|
1 | 1 | # Changelog |
2 | 2 |
|
| 3 | +## Franklin Agent 3.35.5 — the last three OpenAI ceilings, measured |
| 4 | + |
| 5 | +| model | was | now | |
| 6 | +|---|---|---| |
| 7 | +| `openai/gpt-5.5` | 32,768 | 128,000 | |
| 8 | +| `openai/gpt-5.4` | 32,768 | 128,000 | |
| 9 | +| `openai/gpt-5-mini` | 16,384 | 65,536 | |
| 10 | + |
| 11 | +These took three attempts to get right, and the story is worth keeping because |
| 12 | +the table is where a wrong belief gets frozen. |
| 13 | + |
| 14 | +The first attempt set gpt-5.5 and gpt-5.4 to 100,000, citing "upstream's stated |
| 15 | +maximum". There was no upstream response: both BlockRun SDKs rejected |
| 16 | +`max_tokens > 100000` client-side, so 19 probes at 128,000 never reached a |
| 17 | +provider. The uniform limit across three unrelated vendors should have been the |
| 18 | +tell. That PR was closed. The second attempt put them back at 32,768 and |
| 19 | +recorded the catalog's 128,000 as unrefuted but unverified — honest, but still |
| 20 | +not measured. |
| 21 | + |
| 22 | +Re-probed with the SDK guard bypassed: every model advertising a ceiling above |
| 23 | +100,000 accepts it, including `zai/glm-5.2` at 262,144. The catalog was right |
| 24 | +the whole time; the measurement was broken. Fixed at the root in |
| 25 | +blockrun-llm 1.8.1 and @blockrun/llm 3.8.1, which raise that guard to a |
| 26 | +1,000,000 typo bound and reword the message so it no longer reads like a |
| 27 | +provider response. |
| 28 | + |
| 29 | +Franklin never ran that guard — both request paths use raw fetch and import |
| 30 | +only payment helpers from the SDK — so these values do not depend on the SDK |
| 31 | +release. |
| 32 | + |
| 33 | +Effect is on the escalation path. A normal turn still sends |
| 34 | +`min(16384, ceiling)`; after a `max_tokens` stop the ceiling for gpt-5.5 and |
| 35 | +gpt-5.4 goes from 32,768 to 65,536 (`ESCALATED_MAX_TOKENS`), so recovery |
| 36 | +doubles instead of clamping back below where it started. |
| 37 | + |
3 | 38 | ## Franklin Agent 3.35.4 — Haiku and Sonnet 4.6 get their real output ceilings |
4 | 39 |
|
5 | 40 | **Haiku 4.5 goes from 16,384 to 64,000 max output. Sonnet 4.6 from 64,000 to |
|
0 commit comments