shumengya/sproutclaw
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(ai): correct Azure gpt-5.4/5.5 context window and gpt-5-pro maxTokens

Browse Source 
Azure Foundry deploys gpt-5.4 and gpt-5.5 with a 1,050,000 context
window, but the Azure provider cloned OpenAI-direct's 272k API limit.
Override the context window in the Azure derivation.

Also fix gpt-5-pro maxTokens, which upstream metadata set to 272000 (a
duplicate of the input sub-limit) instead of the actual 128000 max
output; corrected at the source so OpenAI-direct and Azure both match.

closes #5559
This commit is contained in:
Armin Ronacher
2026-06-09 22:57:20 +02:00
parent 5a9d72ea02
commit 6b5923f107
2 changed files with 14 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
packages/ai/CHANGELOG.md
View File

@@ -13,6 +13,8 @@
- Fixed OpenCode completions model metadata to send explicit `maxTokens` as `max_tokens` ([#5331](https://github.com/earendil-works/pi/issues/5331)).
- Fixed Moonshot Kimi thinking-off requests to send the provider's `thinking: { type: "disabled" }` compatibility parameter ([#5531](https://github.com/earendil-works/pi/issues/5531)).
- Fixed Azure OpenAI Responses requests to disable server-side response storage ([#5530](https://github.com/earendil-works/pi/issues/5530)).
- Fixed Azure GPT-5.4 and GPT-5.5 context window metadata to 1,050,000 tokens, matching Azure Foundry deployments instead of OpenAI's 272k limit ([#5559](https://github.com/earendil-works/pi/issues/5559)).
- Fixed OpenAI and Azure GPT-5 Pro `maxTokens` metadata to 128,000, correcting an upstream value that duplicated the 272,000 input sub-limit as the output limit ([#5559](https://github.com/earendil-works/pi/issues/5559)).
## [0.79.0] - 2026-06-08

12
packages/ai/scripts/generate-models.ts
View File

@@ -1366,6 +1366,11 @@ async function generateModels() {
candidate.contextWindow = 272000;
candidate.maxTokens = 128000;
}
// models.dev reports gpt-5-pro output as 272000 (a duplicate of the input sub-limit);
// the actual max output is 128000. Also propagates to the derived Azure clone.
if (candidate.provider === "openai" && candidate.id === "gpt-5-pro") {
candidate.maxTokens = 128000;
}
// Keep selected OpenRouter model metadata stable until upstream settles.
if (candidate.provider === "openrouter" && candidate.id === "moonshotai/kimi-k2.5") {
candidate.cost.input = 0.41;
@@ -2064,6 +2069,12 @@ async function generateModels() {
];
allModels.push(...vertexModels);
// Azure Foundry deploys these with larger context windows than OpenAI's own API,
// which caps gpt-5.4/gpt-5.5 at 272k. See models-sold-directly-by-azure docs.
const AZURE_CONTEXT_WINDOW_OVERRIDES: Record<string, number> = {
"gpt-5.4": 1050000,
"gpt-5.5": 1050000,
};
const azureOpenAiModels: Model<Api>[] = allModels
.filter((model) => model.provider === "openai" && model.api === "openai-responses")
.map((model) => ({
@@ -2071,6 +2082,7 @@ async function generateModels() {
api: "azure-openai-responses",
provider: "azure-openai-responses",
baseUrl: "",
contextWindow: AZURE_CONTEXT_WINDOW_OVERRIDES[model.id] ?? model.contextWindow,
}));
allModels.push(...azureOpenAiModels);