What changed.

Release tracker: record the rerun HF sweep and the org list in the protocol

Regenerate the changelog

Release tracker: rowcheck sweeps HF repos reopened in the window

Regenerate the changelog

Release tracker: 9 November journal records the repeat sweep that closes the 08:52:41–08:53:20Z gap (0 new rows)

Regenerate the changelog

Release tracker: 9 November journal completed; PROTOCOL §6 states how rowcheck treats HF errors and Wayback

Regenerate the changelog

Release protocol: rowcheck.py checks every grok row against the maker's page (§6), with fixtures; snowflake.py exposes when() for it

Release tracker: 9 November rumor pass; Frantic Ferret added, six cards noted, seal 5–11 Nov

Regenerate the changelog

Release protocol: grok signs in through grok.com OAuth with the xAI key variable unset; xAI handle @SpaceXSI

Release tracker: 8 November rumor pass; Qoduck added, ERNIE 5.5 Flash and Hy4 weights raised to high, seal 5–11 Nov

Regenerate the changelog

Jeep: g1 by Qoduck (Qwen4 Flash/Plus/Max?) (xhigh) (owner-supplied, 7 November)

Regenerate the changelog

Release tracker: 7 November rumor pass; 8 releases added, Muse Realtime Avatar shipped, Inkling 2 Small weights and ERNIE 5.5 Flash added as cards, seal 5–11 Nov

Regenerate the changelog

EvalMapBench: Mercury 3 (high), 77.0% [72.6, 81.3], strict 41 (OpenRouter, Inception)

Regenerate the changelog

Jeep: Mercury 3 g2 detail credits the effort to the owner, not the session log

Regenerate the changelog

Jeep: g2 by Mercury 3 (high) (owner-supplied, 6 November)

Regenerate the changelog

Jeep: g1 by Devin SWE-2 Flash (max) (owner-supplied, 6 November)

Regenerate the changelog

EvalMapBench: Devin SWE-2 Flash (max) at 45 minutes, 83.7% [80.0, 87.5], strict 60 (9 re-runs)

Regenerate the changelog

EvalMapBench: Devin SWE-2 Flash (max), 129/129 on 6.7 in a bubblewrap sandbox, 80.9% [76.2, 85.5]

EvalMapBench audit: a partial point's SVG label and key number are the unranked "—", its key entry reads "(partial, unranked)"

Regenerate the changelog

Release tracker: label the Hy4 Flash card ‘Hy4 Flash announced’, like the other announced cards

Regenerate the changelog

Release tracker: 5 November rumor pass; 7 releases added, Mistral Large 4 weights, Kimi K3.1 Pro weights and GPT Image 2.6 shipped, Hy4 weights and Hy4 Flash added as cards, close week 29 Oct–4 Nov, seal 5–11 Nov

Regenerate the changelog

Atlas: AA snapshot of 4 November 2026 (Intelligence Index v5, 88 configurations)

Regenerate the changelog

EvalMapBench: Hy4 (high), page-only row (129/129) through OpenRouter's Tencent endpoint

Regenerate the changelog

Releases: Hy4 weights still not out at 08:14 UTC on 4 November (Hugging Face, ModelScope, GitHub checked)

Regenerate the changelog

Releases: Hy4 (4 Nov), the general release of Hy4 preview, with the cutoff moved to 4 November for that row only

Regenerate the changelog

Releases: LongCat-2.5 (2 Nov), the general release of LongCat-2.5-Preview, with the cutoff moved to 3 November for that row only

Regenerate the changelog

Jeep: game31 by LongCat 2.5 (max) (owner-supplied, 3 November)

Regenerate the changelog

Add LongCat 2.5 to EvalMapBench

Regenerate the changelog

Release tracker: the 29 October second-pass journal says the open week was reprinted for the GPT Image 2.6 card and notes where the §0 grep list was kept

Regenerate the changelog

Release tracker: 29 October second pass; no new releases, GPT Image 2.6 added as an image card, notes added to DeepSeek V4.1 Pro, OpenAI's next model and Mistral Large 4 weights

Regenerate the changelog

Odds: a settled question keeps its pointer after its card is counted and removed; odds.json as on 28 October

Regenerate the changelog

Release tracker: the 29 October journal says leftOut does name Granite Speech 5.0 2B and Raon-1-mini-preview

Regenerate the changelog

Release tracker: 29 October pass; close week 22–28 Oct, seal 29 Oct–4 Nov

Regenerate the changelog

Release tracker: 28 October rumor pass; 7 releases added, Genie 4 added as a later card, Gemini 4 Carbon announced, week 22–28 Oct rechecked

Regenerate the changelog

Jeep: the GPT-6.1 Terra card names game1 by its archive and project, not a folder

Regenerate the changelog

Jeep: game1 by GPT-6.1 Terra (max) (owner-supplied, 27 October)

Regenerate the changelog

Release tracker: the 27 October journal keeps Fern 0.1 on the watch list and names what stayed out of leftOut

Regenerate the changelog

Release tracker: the 27 October note and journal count Mirendil Fern 0.1 as the pass's new card

Regenerate the changelog

Release tracker: Mirendil Fern 0.1 announced 26 Oct, added as a later card

Regenerate the changelog

Release tracker: 27 October rumor pass; GPT-6.1 Terra and Ling-3.1-flash-1019 shipped, 4 releases added, week 22–28 Oct rechecked

Regenerate the changelog

EvalMapBench: GPT-6.1 Terra (max), page-only row (129/129) with its 45-minute re-runs

Regenerate the changelog

EvalMapBench: GPT-6.1 Terra (high), page-only row (129/129)

Regenerate the changelog

Release tracker: record the 25 October pass's deploy version and recheck the evening of 24 October

Regenerate the changelog

Release tracker: 25 October rumor pass; Nemotron 3.5 Bolt E4B and Kimi K3.1 Pro shipped, 9 releases added, week 22–28 Oct rechecked

Regenerate the changelog

EvalMapBench: K3.1 Pro note says the default was high by 24 October

Regenerate the changelog

EvalMapBench: note that Kimi K3.1 Pro ran at max while Kimi's documented default moved to high

Regenerate the changelog

EvalMapBench: Kimi K3.1 Pro (max), page-only row (129/129)

Regenerate the changelog

Jeep: game23 by Kimi K3.1 Pro (max) (owner-supplied, 23 October)

Regenerate the changelog

Release tracker: 23 October rumor pass; MiniMax M3.1 Flash weights shipped, week 22–28 Oct rechecked

Release tracker: 21 and 22 October rumor passes; close week 15–21 Oct, seal 22–28 Oct

EvalMapBench: Ornith 1.5 122B (high), page-only row (129/129)

Regenerate the changelog

EvalMapBench: MiniMax M3.1 Flash, page-only row (129/129)

Regenerate the changelog

Jeep: game19 second build by Kimi K2.7 Code (owner-supplied, 19 October)

Regenerate the changelog

Jeep: game17 by MiniMax M3.1 Flash (owner-supplied, 19 October)

Regenerate the changelog

Atlas: DeepSeek V4.1 Flash 1017 point and jeep link

Regenerate the changelog

Release tracker: 19 October rumor pass (with pending 16-18 October updates)

EvalMapBench: DeepSeek V4.1 Flash 1017 (max) through Terminus-2 on DeepSeek's API

EvalMapBench: Pareto 26.10 through Terminus-2 on Unbiased's API

EvalMapBench: Nemotron 3.5 Super VL through Terminus-2 on paid OpenRouter

EvalMapBench: Ultramarine 1 Flash (max) through Terminus-2 on Paper Instruments' API

Jeep: Nemotron 3.5 Super VL, owner-supplied game9, failed (crashes on the first frame); rebuilt site

Releases 15 Oct: Nemotron 3.5 Bolt 74B (card added, then released) and Nemotron 3.5 Super VL

Snapshots of 15 October: AA through the Data API, fresh open feeds; method-text counts and the Pareto front follow the data

Releases: 15 Oct rumors pass — Step 5 Preview weights, Ling-3.1-flash weights, Grok Voice 3.0; forecast week 15–21 Oct

EvalMapBench: note that Gemini 4 Argon (max) barely used its 1M output budget

EvalMapBench: Gemini 4 Argon (max) through cli2027

Jeep: Gemini 4 Argon (max, cli2027), owner-supplied game6

validate: every jeep shotNote must reach the home banner or be listed in OFF_BANNER

Home banner: show owner-supplied jeep builds with race/countdown frames

Jeep: Gemini 4 Argon (max, 1M output), owner-supplied game5

Odds: 14 Oct (no new maker dates; Muse Spark next and Grok 4.8 follow markets down)

Ledger record: score only the cards named for the week; a further-out card that shipped is said apart

Ledger: close the week of 4-10 Oct (1 of 4 named shipped: Claude Haiku 5.5)

EvalMapBench 6.7: all 43 tasks ranked

Releases: odds table links only to cards still shown

Home banner: EvalMapBench ranks every full row, priced or not

Strip: shipped cards (Argon, Watermelon) leave the expected side

AA snapshot 13 Oct 2026: Muse Wonder in; Ling-3.0-flash-VL kept without AA price and speed

Jeep: Devin SWE-2 (max), owner-supplied game4

Releases: pass 2026-10-13 (Muse Wonder, Gemini 4 Argon, MiniCPM-V 4.7, LFM2.5-24B-A2B)

Jeep: Gemini 4 Argon tile captured natively on the RTX 5090

Jeep: Gemini 4 Argon (high), owner-supplied game3

EvalMapBench 6.6: Devin SWE-2 (max) complete, ranked

EvalMapBench point key: Muse Wonder Full keeps #5, no per-token price so no point

EvalMapBench: rank full rows without a price (Muse Wonder Full)

EvalMapBench 6.6: Muse Wonder Full (max) complete, ranked

EvalMapBench 6.6: all 40 tasks ranked

EvalMapBench matrix: a column with the number of tasks each row solved

EvalMapBench matrix: rows sorted by tasks solved, most first

Regenerate the changelog

EvalMapBench: Claude Fable 5.1 (high) through Claude Code, on all forty tasks

EvalMapBench matrix: every row on the four new tasks (Opus 5.5 xhigh, Gemini 3.8 Flash and Kimi's re-run last)

EvalMapBench matrix: Sonnet 5.5 xhigh and Qwen3.8 Max Prime on the four new tasks

EvalMapBench matrix: Haiku 5.5, GLM-5.3-Prime and Qwen3.8 27B on the four new tasks

EvalMapBench matrix: the four new tasks on most rows so far; a dot marks cells run through Claude Code or Codex

Releases 12 Oct: cutoff and forecast check to 12 October; Watermelon market back to a 15 Oct median

Releases 11 Oct: Pixel Canary is Callstack's Apex; Qwen teaser, MiniCPM-V 4.7 loader and Watermelon market on their cards

Home: What's new expanded on load; audit expects it open

EvalMapBench matrix: models as rows, tasks as columns

EvalMapBench matrix: compact squares (green solved, red not solved, grey not run)

EvalMapBench: per-task matrix of every model on every task

Home UI: compact first screen, collapsible banners, readable EvalMapBench, two-model comparison

Regenerate the changelog

EvalMapBench chart: plot output tokens per trial instead of minutes

Regenerate the changelog

EvalMapBench star: price counts half as much as time and detail share

Regenerate the changelog

EvalMapBench 6.5: Muse Spark 1.3 Contributor at effort max joins

EvalMapBench star: leave the 0% baseline out of the frontier and its scales

EvalMapBench: star the frontier model nearest the cheap, fast, highest-share corner; partial rows out of the frontier

EvalMapBench 3D chart: star only the best model on the Pareto frontier (highest detail share)

EvalMapBench 3D chart: numbered circles replace dots and callout boxes; Pareto star in key and table

EvalMapBench 3D chart: perspective with drop lines and floor shadows instead of the isometric view; flat Price/Time presets unchanged

Regenerate the changelog

EvalMapBench 6.5: ten new tasks, every row re-measured on thirty-six

EvalMapBench: the detail chart as a turnable 3D cube of share, price and time; Price/Time become camera presets

EvalMapBench: Price/Time toggle on the chart, median minutes per trial on a log axis

EvalMapBench 6.5: ten new tasks, every row re-measured on thirty-six

EvalMapBench 6.5: add 'The middle is a tie' limitation

Regenerate the changelog

Generate SEO/LLM metadata: Dataset, ItemList and FAQ JSON-LD, llms-full.txt, markdown pages

EvalMapBench 6.4: three new tasks, Qwen3.8 Max Prime high, Qwen3.8 27B on Cerebras (#13)

Snapshots of 10 October: AA 196 configurations, fresh open feeds, odds day, releases through 10 Oct

EvalMapBench: pointing at a key row lights its point on the chart

EvalMapBench 6.3: Gemini 3.8 Flash at 61 of 69 trials

EvalMapBench 6.3: Grok 4.7 high; Gemini 3.8 Flash at 56 of 69 trials

EvalMapBench: chart axes fit the data, points coloured by maker with a legend

EvalMapBench 6.3: Qwen3.8 Max Prime at 45 minutes; Gemini 3.8 Flash at 52 of 69 trials

EvalMapBench 6.3: Grok 4.7 xhigh; Gemini 3.8 Flash at 47 of 69 trials

EvalMapBench 6.3: Step 5 Preview at 45 minutes; Gemini 3.8 Flash at 32 of 69 trials

EvalMapBench 6.3: Gemini 3.8 Flash, published while its run is still going

EvalMapBench: results table fits the page width without scrolling (#12)

EvalMapBench: ship table sorting as sort.js, the CSP blocks inline scripts (#11)

EvalMapBench 6.3: scored beyond a gpt-oss-120b baseline

EvalMapBench: tidier tier table, click-to-sort headers on both tables (#10)

EvalMapBench 6.2: 45-minute figures for Sonnet 5.5 xhigh and high and Haiku 5.5 xhigh

EvalMapBench 6.2: Step 5 Preview joins, first 45-minute re-runs land

EvalMapBench: tidier results table (notes under numbers, right-aligned headers)

EvalMapBench 6.2: Claude Mythos 5.1 joins, tokens and 45-minute columns

EvalMapBench 6.1: three planning tasks join the twenty, every row re-measured on them

Read the EvalMapBench version and task count from the export

EvalMapBench 6.0: Kimi K3 at max on Moonshot's own endpoint

EvalMapBench 6.0: GLM-5.3-FlashX

EvalMapBench 6.0: GLM-5.3-Prime

EvalMapBench 6.0: GPT-6.1 Sol xhigh on OpenAI's ultrafast tier through Codex

EvalMapBench 6.0: GPT-6.1 Sol high on OpenAI's ultrafast tier through Codex

AA snapshot of 8 October: 187 configurations, Claude Haiku 5.5, and CritPt as a metric

EvalMapBench 6.0: DeepSeek V4.1 Flash at max, the default run becomes its own high row

Odds: the 8 October reading, eleven open questions to 31 Oct, 30 Nov and 31 Dec

Ledger: close the seven weeks that ended by 7 October

BridgeBench: re-read on 8 Oct with a new Onion key; 27 models, values unchanged

Method text: the text arena reaches 71 configurations, not 68

LMArena: full refresh of all ten feeds (8 Oct)

EvalMapBench 6.0: Claude Haiku 5.5 at xhigh, on the bench page until the atlas has it

EvalMapBench 6.0: DeepSeek V4.1 Flash ran at the API default, which is the high level

EvalMapBench 6.0: eleven rows, Anthropic and OpenAI efforts side by side

Publish game22: GLM-5.3-Prime Jeep build, owner-supplied and unscored

Vals: read the boards from the new page layout, refresh all six

8 October reading: forecast 8–14 Oct, open sources refreshed

Banner: admit builds whose frame is from the RTX 5090 measured drive

game21 card: say where the rainbow ground comes from

EvalMapBench page: a compact design

Score game21 Step 5 Preview (high): 28/100, rubric v4, drives

Publish game21: Step 5 Preview (high) Jeep build, owner-supplied and unscored

Publish Claude Haiku 5.5 Jeep build and release, GPT-6 October in ChatGPT

EvalMapBench 6.0: allow rows the atlas has no entry for yet

Bench-page audit: admit the disclosed Cloudflare page counter, as the main audit does

Publish EvalMapBench 6.0 and retire the earlier tracks

Publish Mythos 5.1 and Fable 5.1 xhigh Jeep builds and ribbon update

Publish reviewed Bench UI, Mistral Large 4 and Nano Banana 2.1

feat: publish EvalMapBench aggregates and dated AA agent steps

Releases: Banbury Road's Kardashev-0.7, announced 5 Oct, API beta waitlist, no date

Releases: Reflection AI's Beam, announced 5 Oct, weights promised later in October

Audit: expect the 6 Oct API snapshot (181 rows, SWE for all, 41 on the frontier); unplotted-model check on GPQA

Take the 6 October AA snapshot from AA's official Data API (181 rows, SWE domain for all); steps for 29 of 31 agents; attribution and dates (deployed as e2200f6f)

Jeep: publish Kolibri 1 (oc2) as a failed build, metadata only (deployed 5 Oct as d1f3f301)

Re-read BridgeBench on 5 Oct (data unchanged); rebuild; dates

Header: build keeps the agent steps date beside the AA snapshot date

Y axis takes any metric; add AA Coding Agents mean steps (18 rows, read 5 Oct 2026); update dates; carry the uncommitted 4-5 Oct working tree (releases, EvalMapBench v5, jeep, rumors)

EvalMapBench v5: refresh Fable and GLM, publish partial Flash and Qwen Prime

Jeep: publish Saltwind Rally from Grok 4.7 Fast xhigh

Jeep bench: grok47-xhigh-oc2 failed the brief (main.js does not parse)

Jeep bench: fledge-alpha-oc2 failed the brief (three.tsl.js 404); declare free Zen models in the harness

Jeep bench: fledge-alpha-oc2 (unscored: three/tsl 404), bench fixes for opencode 2.0.14

Regenerate the changelog

Fetch BridgeBench through the Onion account API

Regenerate the changelog

Bring the hand-written counts in line with the 2 October snapshot

Regenerate the changelog

Refresh the open sources and Artificial Analysis to 2 October

Regenerate the changelog

Redesign the releases page: an editorial hero and a timeline of what shipped

Regenerate the changelog

Deploy to Cloudflare after a green check on main

List Ideogram 4.5 and FLUX 3 Image on the releases page

Regenerate the changelog

Leave merge commits out of the changelog

Regenerate the changelog

Tell search engines what each page is: titles, structured data, a fetchable icon and llms.txt

Redesign the atlas and the jeep gallery, add a dark theme, and take the 1 October snapshots

Regenerate the changelog

Add Ling-3.1-flash's jeep build at 37, run on the bench at its highest effort

Regenerate the changelog

Add GPT-6.1 Sol to EvalMapBench: 142 of 147 tasks, run as GPT-6 Sol was

Regenerate the changelog

Take the 30 September Artificial Analysis snapshot: GPT-6.1 Sol on the atlas, with its blitz and knowledge

Regenerate the changelog

Add the hard blitz (v5.1) and take the 30 September blitz export: Sonnet 5 out by age

Regenerate the changelog

Add GPT-6.1 Sol's jeep build at 45, made in the owner's own Codex the day it came out

Regenerate the changelog

Add GPT-6.1 Sol, released at DevDay on 29 September, which settles the DevDay question

Regenerate the changelog

Bring the releases to 29 September: Kling 4.0 Flash, Kling 4.0 announced, OpenAI drops GPT-6.1 Astra

Regenerate the changelog

Refresh AA and BridgeBench, read Sonnet 5.5's SWE index off AA's chart, and say on the chart when a picked model is not on it

Regenerate the changelog

Open each build straight from its thumbnail on the atlas banner

Regenerate the changelog

Take the 29 September blitz export: Sonnet 5.5 in progress, seven rows on the atlas, five older ones out

Regenerate the changelog

Add Claude Sonnet 5.5: its release, five configurations from AA, its jeep build at 48, its knowledge

Regenerate the changelog

Add the evening's releases: NaiveAI's Naive-N0.5-Flash and ElevenLabs' Eleven v4

Regenerate the changelog

Bring the releases to 28 September: MiMo-V2.6 MOPD weights, Grok 4.7 from 20 Sep, OpenAI's pause

Regenerate the changelog

Add Pixel Canary's jeep build: Badlands Rally, 12/100, breaks at the rim

Regenerate the changelog

Put the EvalMapBench blitz on the atlas: 146 small pieces, five rows with 90% of a verdict

Run Gemini 3.8 Flash on an eighth harness, Google's Antigravity CLI: 12, breaks

Bring the releases to 27 September: MiniMax-M3.1-Flash-Preview, and a strip that keeps its days

Rerun the three builds NanoGPT's limit and a stalled provider cut off: one drives

Publish the 27 September reruns: thirteen builds, two that drive

Regenerate the changelog

Re-export EvalMapBench: Grok 4.7 and Kimi K3 join, on 60 and 52 of the 150 tasks

Turn the BridgeBench button to one shot against overall, as the owner asked

Note OpenCode Go's data terms and request limits, and the first MiniMax Code run's use

Publish MiniMax-M3.1-Flash-Preview's jeep build from MiniMax Code, held to 12

Add BridgeBench as a source, and turn the chart with three large buttons

Add a seventh jeep bench harness, MiniMax Code

Let the jeep box's models test in a browser and look up package files

Regenerate the changelog

Give EvalMapBench a banner of its own, a preset and a way into the 3D chart

Regenerate the changelog

Link the EvalMapBench page from everywhere a reader looks

Put the jeep box's Chromium on the PATH as chromium

Regenerate the changelog

Take the 27 September Artificial Analysis snapshot: 160 new speeds, two now EvalMap's

Regenerate the changelog

Re-measure EvalMapBench v3 after fixing six tasks and three grading faults

Add Space Bunny Alpha's second jeep run, through OpenCode Zen: it fails at load

Add a sixth jeep bench harness, Freebuff, and its first eight builds

Regenerate the changelog

Add EvalMapBench: private tasks from one developer's own work, and a page of solved against price

Regenerate the changelog

Link LongCat-2.5-Preview's release row to its jeep build

Regenerate the changelog

Add LongCat-2.5-Preview to the jeep bench: 32, on the day OpenCode made it free

Regenerate the changelog

Score the jeep bench on rubric v4, and add five builds and three failed cards

Regenerate the changelog

Bring the releases to 26 September: Pixel Canary, LongCat-2.5-Preview, Perceptron Mk1.5

Regenerate the changelog

Take the odds for 26 September: part of the way toward the market on Sonnet 5.5

Regenerate the changelog

Add GPT-5.6 Luna and Qwen3.7 Plus, and failed cards for Inkling Small and Gemini 3.5 Flash

Regenerate the changelog

Add Claude Opus 5, Gemini 3.8 Flash and GPT-5.6 Terra from the Command Code bench

Regenerate the changelog

Put the odds on the atlas, under the jeep banner

Note what Command Code charges and which models refuse --effort max

Regenerate the changelog

Add four more Command Code jeep builds: GPT-5.6 Sol, Qwen3.7 Max, Inkling, Gemini 3.6 Flash

Regenerate the changelog

Add Claude Sonnet 5 from the Command Code bench, and stop counting its cache twice

Regenerate the changelog for Qwen3.5 397B's failed jeep run

Run Qwen3.5 397B A17B on the jeep bench: it failed the brief, a HUD over a white screen

Regenerate the changelog for three more failed jeep runs through NanoGPT

Run DeepSeek V4 Flash 0731, Ling-3.0-flash and Kimi K2.6 on the jeep bench: all three failed

Regenerate the changelog for Hy3's failed jeep run

Run Hy3 on the jeep bench: it failed the brief, "pow is not defined"

Regenerate the changelog for MiMo-V2.5-Pro's failed jeep run

Run MiMo-V2.5-Pro on the jeep bench: it failed the brief, stuck on its loading screen

Let the banner check skip a build whose frame carries a note, as the banner does

Regenerate the changelog for Kimi K3's jeep run

Run Kimi K3 on the jeep bench through Command Code: 33, scored on its WebGL2 fallback

Price Command Code runs at list price, and let the capture tools hide WebGPU

Regenerate the changelog for three failed jeep runs through NanoGPT

Run MiniMax-M3, GLM-5.2 and Kimi K2.7 Code on the jeep bench: all three failed the brief

Regenerate the changelog for DeepSeek V4 Pro 0813's failed jeep run

Run DeepSeek V4 Pro 0813 on the jeep bench: it failed the brief, nothing is ever drawn

Add a Command Code runner to the jeep bench, for the models the owner's subscription carries

Add Antigravity and grok CLIs to the ledger, and Go's DeepSeek V4 Pro now that Global is on

Note in the ledger that OpenCode Go works again, and that DeepSeek V4 Pro there wants the Global region

Start a ledger of subscriptions and CLIs: what each has left and which to run a model on

Let the jeep bench run models through NanoGPT, and take logins OpenCode 2 keeps in its database

Let jeepshots.js press a start button that says "race", as jeepdrive and jeeppilot already do

Regenerate the changelog for GLM-5.3-Prime's failed jeep run

Run GLM-5.3-Prime on the jeep bench: it failed the brief, only the HUD is drawn

Refresh OpenCode's catalogue from models.dev before every jeep bench run

Describe the AA fetch as it is in the README and the helper scripts, not as a hand save

Say in aa_extract.py how the AA page is actually fetched: one curl request since 21 September

Regenerate the changelog for the 25 September snapshot and releases

Bring the releases to 25 September: five new rows, the week of 25 Sep - 1 Oct, and today's odds

Take the 25 September Artificial Analysis snapshot: three rows in, six out, 155 new speeds

Regenerate the changelog for the MiMo UltraSpeed speed run

Time MiMo-V2.6-Pro-UltraSpeed: 344 tokens/s against 26 for standard Pro, noted on its card

Regenerate the changelog for the Qwen3.8-Max-Prime speed run

Time Qwen3.8-Max-Prime: 109 tokens/s against 49 for 0902, on one client and endpoint

Regenerate the changelog for the Qwen3.8-Max-Prime note on the atlas

Note Qwen3.8-Max-Prime on Qwen3.8 Max (0902) in the atlas, not as a point of its own

Regenerate the changelog for Qwen3.8-Max-Prime's release row

Add Qwen3.8-Max-Prime to the releases: Qwen3.8-Max in Alibaba's fast mode

Regenerate the changelog for Qwen3.8 Max Prime's jeep build

Add Qwen3.8 Max Prime's jeep build: Canyon Ridge Rally, 43/100

Regenerate the changelog for the image test

Add a ten-model image test to the small tests: exact text in three scripts, clocks, hands, a full glass

Regenerate the changelog for the English role-play run

Rerun the Aion 3.5 role-play in English, and show the Russian run beneath it

Regenerate the changelog for the small tests page

Publish two small tests run here: Aion 3.5 on a long role-play, Recraft V4.1 Flash on text

Regenerate the changelog for Ling-3.0-flash-Fin's failed jeep run

Run Ling-3.0-flash-Fin on the jeep bench: it wrote no file

Regenerate the changelog for the Ling-3.0-flash-VL build

Run Ling-3.0-flash-VL on the jeep bench: 22, a mountain lake without a jeep in view

Regenerate the changelog for the DeepSeek-V4-Flash-Vision-Exp build

Run DeepSeek-V4-Flash-Vision-Exp on the jeep bench: 37, on ground that reads as water

Regenerate the changelog for Granite 4.2 8B's failed jeep run

Run Granite 4.2 8B on the jeep bench: it failed the brief, and its card says why

Regenerate the changelog for Mercury 2.5's failed jeep run

Run Mercury 2.5 on the jeep bench: it failed the brief in 18 seconds, and its card says why

Regenerate the changelog for Qwen3.8-Flash-Next's jeep build

Run Qwen3.8-Flash-Next on the jeep bench and publish its build: 39, drives

Regenerate the changelog for Solar Mini 4's failed jeep run

Run Solar Mini 4 on the jeep bench: it failed the brief, and its card says why

Regenerate the changelog for the failed jeep mark

Show a failed jeep build red and struck through on the release line and list

Regenerate the changelog for the odds

Put a number on the forecast: the odds by the next three month-ends, beside the markets

Regenerate the changelog for the failed-build cards

Show builds that failed the brief: AliceAI-Foundation gets a struck-through jeep and the reason

Regenerate the changelog for the Meta recheck

Recheck Meta after the Connect keynote: no model shipped, Muse Realtime Avatar announced

Add EvalMapBench: thirteen models on one developer's own work, measured here

Regenerate the changelog for the single pick line

Keep one line on the 3D chart, from the previous pick to the current one

Regenerate the changelog for the selection trail

Join the models picked on the 3D chart with one trail, in the order picked

Regenerate the changelog for Space Bunny Alpha's jeep build

Publish Space Bunny Alpha's jeep build from the OpenCode bench, run twice

Regenerate the changelog for wheel zoom

Zoom the 3D chart with the mouse wheel, within the bounds of the + and − buttons

Regenerate the changelog for Space Bunny Alpha

Add Space Bunny Alpha, a stealth model that appeared on OpenRouter and OpenCode on 23 September

Regenerate the changelog for the 23 September release update

Update the release list and forecast to 23 September

Regenerate the changelog, which had fallen behind the history

Take the 23 September Artificial Analysis snapshot: new speeds, and 16 rows AA now marks deprecated

Retake five jeep tiles with the pilot on the course, not the jeep off it

Score Claude Code and Codex builds with run.py, not a rubric v1 of their own

Steer the jeep pilot along the course and re-judge the builds it could follow

Run every jeep-bench model in a container with no way out but its API and npm

Hold a jeep build that cannot be driven to 20 or 12, after a minute of steering it on the graphics card

Score every jeep build on rubric v2 after a measured drive on a graphics card, and retake the gallery there

Take Artificial Analysis's speed for all twelve GPT-6 Sol and GPT-6 Luna configurations

Regenerate the changelog, which had fallen eight changes behind the history

Open the README with a compact index of its eleven sections

Make the jeep rubric v2: score what plays on screen, and put the measured frame rate on the card

Take Artificial Analysis's speed for Claude Opus 5.5 at xhigh, high, medium and low

Put GPT-6 Sol and GPT-6 Luna on the chart, all twelve configurations from Artificial Analysis

Put Claude Opus 5.5 on the chart, and publish the jeep builds of Opus 5.5, GPT-6 Sol and GPT-6 Luna

List GPT-6 Sol and GPT-6 Luna, out on 22 September, and take them off the week's forecast

Move Claude Opus 5.5 from the forecast to the releases, out on 22 September

Put the scoring commit into the changelog it left out

Score every build in the gallery, not only the ones the bench ran

Score every bench build from 0 to 100, where 100 is a game on the level of GTA VI

Run the jeep brief ourselves on an OpenCode bench, starting with GLM-5.3 at max effort

Let a build's page say the language its words are actually in

Keep the two presets that name a job, and publish the build MiMo-V2.6-Pro wrote

Stop the checks running on a version of Node the runner is about to withdraw

Give every page the audit opens a timeout a machine without a graphics card can meet

Keep the forecast to its word, and publish every change this site has been through

Date every source by its own download, and stop a stale checkpoint setting a price

Mark DeepSeek V4.1-Pro, Watermelon and Grok 4.8 as announced, and keep each announcement visible at any width

Add MiMo-V2.6's UltraSpeed mode and 9B distillation, and Apsara's image and speech announcements

Put OpenAI's launch on Tuesday, and add three models due later from a rumour check

Keep an announcement on the strip until the model ships

Put Qwen4 on the release strip on the day it was announced

Correct the Apsara note after a grok check, and add two Qwen speech models it turned up

Put MiMo-V2.6-Pro on the chart, list the MiMo-V2.6 release, and record Qwen4 as announced

Take Grok 4.7's speed and price from Artificial Analysis, and retire the stand-ins

Move each jeep build down under its strip, and add GLM-5.3-FlashX and Fugu Max

Count Cloudflare's own beacon apart in the strip check

Add Vals AI as a source, and give every jeep build a strip of its own

Take Grok 4.7 onto the chart with its own speed, price and SWE index

Rework the LMArena join and add five arena boards

Count the WebDev arena at 49 on the licences page too

Give Qwen3.8 Max (0902) its own arena entry, not the older checkpoint's

Put a model released on the snapshot's day on the Released line

Show the jeep builds only here, and give the model out today a full chip

Add Grok 4.7's jeep race, published only here

Write down how to gather releases and rumours with the grok CLI

Mark each release's place on the chart and its jeep build, and add Grok 4.7

Set the releases against today, lay out the week, and ring new models on the chart

List the month's model releases and this week's forecast on the site

Keep the month's model timeline and this week's forecast

Call two builds by the names their makers gave the models

Open the gallery with an index of every build

Add the five jeep builds that were published as Workers

Make the audits count the injected counter, as the last commit said

Say that the zone's counter reaches the game pages too

Bump the licences page to the same stylesheet token

Take the frontier table off the page

Host every jeep build here, in English, on a timeline that reads downward

Run the ribbon newest first, and sign the banner thumbnails

Divide the jeep rail by week, and say it is an order

Show what 21 models made of the same jeep brief

Add nine metrics and two price sources, all openly licensed

Say that Cloudflare's edge adds a page counter, because it does

State plainly what the site takes and collects: nothing

Rebuild the page around the chart, on Fable's design review

Drop the 2D charts and leave the cube as the only view

Darken the markers above the price limit and publish a contact address

Put the green reference plane back and let it move along axis Z

Swap the X and Z metrics so the benchmark reads vertically

Put output speed back on an axis and price back in the cloud

State precisely which Artificial Analysis terms apply and why none of them cover this site

Label every value with its source and let the visitor choose it

Publish complete EvalMap source and Cloudflare deployment configuration

Initialize EvalMap website repository