Coding intelligence got ~118× cheaper in 30 months.
From April 2024 to November 2026, the Chatbot Arena coding score–price Pareto frontier marched up and left: peak quality rose, and the cost of clearing yesterday’s published score collapsed. The April 2024 coding peak (~1267 at $15/MTok blended, GPT-4 Turbo 126×); that promotion ended September 9. This repo packages the weekly panels, methods, caveats, exports, and a chaptered movie of that migration.1106) is cleared the week of November 9, 2026 by gpt-oss-20b at $0.1275/MTok standard list price — about 118× lower for the old score threshold. The low point was August 31, 2026, when glm-5.3-flash cleared it at a launch-promotion $0.11875 (
“Cost of intelligence” is deliberately provocative shorthand, not a claim that Arena is an absolute intelligence meter. The exact factor depends on the benchmark slice, dates, price mix, match coverage, and pricing schedule. The claim is narrower: over 30 months, the list-price cost of exceeding one published human-preference score threshold fell by roughly two orders of magnitude.
Each week we plot matched models in (blended $/MTok, Arena Text/coding score) space. A model is on the Pareto frontier if no other matched model that week is both cheaper (or equal) and higher-scoring (or equal) with at least one strict inequality — i.e. maximum published score for a given price, and minimum price for a given score. Dominated interior points are ignored in the story charts; only the frontier polyline matters.
Blended price = (3 × input + 1 × output) / 4 USD per 1M tokens (3:1 input-heavy mix).
| Anchor | Published Text/coding score | Cheapest blended $/MTok clearing that score | Models (illustrative) |
|---|---|---|---|
| 2024-04-01 | ~1267 | $15.00 | gpt-4-1106-preview |
| 2026-08-31 snapshot | clears the old bar | $0.11875 (~126×) | glm-5.3-flash, launch promotion |
| 2026-11-09 latest week | clears the old bar | $0.12750 (~118×) | gpt-oss-20b (Fireworks); step-3.5-flash at $0.15 is ~100×, Claude Haiku 5.5 at $0.20 is ~75×, glm-5.3-flash at $0.2375 is ~63× |
Peak ceiling also rose: ~1267 → ~1542 weekly maximum priced score (claude-opus-5.5-high at $8, week of November 9, 2026, on 3,339 votes), with running-max envelope ~1553. The price of near-peak coding halved in October: gemini-4-argon-high scores ~1540 at $4 (introductory rate through November 30), 2 points below Opus 5.5. Next: Gemini 4 Carbon (~1563, #1 on the November 9 board) became generally available on November 10 at the same $4 blended introductory rate. That is one day after the latest Arena snapshot, so under the panel's timing rule it is not plotted yet. With the next snapshot it should become the top of the priced frontier: about 21 points above Opus 5.5 at half the price. See data/frontiers/coding/_summary.json and media/coding/.
Watch:
media/coding/frontier_evolution_chaptered.mp4— nine chapter freezes from the GPT-4-era stub through the ~$0.12 launch-promo tip and Opus 5.5's peak at $8 to Gemini 4 Argon at $4.
Take the model that led the coding board at some past date, and ask each week since: what is the cheapest API model rated at least as high as it, on that week's board? The answer may be a newer model — that is the point.
OpenAI o1 topped the Text/coding board in January 2025 at $26.25/MTok blended. Six weeks later o3-mini-high matched it at $1.93; by August 2025 Gemini 2.5 Flash-Lite did so at $0.175. The week of November 9, 2026 it costs $0.15 (step-3.5-flash) — about 175× cheaper in 21 months. Nearly all of that fall came in the first seven months; from August 2025 the price of o1-level coding sat at $0.175, apart from GLM-5.3 Flash's one-week $0.119 launch promotion, until Step 3.5 Flash gained a list price in our sources in October 2026.
The factor depends heavily on which past leader you pick, because some leaders were themselves bargains:
| Past board leader | Anchor week | Price then | Cheapest equal-or-better now | Decline |
|---|---|---|---|---|
o1-mini |
2024-09-16 | $1.925 | $0.15 (step-3.5-flash) |
~13× in 25 months |
o1 |
2025-01-13 | $26.25 | $0.15 (step-3.5-flash) |
~175× in 21 months |
claude-3-7-sonnet |
2025-02-24 | $6.00 | $0.1375 (gpt-5-nano-high) |
~44× in 20 months |
chatgpt-4o-latest (Mar 2025) |
2025-03-31 | $4.375 | $0.15 (step-3.5-flash) |
~29× in 19 months |
Read the o1 number with care: o1-mini sat ~4 rating points below o1 that week (well inside the noise) at under a tenth of the price, so 175× measures the fall from a flagship's sticker price, while ~13× measures the fall from the best deal available at nearly the same level. Both are true. o1-preview and gpt-4.5-preview also led the board in this window but have no dated list price in our sources, so they cannot anchor a series.
Watch:
media/coding/fixed_intelligence.mp4(~24s) · all four anchors:fixed_intelligence_anchors.png· data:fixed_intelligence_cost.csv
These are not forced onto one axis. Text/coding and WebDev are separate Arena boards, and a model is only placed on a board where it has a published rating.
| Model | Independent capability signal | 3:1 blended list price | Status |
|---|---|---|---|
| Gemini 4 Carbon | Text/coding 1563.4, rank 1 (1,072 votes); WebDev 1768.0, rank 3 | $4.00 intro ($8.00 from Dec 1) | Generally available Nov 10, the day after the latest snapshot; joins the panel with the next one |
| Claude Opus 5.5 | Text/coding 1542.2, rank 2 (-high, 3,339 votes); WebDev 1817.7, rank 1 (-max) |
$8.00 | Top of the priced coding frontier in the panel |
| Gemini 4 Argon | Text/coding 1540.0, rank 3; WebDev 1682.6 | $4.00 intro ($8.00 from Dec 1) | On the frontier, 2 points below Opus 5.5 at half the price |
| GLM-5.3 Flash | Text/coding 1509.7; WebDev 1608.6 | $0.2375 | On the frontier; cheapest model above 1500 |
| Claude Sonnet 5.5 | Text/coding 1500.2, rank 25 (-xhigh, down from 1512); WebDev 1774.3, rank 2 |
$4.00 | Inside the coding frontier |
| Kimi K3.1 Pro | Text/coding 1476.7 (Max, 774 votes, first rating); WebDev 1687.8, rank 10 |
$4.00 | Inside the frontier: Argon scores 63 points higher at the same price |
| Inkling 2 Small | Text/coding 1456.6 (616 votes, first rating); WebDev 1649.5 | $0.475 | Announced Nov 5 on Tinker, open weights planned; inside the frontier |
| Claude Haiku 5.5 | Text/coding 1454.1; WebDev 1595.1 | $0.20 | On the frontier at $0.20 |
| Muse Ember | WebDev 1608.1, rank 34 (1,345 votes); no Text/coding score yet | $0.20 | Released Nov 9; above Haiku 5.5 and GPT-6 Luna on WebDev at the same price. The model to watch at the cheap end |
| Step 3.5 Flash | Text/coding 1435.6 | $0.15 | Cheapest paid model at o1 level |
| Claude Fable 5.2 | not rated yet | $16.00 | Released Nov 9 with a 20% price cut ($8/$40) |
| MiMo v2.6 Pro | Text/coding 1533.1, rank 5 (1,091 votes) | no list price found | Excluded from the price axis until Xiaomi publishes an API price we can cite |
All scores are from the November 9 snapshots. The auditable snapshot is data/current/current_model_watch.csv. The normalized November 9 WebDev board is data/arena/webdev/arena_webdev_latest.csv.
- LMArena / Chatbot Arena leaderboard — Hugging Face
lmarena-ai/leaderboard-dataset- Primary spine: subset
text, categorycoding - Raw Text is used for longitudinal continuity: the available style-controlled coding series begins in September 2024 and would remove the April anchor
- Optional:
text/ overall (weekly CSV included; full raw Parquet not shipped) - Secondary tip chart: Code Arena WebDev (
webdev/overall) when a Text/coding score is missing
- Primary spine: subset
- Dated API prices
- RoninForge AI Price Index — https://roninforge.org/ai-price-index/ / open data (CC BY 4.0)
- Model Price Watch — https://modelpricewatch.com (supplement for models not in RoninForge)
- Blended price =
(3*input + output)/4as above - Optional secondary: Code Arena tip visuals for current models without a Text/coding score
Normalized CSVs live under data/. Huge raw JSON/Parquet dumps are not included; rebuild from upstream if you need them.
- Weekly Mondays. Arena snapshots and prices are aligned to ISO week starts (Monday). Panel coverage for the coding × price join: 2024-04-01 → 2026-11-09 (125 weeks; no snapshot was published in the week of September 14). Each week uses its last published snapshot and the Monday list price, with one rule for mid-week changes: if a price took effect before the week's snapshot, that price is used (GLM-5.3 Flash's standard rate for week 2026-09-07; the launch prices of Opus 5.5, GPT-6 Sol/Luna and Grok 4.7 for week 2026-09-21, whose snapshot is September 25). Models removed from a vendor's price list leave the panel from the first week whose snapshot follows the removal (Gemini 3.5/3.7 Flash from week 2026-10-05, retired Qwen3 Max/Coder SKUs from week 2026-10-12, DeepSeek V4.1 Flash 0910 from week 2026-10-19). DeepSeek is priced at its peak (standard) rate; off-peak is a 50% discount.
- Arena name change. From the October 20, 2026 snapshot, LMArena publishes display names (
Claude Opus 5.5 (High)) in place of slugs (claude-opus-5.5-high).scripts/update_arena_snapshots.pymaps them back throughdata/arena/display_name_map.csv(409 rows: exact token matches plus 53 hand-checked pairs) so each model keeps one continuous series; ratings for mapped models move a median 0.05 points across the switch. Models first published under a display name keep it. - Name matching. Arena display names ↔ price catalog IDs via normalization + a checked crosswalk (
data/model_name_crosswalk.csv,data/frontiers/arena_price_name_matches.csv). No invented scores or prices — unmatched models are listed, not guessed. - Score terminology. The source calls the published value an Arena rating/score. Post-January 2024 leaderboard ratings use Bradley–Terry estimation rather than classic Elo; legacy filenames and some rendered labels still say “Elo.”
- Pareto. Per week,
on_frontiermarks non-dominated points in published score ↑ / blended price ↓. - Same-score cost. For each early anchor week’s maximum published score, track the cheapest later blended price among models that meet or exceed that threshold.
- Fixed intelligence (head-to-head bar). For a named past leader, the bar each week is that model’s score on that week’s board, not its frozen launch score. Arena ratings drift as the pool changes — a frozen bar leaves
o1-minifailing its own threshold for 23 weeks — so comparing within one snapshot is the like-for-like test. The frozen-bar result is kept as a cross-check column (o1: ~175× under both bars this week). - Free and promotional prices. $0 rows remain visible in the frontier data and are floored at $0.05 only for log-axis display. The headline same-score calculation requires a positive paid price. Promotions are dated snapshots, not silently carried forward as current rates.
- Chaptered movie. Quiet weeks advance quickly; nine shape-shift weeks hold with callouts (GPT-4 → GPT-4o → o1-mini → Gemini 2.5 Pro → Opus 4.5/4.6 → Aug 2026 cheap tip → Opus 5.5 peak at $8 → Gemini 4 Argon at $4). See
scripts/viz/make_chaptered_movie.py.
| File | Description |
|---|---|
media/coding/frontier_evolution_chaptered.mp4 |
Primary chaptered movie (~65s, 9 chapters) |
media/coding/frontier_evolution_chaptered_last_frame.png |
End card still |
media/coding/frontier_evolution_chaptered_ch7_hold.png |
Chapter 7 hold (Aug 2026 cheap tip, launch promo) |
media/coding/fixed_intelligence.mp4 |
Animated cost of o1-level coding, Jan 2025 → Nov 2026 (~24s) |
media/coding/fixed_intelligence.png |
Same, static, with each record-low model labeled |
media/coding/fixed_intelligence_anchors.png |
Four past leaders side by side (13×–175×) |
media/coding/migration.png |
Early vs latest frontier |
media/coding/price_collapse.png |
Same-score-threshold cost over time |
media/coding/storyboard.png |
One-pager |
media/coding/code_arena_tip.png |
Code Arena (WebDev) tip — includes Astra-scale board |
- Arena score is a relative human-preference ranking, not a fixed absolute IQ score. Pairwise battles, estimation methods, sampling, and leaderboard composition change over time; small week-to-week dips are often noise, not capability loss. It is not a frozen exam and ≠ SWE-bench (or other absolute coding benchmarks).
- Price coverage is sparse early (especially 2023–early 2024). Open / self-hosted models without list API prices are excluded from the price axis.
- New entrants sit on few votes. Gemini 4 Carbon's 1563 (1,072 votes), MiMo v2.6 Pro's 1533 (1,091 votes), Kimi K3.1 Pro's 1477 (774 votes) and Inkling 2 Small's 1457 (616 votes) will move as votes accumulate. Gemini 4 Carbon was priced on November 10, after the latest snapshot, and MiMo v2.6 Pro has no citable list price, so neither is on the price axis. Gemini 4 Argon's and Carbon's $4 is an introductory rate that doubles on December 1, 2026. Each Arena entry is matched to one price ID.
deepseek-v4-flashis matched to DeepSeek's own API, delisted in September, so DeepInfra's $0.09 host price for it (score 1456) is not on the chart. - Match rate is incomplete (~47% of coding Arena names matched to prices overall; lower in 2024). Results describe the matched API-priced subset.
- List prices ≠ discounted enterprise rates; blend assumes a 3:1 input:output mix. Free tiers and promotions are real but unstable, so they are labeled explicitly rather than presented as ordinary durable prices.
cd cheap-intelligence
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Rebuild coding frontiers from weekly CSVs already in data/
python scripts/build_coding_frontiers.py
python scripts/build_fixed_intelligence.py
python scripts/build_export.py
# Story statics + (optional) full animation; needs ffmpeg for mp4
python scripts/viz/make_story_viz.py --skip-anim
python scripts/viz/make_fixed_intelligence.py
python scripts/viz/make_chaptered_movie.pyPre-built tables and media are committed so you can inspect findings without re-running. FFmpeg is required only to regenerate MP4s.
See ROADMAP.md: fixed-cost intelligence, declining-cost update series, size vs intelligence, richer illustrated pauses.
- Code (scripts, notebooks scaffolding): MIT © 2026 Sidney VanNess
- Third-party data:
DATA_ATTRIBUTION.md— LMArena and RoninForge CC BY 4.0; Model Price Watch per upstream terms
data/prices/ dated events + weekly blended prices
data/arena/ weekly Arena scores (coding + overall) + latest WebDev board
data/current/ dated cross-board current-model watch
data/frontiers/ Pareto panels + same-score-threshold cost series
data/export/coding/ xlsx + frontier_only CSVs for charts
scripts/ rebuild + viz
media/coding/ chaptered mp4 + key statics
