[FEATURE]: Expose opt-in 1M context variants for GPT-5.6 OAuth #46527

Activity

1. github-actions

   assigned

   jlongster

   on Sep 1, 2026
2. github-actions commented on Sep 1, 2026

   

   github-actionsBot

   on Sep 1, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • fix(openai): OAuth transform treats Codex product budget as endpoint context limit #44821: Reports that OpenCode's OAuth transform uses Codex product budget as the endpoint context limit for GPT-5.6 Sol, covering the same underlying problem (incorrect/conservative OAuth context limits) and overlapping proposed resolutions (dynamic or override-based limit resolution vs. opt-in 1M variants).
3. rhan637 commented on Sep 1, 2026

   

   rhan637

   on Sep 1, 2026

   Author

   More actions

   Thanks for pointing this out. I believe #46527 is related to #44821, but it proposes a narrower and different change.

   #44821 investigates the physical maximum input accepted by the OAuth endpoint and discusses dynamic or override-based limit resolution. This request does not attempt to determine or expose that physical endpoint maximum.

   Instead, it proposes exposing the opt-in long-context tier already defined by Codex itself:

   • The bundled GPT-5.6 Sol, Terra, and Luna metadata uses context_window: 272000 and max_context_window: 872000.
   • Codex has a test confirming that setting model_context_window to 1_000_000 resolves these three models to 872_000.

   Sources:

   • https://github.com/openai/codex/blob/6be2a6ca952ac9f70676ce4dd07fda27175aa9dd/codex-rs/models-manager/models.json#L30-L31
   • https://github.com/openai/codex/blob/6be2a6ca952ac9f70676ce4dd07fda27175aa9dd/codex-rs/models-manager/src/manager_tests.rs#L731-L751
   • https://github.com/openai/codex/blob/6be2a6ca952ac9f70676ce4dd07fda27175aa9dd/codex-rs/models-manager/src/model_info.rs#L25-L33

   OpenCode currently exposes only the conservative OAuth limits. #46527 proposes mirroring Codex's two-tier behavior:

   • keep the existing 272K input budget as the default;
   • add an explicit opt-in model entry using a 1M total context budget, with 872K input and 128K output;
   • leave API-key authentication and existing defaults unchanged.

   Therefore, #44821 addresses the broader endpoint-capacity and dynamic-resolution problem, while #46527 proposes a specific selectable model-catalog feature based on Codex's existing long-context ceiling.
4. mary-ext commented on Sep 4, 2026

   

   mary-ext

   on Sep 4, 2026

   Last edited by mary-ext

   More actions

   if anyone comes into here looking for a plugin form

   import type { Plugin } from "@opencode-ai/plugin";
   
   const plugin: Plugin = async () => {
     return {
       provider: {
         id: "openai",
         async models(provider, ctx) {
           if (ctx.auth?.type !== "oauth") {
             return provider.models;
           }
   
           const models = { ...provider.models };
           for (const [modelId, model] of Object.entries(provider.models)) {
             if (!modelId.startsWith("gpt-5.6-")) {
               continue;
             }
   
             const longModelId = `${modelId}-1m`;
   
             models[longModelId] = {
               ...model,
               name: `${model.name} (1M context)`,
               limit: {
                 context: 1_000_000,
                 input: 872_000,
                 output: 128_000,
               },
             };
           }
   
           return models;
         },
       },
     };
   };
   
   export default plugin;

   ❤️React with ❤️2Reacted by Rodrigo Vargas and ChrisLazaridis
5. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   Thanks for the detailed proposal and the Codex references! You're asking for opt-in -1m versions of the GPT-5.6 Sol/Terra/Luna models (and their Fast versions) on ChatGPT sign-in. They would have a 1M context / 872K input / 128K output limit, while the current models keep their 272K input limit. Maintainers will consider it.

   This was filed while v1 was the main line, and #46574 was closed asking for a port to v2. The request still applies to v2, where the cap now lives in a separate built-in plugin, opencode.provider.openai-context-caps (#58194).

   v2 already has two documented ways around the cap:

   • Turn the cap off for all models: "plugins": ["-opencode.provider.openai-context-caps"]
   • Override a single model's limit in config. feat(core): move ChatGPT context caps into a separate plugin #58194 says config is applied after the cap plugin.

   To get a long-context entry next to the default one, you can also try a config alias with modelID (see "Aliases" in the models docs). I haven't checked that an alias's limit takes precedence over the cap, so please confirm it shows the higher limit:

   {
     "providers": {
       "openai": {
         "models": {
           "gpt-5.6-sol-1m": {
             "modelID": "gpt-5.6-sol",
             "name": "GPT 5.6 Sol (1M)",
             "limit": { "context": 1000000, "input": 872000, "output": 128000 }
           }
         }
       }
     }
   }

   Related issues:

   • models: allow model variants to override limit (context/output) #56582: lets model variants override limit, which would allow a 1m variant instead of a separate model
   • [FEATURE]: Allow 1M context window for GPT 5.6 #43041: earlier request for a 1M context setting for GPT-5.6
   • openai: ChatGPT OAuth reports 400k context for long-context models #47646 and fix(openai): OAuth transform treats Codex product budget as endpoint context limit #44821: discussion of the ChatGPT OAuth context caps
6. rhan637

   closed this as completedon Oct 29, 2026