Skip to content

About

Coding IQ

Coding IQ compares chat models on published coding benchmarks. On this scale, 100 was the average for models with enough coding results on 23 Sep 2026.

We calculate it from the published coding benchmarks we collect, and it currently ranks 303 models on 215 boards across 60 benchmark families from 33 sources. Each score uses the model's best published result, usually at its highest reasoning setting.

It has its own scale, not the llminfo IQ scale. 15 points is one standard deviation of the group that set the average. Today's frontier lands around 120 to 140. The scale is fixed to that group, so a model's number stays put when new models or new boards arrive. It only moves when its own evidence changes.

Last calculated 22 Nov 2026. Every score is shown with this date. A model's score can change as new coding results are published.

Current frontier: about 120–140557085100115130145160100 average on 23 Sep 2026
The ruler shows 55 to 160. That is the range drawn, not a limit of the score. Each tick is 15 points, one standard deviation of the group that set the average.

How to read it

  • 140 ±10 — the number and its interval. The ± is half of a 95% interval built from how much the model's own boards disagree. Two models whose intervals overlap are not separated by this index. Where a table shows the score without a ±, the interval is on the model's own page.
  • ≈ 112 — an estimate with little data behind it. Three or four benchmark families is enough to rank but not to be precise. So the score is muted, rounded and held inside the range of models with more data behind them.
  • No score — not enough independent coding evidence. A model needs at least 3 benchmark families from at least2 independent sources. Preview releases and maximum-compute tiers often fall short because nobody has run the agentic boards on them. Absence is never a verdict, and the general llminfo IQ is never shown in its place.
  • Stale — evidence over a year old. Coding evidence ages in months. A score resting only on results more than a year old is marked, not hidden.

What counts as evidence

Three kinds of published result are fitted together as one factor. A benchmark family counts once per source whoever ran it and under whatever version, and rows that measure cost, speed or token use never enter.

  • Code generation. Function, competition and snippet tasks scored by running tests.

    ale-bench · BigCodeBench · Bird-SQL · Codeforces · HumanEval · ioi · LiveBench Coding · LiveCodeBench · LiveCodeBench Pro · MBPP · multipl-e · natural2code · SciCode · usaco

  • Software engineering. Agents working on real repositories and terminals, always model plus scaffold.

    Aider Polyglot · algotune · ArtifactsBench · Claw · code migration · Cognition (vendor harness) · Cursor (vendor harness) · DeepSWE · frontierswe · gso · internal research debugging evaluation · kernelgen 1p · Kilo code (vendor harness) · mirrorcode · nl2repo · nl2repo-bench · Openhands (vendor harness) · PinchBench · ProgramBench · qwenqoderbench · qwenreactbench · Roomote (vendor harness) · skillsbench · skillsbench · self-contained subset · skillsbench · text-only · skillsbench · text-only · openhands · skillsbench · vals · with skills · accuracy · skillsbench · vals · with skills · mean duration / test · skillsbench · vals · without skills · accuracy · skillsbench · vals · without skills · mean duration / test · skillsbench 1.1 · with skills · skillsbench 1.1 · without skills · swe atlas · swe-atlas · codebase q&a · SWE-bench · SWE-bench Multilingual · SWE-Bench Pro · SWE-fficiency · swe-lancer · swe-marathon · SWE-rebench · Terminal-Bench · Vibe Code · vision2web

  • Human preference. Head-to-head votes on code answers in public arenas.

    Arena Coding · WebDev Arena

Re-fitting without the vendor-harness boards keeps the ranking (rank correlation 0.999, 10 of the top ten retained, median shift 0.4 points across 25 boards removed).

Kept out of the fit and used only as cross-checks: Artificial Analysis Coding Index (relayed) — convergent check only; Artificial Analysis SciCode / LiveCodeBench / Terminal-Bench Hard (relayed constituents); Epoch Capabilities Index; OpenHands Index (composite).

What it is not

  • Not on the llminfo IQ scale. The models people run coding benchmarks on are a stronger group than the models behind llminfo IQ, so the same frontier model reads lower here than on IQ. That is a property of the two populations, not a weakness at coding. Subtracting one from the other says nothing calibrated.
  • Not a measurement of any single benchmark. It combines results from many tests. A model can hold a high Coding IQ and still lose a specific board to a lower-ranked one. It says nothing about your codebase.
  • Not the default configuration. Every score uses the model's best published result. At the top, that usually means high or maximum reasoning effort and, on agent boards, a specific harness.
  • Not harness-free. Agentic coding scores always measure the model plus its scaffold. Boards run inside a vendor's own product count as one vote per vendor, and the index is re-fitted without them on every refresh to check they do not drive the ranking.
  • Not independent of lab reporting. Numbers a lab publishes about its own model count as one vote however many sites relay them. The two-source floor keeps a model from being ranked on its own report card alone.
  • Not immune to age. Older models were mostly measured on older, easier boards. Scores resting on evidence over a year old are marked stale rather than hidden. The bottom of the ranking is the least trustworthy part.
  • Not human IQ. The name borrows the 100 and 15-point convention and nothing else.

How much data sits behind a score

Select a score in the table to see how much data sits behind it. The exact number of families and sources is on each model's page.

  • highPlenty of data behind it: ten or more benchmark families, including trusted independent boards.
  • mediumA fair amount of data behind it: five to nine families.
  • lowLittle data behind it: three or four families, shown muted as an estimate.