dated-break-eval / RESULTS.md
CyberNative AI LLC
Add gemma-4-26B results (version 2)
c2b857d
|
Raw History Blame Contribute Delete
7.08 kB

Do coding models walk into known dated breaks?

Run date 2026-11-14 · four models · one sample per prompt.

We asked each model for 58 ordinary pieces of code where a published upstream change breaks the obvious answer on a known date: an OpenAI or Gemini model that is being shut down, a Docker tag that moves, a GitHub runner image that is retired, a data API that is being switched off. We then gave the same prompts with today's date, and with a list of the dated breaks.

The headline

Share of scored replies that chose a route already broken or dated to break (the prompt's own calendar row):

Model A: plain B: + today's date C: + calendar
claude-sonnet-5.5 18 of 49 (37%) 16 of 52 (31%) 6 of 37 (16%)
gpt-6.1-sol 20 of 53 (38%) 19 of 53 (36%) 10 of 49 (20%)
gpt-oss-20b (MXFP4 GGUF) 19 of 48 (40%) 20 of 48 (42%) 19 of 41 (46%)
gemma-4-26B-A4B-it (UD-Q4_K_M GGUF) 17 of 48 (35%) 21 of 53 (40%) 19 of 52 (37%)

Paired by prompt (only prompts scored in both arms), with 95% intervals. "Headline" is the rate above; "wider" also counts routes already shut down or dated to break that are not on the prompt's own row.

Model Pairs A/C A minus C, headline A minus C, wider A minus B, headline A minus B, wider
claude-sonnet-5.5 35 0.29 (0.09 to 0.48) 0.26 (0.05 to 0.47) 0.06 (−0.07 to 0.20) 0.13 (0.02 to 0.24)
gpt-6.1-sol 48 0.15 (0.02 to 0.30) 0.15 (0.02 to 0.30) 0.00 (−0.06 to 0.05) 0.00 (−0.06 to 0.05)
gpt-oss-20b 37 0.00 (−0.08 to 0.08) 0.03 (−0.06 to 0.12) −0.02 (−0.14 to 0.10) −0.07 (−0.16 to 0.03)
gemma-4-26B 46 0.02 (−0.07 to 0.13) 0.02 (−0.07 to 0.13) 0.00 (−0.07 to 0.05) −0.02 (−0.07 to 0.00)

The date alone did not move the headline rate for any of the four models. On the wider rate, Sonnet dropped a little with the date (0.13, interval 0.02 to 0.24); we claim nothing beyond that.

Among replies that could be scored in both arms, the calendar text helped the two closed models, not the two open-weight models. For Sonnet that is 35 of 58 prompts, because 21 of its arm C replies could not be scored (see below). gpt-oss-20b's arm C rate looks higher (46%) only because a different set of replies was scorable; on the 37 prompts scored in both arms the rate is the same. On gemma-4-26B's 46 prompts scored in both arms, the calendar moved two replies off a break (the two Python prompts below) and one onto it.

Where they walked in

  • OpenAI speech and transcription. Asked to turn text into speech or transcribe a file with OpenAI's API, every scored reply (46 of 46), from every model, in every arm, used a model or endpoint OpenAI has dated for removal (whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, or the Speech endpoint with gpt-4o-mini-tts or tts-1). That includes arm C, where the calendar row was in the context. For speech, OpenAI's replacement is a different API (Realtime).
  • Already gone. Asked for an OpenAI image, all 23 scored replies named a model already shut down (gpt-image-1 on 23 October 2026, dall-e-3 on 12 May 2026). These are outside the headline (they are not on that prompt's row) and counted in the wider rate.
  • Already ended, still written. In every arm, gpt-oss-20b wrote old Android back-button code, which Google says is not called when an app targets API 36 on Android 16 (Google Play's last extension for that target ended on 1 November 2026; 6 of 6 replies on two prompts). Asked for a test matrix from Python 3.8 to 3.13, it put 3.8 on ubuntu-latest, where setup fails once the label moves to Ubuntu 26.04 (rollout 19 October to 19 November 2026; 3 of 3). On those three prompts the closed models chose the current route, apart from one arm C reply from gpt-6.1-sol that used both the old and the new back-button code. gemma-4-26B chose the current back-button code every time, but broke the same Python 3.8 prompt in arms A and B, and a second Python prompt (3.9) the same way.
  • node:lts and Yarn. On one prompt (GitLab CI for a Yarn project), all four models used node:lts in arms A and B. That image dropped Yarn on 28 October 2026. With the calendar, three of them installed Yarn or Corepack themselves; gemma-4-26B did not. On a second Yarn prompt (a Dockerfile), gpt-oss-20b (arms B and C) and gemma-4-26B (all three arms) used an lts Node image, calendar included.
  • Where the calendar worked. For the closed models, USGS WaterServices (switched off 22 February 2027), Socrata rows.csv downloads (16 December 2026) and XSLT in Chrome (17 November 2026) broke in arm A and did not break in arm C. For gemma-4-26B, among the 58 headline prompts, the two Python prompts were the only ones where the calendar turned a break into a live route. Outside the headline, it did the same on gemma-4-26B's two Rails prompts.

The per-row table is in row-table.md.

What this does not show

  • PASS means a live route was chosen, not that the code runs. A lower break rate is not a higher success rate.
  • Arm C has more unscored replies for three of the four models. With the calendar, replies more often left the model name out of the code: they read it from an environment variable, or picked one at run time from the provider's model list. 21 of 58 Sonnet replies in arm C are unscored (9 for Sol, 17 for gpt-oss, 6 for gemma, one of them an unknown ID). Sonnet's arm C rate is 10% if none of them is a break and 47% if all are (52% if its 3 mixed replies also count as breaks).
  • Echoing the calendar is not following it. Arm C replies that repeated an ID or term from the calendar still broke in 3 of 7 (Sonnet), 3 of 9 (Sol), 4 of 6 (gpt-oss) and 3 of 9 (gemma). gemma-4-26B's one reply that broke only with the calendar used a Socrata rows.csv download link, the route the calendar line says stops working.
  • Small and single-shot. 64 prompts from 32 calendar rows, one sample per prompt. The prompts cover only breaks our calendar lists, and we make that calendar.
  • Four models. Ornith-1.5-35B-A3B and Qwen3.8-27B were started and did not finish within the run's 8 GPU-hour budget. A part-run is not scored, so they are not here.
  • One ID was not scored. For the Gemini Live prompt, gemma-4-26B wrote gemini-2.0-flash-exp in all three arms; that ID is not in our status table, so those 3 replies are outside every rate.

Check it yourself

Prompts, checks, per-reply outcomes and the scorer are in this dataset. To see which model IDs in your own code have a shutdown date, paste it into the model shutdown checker. Every dated break we track, with the fix, is on the Breakage Calendar.

Prepared by CyberNative AI LLC, an AI-run company, using AI agents. Closed models were called through OpenRouter for evaluation only.