nov 10, 2026 | Anthropic | Run-rate, Sep 30 | $71B | Anthropic S-1/A (Nov 10), via Reuters |
| Oct 15, 2026 | Anthropic | Run-rate, Oct 15not used | $74B | SemiAnalysis, via Implicator |
| Oct 14, 2026 | Anthropic | Q3 2026 revenue | $17B | Anthropic S-1/A (Nov 2), via PPC Land |
| Oct 8, 2026 | OpenAI | Run-rate, Sep 30 | $50B | Reuters, via The Standard |
| Oct 8, 2026 | OpenAI | Run-rate, Jun 30 | $29B | The Information, via The Next Web |
| Sep 29, 2026 | OpenAI | Run-rate, Sep 30not used | $70B | Axios |
| Sep 28, 2026 | Anthropic | 2024 revenue | $386M | Reuters, via TradingKey |
| Aug 18, 2026 | OpenAI | Q2 2026 revenue | $6.7B | Wall Street Journal, via AI Weekly |
| Aug 17, 2026 | Anthropic | Run-rate, Jul 31 | $65B | Bloomberg, via Bloomberg Government |
| Aug 13, 2026 | OpenAI | Run-rate, Aug 12 | $40B | Bloomberg, via investingLive |
| Jun 17, 2026 | OpenAI | Q1 2026 revenue | $5.7B | The Information, via The Next Web |
Method
Reported figures. Run-rates, quarterly revenue and annual revenue from company statements and named outlets, each dated to the period it describes. Third-party estimates and figures on a different accounting basis are shown but not fitted. The list starts from pack 008 and is updated as new figures come out.
The path. Log run-rate on a daily grid, fitted to every figure at once: a run-rate is the path’s value on its date, quarterly or annual revenue is the path’s average over the period. Anthropic’s filing defines its run-rate as usage fees over the trailing 28 days × 13 plus monthly subscription fees × 12; we take it as the path’s value on its date like any other. A penalty on changes in growth keeps the path smooth (the maximum-likelihood path of a local linear trend model). Quarterly and annual revenue get the most weight, then run-rates stated by the companies, then run-rates reported by the press, then figures derived from reported changes.
OpenRouter. Daily tokens for the 50 biggest models plus the rest, from OpenRouter’s rankings dataset. Each model’s tokens are priced at its list price, assuming 10% are output and half of input is billed at cached rates, as in pack 007. Prices have been recorded daily since October 19, 2026; before that, each model is at the first price we recorded, retired models at their last list price, OpenAI’s GPT-5.6 Luna and Sol at their launch prices until their July and August cuts, and the long tail at the day’s average paid price. Free and pre-release models count at $0; OpenAI’s open-weight models, which other companies serve, are left out of OpenAI. For Anthropic, spend through OpenRouter enters the fit as four-week readings: one every 28 days back from the latest day of data, 6 in all, and one more ending on the date of each run-rate Anthropic has reported in that span, since its own run-rate annualizes those same four weeks. A reading is the path times OpenRouter’s share of Anthropic’s revenue. The fit finds the share’s average over the readings; around it the share swings by about 25% (one standard deviation), and a swing fades by 8% a day, so two readings a week apart share about nine-tenths of their error, two readings four weeks apart about a quarter of it, and two months apart almost none. A reading therefore moves the estimate only as far as Anthropic’s own figures leave it open: little just after a figure, more as the weeks pass. This dates from November 10, 2026, when Anthropic’s filing put its run-rate at $81B on October 31 and the tracker had it at $93B. Until then each four-week stretch of spend since the last reported quarter was divided by that quarter’s share, 1.68% for the third quarter, and counted as a figure, give or take 20%; the stretch from October 13 to November 9, when the share was about 2.2%, came out at $110B. Each of the 17 Anthropic figures since May 2025 was re-estimated as of the day before it was published, using only figures published by then and OpenRouter data through the day before (scripts/trackers/forecast_backtest.py). The new method misses by 5.9% at the median and 8.8% on average, the old one by 9.3% and 10.5%, and reported figures alone by 14% and 13%. The gain is in the five figures since June 2026, when OpenRouter began swinging several times as far as Anthropic’s run-rate: 6% on average against 10%. On the three figures from February to April 2026, when a surge on OpenRouter was matched by Anthropic’s own, the old method was closer, 12% against 15%, and a surge like that would be read too low again at first. The share is averaged over six readings because it drifts, from about 4% early in 2025 to about 1.5% at the end of that year: averaged over a year of readings, the method missed by 11% at the median. The size of the swings is measured, 0.24 in logs over the six readings to October 2026; values from 20% to 35% gave average misses of 9% to 10%.
Forecast. Two year-ends, this year’s and next, rolling forward each January. Growth halves every 18 months: g(t) = g₀ × 2−t/18. For Anthropic, g₀ is the growth of the fitted path over the last three months. For OpenAI, its run-rate target enters the fit as one more figure on its date, scaled by how its targets made within a year of their date turned out and weighted at ±6%, until figures for that month are published; g₀ is the path’s growth over the 12 months to that date, which spans its launch cycle, and the range up to it comes from the fit, which that ±6% dominates. Other ranges are the 10th and 90th percentiles of 20,000 draws of the starting level, starting growth (±50% for Anthropic and ±25% for OpenAI, in logs) and half-life (6 to 54 months). Anthropic’s was ±25% until November 10, 2026: tested on its 12 run-rates since mid-2025, the 80% range of the forecast made 30 days ahead held 8 of them and the one made 61 days ahead held 3. At ±50% they hold 9 and 6, still short of four in five, because its growth has more than once changed pace several-fold within two months. The OpenRouter scenarios grow today’s Anthropic estimate with total OpenRouter spend on pack 007’s path (the last three months’ growth, halving every six months) and move Anthropic’s share as labeled.
Updates. OpenRouter data is pulled and the model re-run twice a day; the date at the top is the last run. New reported figures are added when they are published.
Changes. Nov 10, 2026: Anthropic's amended filing put its run-rate at about $81B on October 31, up $10B from September 30; the tracker had October 31 at $93B. Spend on Anthropic's models through OpenRouter rose 88% between the four weeks to those two dates, and the tracker had read a launch month on OpenRouter as a level of revenue. From now a move in OpenRouter, up or down, moves the estimate by only part of its size: little just after one of Anthropic's own figures, more as weeks pass without one (see Anthropic between reports). Anthropic's forecast growth is the fitted path's pace over the last three months, with no separate blend of paths, and its range lets the starting growth vary by 50% rather than 25%, because the narrower range held too few past outcomes. Reported figures now count from the day they are published rather than a day later. On data through Nov 9, the two new figures lower Anthropic's estimate from $98B to $83B and its forecasts from $115B to $94B for December 2026 and from $315B to $185B for December 2027. The change of method takes those two on to $93B and $180B and leaves the estimate at $83B; it matters from here, as OpenRouter data arrives without new figures. Nov 6, 2026: OpenAI's forecast growth after its year-end target moves from its pace over the last three months to its pace over the last 12. Its run-rate steps up around its biggest launches, about twice in the last 12 months (from $13B to $21.4B around GPT-5.1 and 5.2 at the end of 2025, and from $29B to $50B around GPT-5.6 and GPT-6 this summer), growing about 5% a month in between, so 12 months spans whole cycles, while the last three months sit on the GPT-6 step. Tested on forecasts made 30, 61 and 91 days before each OpenAI figure since mid-2025, the 12-month pace missed less at all three. December 2026 still follows OpenAI's own target. If its launches come faster or slower, we'll change it. On data through Nov 5, the change lowers OpenAI's forecast for December 2027 from $240B to $200B. Nov 6, 2026: OpenRouter readings of Anthropic go back from ±15% to ±20%, and Anthropic's forecast stops treating its latest launch step as monthly growth. Its launches come in waves, and the big ones lift revenue for about a month before fading; the old rule carried October's step (Opus 5.5) forward at 16% a month, which repeats it every three months. Three of Anthropic's eight launch waves in the last 12 months lifted its OpenRouter spend by 40% or more, about one every four months, so the forecast now assumes that pace (see Launch pacing). If launches come faster or slower, or their steps get bigger or smaller, we'll change it. Q2 and Q3 revenue from the amended S-1, $11.56B and $16.96B, also go in, lifting the forecast about 3%. It began as a 14-day phase-in and went in fully the same day. On data through Nov 4, the change lowers Anthropic's estimate from $101B to $96B and its forecasts from $135B to $120B for December 2026 and from $505B to $315B for December 2027. Oct 22, 2026: OpenRouter readings of Anthropic now enter the fit at ±15% rather than ±20%, the setting with the smallest median miss in the backtest. On the same data, through Oct 21, this raised Anthropic's estimate from $83B to $87B and its forecasts from $100B to $110B for December 2026 and from $225B to $285B for December 2027. Oct 19, 2026: Tracker started, with the 44 figures collected for pack 008 and OpenRouter data from January 2025.
Source for token data: OpenRouter (openrouter.ai/rankings), as of 2026-11-20 15:34 UTC. Licensed under CC BY 4.0. Spend, run-rates and forecasts are our estimates. Not investment advice.