GPT 5.5 **SMALLER** context limits are not respected

Description

opencode uses hard coded context limits values for gpt-5.5, ignoring values set in opencode.jsonc.

I suspect it's due to PR #24212 which added this to codex.ts:

// gpt-5.5 models temporarily have restricted context window size for codex plans
if (model.id.includes("gpt-5.5")) {
  model.limit = {
    context: 400_000,
    //@ts-expect-error incorrect type for v1 sdk but works
    input: 272_000,
    output: 128_000,
  }
}

It should probably either overwrite only when the values are not specified in the config, or use max(400_000, model.limit.context) (same for input/output) to allow lower values (which is useful because gpt-5.5 is very expensive)

Plugins

None

OpenCode version

1.14.28

Steps to reproduce

Add this to ~/.config/opencode/opencode.jsonc in provider.openai.models:

"gpt-5.4": {
  "limit": {
    "context": 200000,
    "input": 150000,
    "output": 50000,
  },
},
"gpt-5.5": {
  "limit": {
    "context": 200000,
    "input": 150000,
    "output": 50000,
  },
},

Run opencode models openai --verbose and check context limits for both models. gpt-5.4 will respect the overrides, gpt-5.5 will not. Alternatively, start a conversation with gpt-5.5, you'll see that the context window stats are relative to 400k context window not 200k.

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

This issue might be a duplicate of (or closely related to) an existing issue. Please check:

GPT-5.5 Codex context limit dropped to 400k but uses API context limit (1M) causing "exceeds the context window of this model" errors #24171: Same root cause (PR fix: ensure gpt-5.5 compacts at correct context size when using openai oauth #24212 hardcoding gpt-5.5 context limits in codex.ts). That issue reports that the 400k Codex limit overrides the larger API limit, while this issue reports that the hardcoded values override user-configured lower limits. They are distinct symptoms but share the same underlying code.

Confirming this. I just chased the same bug — running openai/gpt-5.5 via Codex OAuth on a Pro/Plus plan, compaction kicks in around ~250k tokens despite the model having 922k input / 1M context.

Root cause is exactly what you identified: CodexAuthPlugin.auth.loader in packages/opencode/src/plugin/codex.ts:394-402 unconditionally replaces model.limit for any gpt-5.5 model:

if (model.id.includes("gpt-5.5")) {
  model.limit = {
    context: 400_000,
    input: 272_000,
    output: 128_000,
  }
}

The user-config override in provider.ts runs after this, but only applies fields the user explicitly set:

if (configLimit) {
  if (configLimit.context) model.limit.context = configLimit.context
  if (configLimit.input) model.limit.input = configLimit.input
  if (configLimit.output) model.limit.output = configLimit.output
}

So a user who only sets { context: 1050000, output: 128000 } (the natural pattern, since input is implied) silently keeps input: 272000 from the plugin hardcode. With usable() = input - reserved = 252_000, compaction triggers at the first 250k+ message.

Debug evidence — the relevant overflow check from a real session before the workaround:

service=session.overflow modelID=gpt-5.5 
  tokens={"total":647998,"input":1,"output":158,"reasoning":0,"cache":{"write":524,"read":647315}} 
  count=647998 limit=252000 overflow=true 
  modelLimit={"context":1050000,"input":272000,"output":128000}

Note modelLimit.input=272000 despite context=1050000 — that's the user's context override applied while input stayed at the codex hardcode.

Workaround: explicitly add all three fields to user config:

"openai": {
  "models": {
    "gpt-5.5": {
      "limit": {
        "context": 1050000,
        "input": 922000,   // <-- this is the field that gets silently overwritten
        "output": 128000
      }
    }
  }
}

Proposed real fix: when user provides any explicit limit override, fall back to the original models.dev snapshot for unspecified fields (not to whatever a plugin auth.loader has mutated the limit to). I have an implementation on a fork: sjawhar@f4ab23a6270e — it captures the original limits before plugin loaders run, then uses them as the fallback in the override loop. The plugin still gets to enforce its restriction when no user override exists, but the user's intent wins when they have explicitly configured the model.

Tested with the same gpt-5.5 + Codex OAuth setup — compaction now correctly fires near the 902k threshold (922k input - 20k reserved) instead of 252k.

this is NOT a bug, this is the context limit for the openai codex oauth model right now, only the api gets the 1M context window

Correction to my earlier comment — my proposed "snapshot fallback" fix was wrong, and I've reverted it.

The 400k context cap on Codex (ChatGPT subscription) gpt-5.5 is server-side enforced by Codex, not just a client-side hardcode. Per OpenAI's launch announcement (Apr 23, 2026):

In Codex, GPT-5.5 is also available for Edu and Go plans with a 400K context window

The context window in Codex is 400K tokens, not the 1M available through the Chat API.

API access is coming "very soon"

So CodexAuthPlugin's hardcode {context: 400_000, input: 272_000, output: 128_000} correctly reflects what Codex actually serves. Letting users override this with snapshot defaults causes the conversation to grow past Codex's hard cap, then triggers reactive context_length_exceeded errors and emergency compaction (verified — the binary I built with the override-bypass fix produces noticeably worse behavior than the default).

The original report (this issue and #24561) appears to actually be the same Codex tier limit, not a config-respect bug. The hardcode is doing its job.

The legitimate underlying problem is: users on the developer API tier (1M context) currently can't use it because gpt-5.5 hasn't been released for general API access yet (per OpenAI's announcement). Once it ships:

API key auth (auth.type === "api"): codex.ts loader doesn't fire, so the snapshot's 1M is used → user can choose to override
OAuth/Codex auth (auth.type === "oauth"): plugin enforces 400k → matches server reality

For now, ChatGPT subscription users have to live with 400k context (252k usable input after the 20k compaction buffer) for gpt-5.5. That's an OpenAI product decision, not an opencode bug.

Apologies for the earlier confusion. Reverted: sjawhar@f3872ba848b2

@rekram1-node

this is NOT a bug, this is the context limit for the openai codex oauth model right now, only the api gets the 1M context window

@rekram1-node I'm not talking about 1 million context window. I'm trying to set a shorter length of 200k, and it gets changed to the hardcoded values.

Ah sorry the other notes confused me, yeah let me look at that

Thanks, I also changed the title to make it clear.
Do I need to open a new issue or is it possible to reopen this one?

re-opened

Navigation Menu

Repository navigation

GPT 5.5 SMALLER context limits are not respected #24751

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

GPT 5.5 **SMALLER** context limits are not respected #24751

Description

Description

Plugins

OpenCode version

Steps to reproduce

Screenshot and/or share link

Operating System

Terminal

Activity

github-actions commented on Apr 28, 2026

sjawhar commented on Apr 29, 2026

rekram1-node commented on Apr 29, 2026

sjawhar commented on Apr 29, 2026

asaf92 commented on Apr 29, 2026

rekram1-node commented on Apr 29, 2026

asaf-genie commented on Apr 30, 2026

rekram1-node commented on Apr 30, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

GPT 5.5 SMALLER context limits are not respected #24751