sysnote
Analysis

Codex GPT-6.1 Sol usage drain: 15 GitHub reports, no OpenAI reply, and no release that lines up

Since November 10, openai/codex #60530, "GPT-6.1 Sol Usage Limits Depleting Much Faster Than Before", has become the place where Codex users report that GPT-6.1 Sol uses up their 5-hour or weekly allowance in minutes. We went through the openai/codex issue tracker from October 21 to November 16 and built a small dataset of every issue that blames GPT-6.1 Sol for a drained allowance. There are 15 such issues. None of them has a reply from OpenAI, none has a linked fix, and they don't point to any one release. The only Codex metering fault anyone has actually measured in that period was on GPT-6.1 Terra, not Sol, and it went away after a day.

Everything below that comes from the issues is user-reported. OpenAI hasn't confirmed a bug, and it hasn't said there isn't one.

The dataset

We counted issues where the reporter names GPT-6.1 Sol as the model and describes the allowance going much faster than expected. We left out "Selected model is at capacity" errors, HTTP 429s, and reports about other models. That leaves 15 issues:

Issue Opened (UTC) Plan Client Effort What the reporter saw
#55876 Oct 21 Plus CLI 0.163.0 medium about half the allowance in under an hour; no subagents
#55997 Oct 21 Plus App 26.1016.51333 high 5-hour allowance gone in 11 minutes; three subagents spawned
#56240 Oct 22 Business (Team) not given not given fast drain "after recent update"
#56244 Oct 22 Pro 200 App 26.1016.61520 ultra 60% used about an hour after a reset
#56618 Oct 24 Plus App 26.1016.62342 low 83% of the 5-hour allowance in two hours
#56672 Oct 24 Plus Work/Codex not given a freshly reset 5-hour allowance gone after 4m21s
#57236 Oct 26 Plus App 26.1021.31345 low/medium fast 5-hour drain
#59235 Nov 3 Plus CLI 0.166.0 xhigh usage while idle
#59639 Nov 5 Business Premium 0.165.0 high/xhigh 8% of the weekly limit in about an hour on Sol
#60421 Nov 9 Pro 100 App 26.1029.61003 max/high allowance "disappearing"; Sol plus Astra
#60530 Nov 10 not stated App 26.1029.70936 unchanged weekly limit at 100.2% for Nov 2–9; one conversation took 88.96%
#60704 Nov 11 Pro 500 App 26.1103.60933 xhigh, Standard speed 17 weekly points in about five Sol-only hours
#61266 Nov 13 not stated Desktop medium workers, Astra high coordinator weekly limit at 99% after a multi-agent run
#61471 Nov 14 not stated Windows ultra the whole 5-hour allowance in one 7m33s task
#61858 Nov 16 €220/month plan not given not given the weekly allowance in about a day

Between them the 15 issues have 44 comments. Fifteen of those are the repository's duplicate-detection bot. The other 29 are users, and none of them is an account associated with the repository. All 15 issues are open, and each carries the same automatic bug and rate-limits labels. They have six reactions in total. #60530, the thread everyone links to, has 11 comments and no reactions at all.

Several things in the table cut against a single, simple cause:

  • It isn't one release. The reports span Codex app builds from 26.1016.51333 to 26.1103.60933 and CLI versions 0.163.0 through 0.166.0, which is about three weeks of releases. The comments add the CLI, JetBrains and VS Code clients (one Plus user says it happens in all three).
  • It isn't one effort level. Every reasoning level from low to ultra shows up. In #60530, one Plus commenter reports the drain on Sol at Low with Fast mode off.
  • It isn't only subagents. Four of the 15 involve subagents. #55876 says none were launched, and the #60530 author writes that the drain began before they started using them.
  • It isn't new in November. Seven of the 15 were filed between October 21 and 26. The next week had none, the week after had three, and the week of November 10 had five. So the November thread is the second wave of the same complaint, not its start.
  • It isn't only Sol. In the same period, issues with the same symptoms name GPT-6 Astra, GPT-6.1 Terra, GPT-6 Luna, GPT-6 Sol and GPT-5.6 Sol. #60425, which the duplicate bot keeps pointing to, is about GPT-5.5.

The volume doesn't show a November spike either. Issues carrying the rate-limits label (which also covers capacity errors and 429s) ran at 109, 194, 157 and 120 a week in September, 77 to 98 a week in early October, 152 in the week of October 20, and 91, 84 and 47 in the three weeks since. By that count, the week #60530 drew the most attention was the quietest week in this tracker since the start of September.

One more thing about the plans. Four reporters are on Pro tiers or a €220 plan. Pro has no 5-hour window: OpenAI's Codex pricing page says so, and Tibo said it again on November 8, as we reported. On Pro, "drained in minutes" can only mean the weekly meter.

What did change, on the record

Two things on the record bear directly on how fast Sol uses an allowance. One is a change, and it happened in October, not November.

On October 11, Tibo raised Sol's default context in Codex and Work:

TTibo@thsottiaux

Day 7/

We have increased default context length for GPT-6.1 Sol to 500k when used through Codex & Work, up from 272k or an ~ 80% increase.

We will go higher in the future, but it helps reduce occurrence of compactions and with tasks that require keeping a lot of context and…

October 11, 2026

A few minutes later he added that OpenAI was "not charging long context fees for GPT-6.1 Sol when drawing usage from the subscription". The pricing page still says this. A day earlier, defending the old default, he had written that Codex could run at 1M "but it would just cost more of the usage."

#59022 measured what that meant on one Pro account. The model_context_window the client reports went from 258,400 to 475,000 tokens between two sessions on October 11, on the same desktop build. Input per call on that user's own work rose from about 93k–130k tokens in September to about 184k–264k in late October, and auto-compaction moved from about 245k to 420k–454k. A larger window means fewer compactions, but every call in a long session resends more history.

The other is the price of a cache miss. OpenAI's rate card charges GPT-6.1 Sol 50 credits per million input tokens, 2.5 per million cached input tokens and 250 per million output tokens. A cached token costs one-twentieth of a fresh one. Here is what that means for a single agent call. The 2,000 output tokens and the 99% cache hit rate are our assumptions, chosen to resemble the logs in these issues:

One Sol call Context Credits
Cache hit 250k about 1.2
Cache hit 475k about 1.9
Cache miss 250k about 13
Cache miss 475k about 24

So at steady state, the bigger default context makes each call about 1.5 times as expensive late in a long session. That is noticeable, but it isn't "ten times faster". A cache miss is different: one miss costs about as much as ten or more hits. Anything that breaks the prompt cache, such as a changed tool list (as with Codex Desktop forks), an edited prefix or a server-side routing miss, multiplies the cost of that call.

#61757, filed on November 16, has the best data on how often that happens to Sol. Across 26,912 requests on CLI 0.167.0–0.167.2, requests within a turn missed the cache 0.16% of the time. The first request after a long turn was different: after turns of 15–30 minutes it missed 10 times out of 39, and after turns of 30 minutes or more, 5 out of 24. That doesn't match what Tibo wrote on October 21 (UTC), that for logged-in users "we keep cache forever" and are "guaranteed to hit the cache". The reporter asks a fair question: does the plan meter follow the cached_input_tokens the client sees? Nobody from OpenAI has answered. Even if it does, one 24-credit miss added to a turn of 60 or more calls makes that turn maybe 20% more expensive. That is real money, but it is still nowhere near the 5–10x in the complaints.

The one fault someone measured was Terra's

The most rigorous report in this period isn't about Sol. In #59690, CoreyHayward took about 21,900 token_count events from ~/.codex/sessions. Each event carries last_token_usage and the account's rate_limits (primary is the 5-hour window, secondary the weekly one). He worked out each call's cost as the change in used_percent since the previous event. On November 5, GPT-6.1 Terra calls suddenly cost 15.5 times more of the 5-hour limit per 100k input tokens than on November 1–4. Cache hit rate, model, effort and service tier were all unchanged, and GPT-6 Luna in the same thread was normal. The numbers lined up with the rate card only if Terra's cached input was being billed at the full input price. By 08:36 UTC on November 6, Terra was back to normal on the same CLI build, so whatever fixed it was server-side. By his estimate, the fault cost about 30 points of his weekly allowance. OpenAI never commented in the thread.

That matters here for two reasons. First, server-side metering faults like this do happen and can be fixed quietly. Second, they can be found from the client. The same report gives a Sol baseline on his Business plan: 0.42–0.50% of the 5-hour limit per call for gpt-6.1-sol at medium effort, over about 3,600 calls between October 15 and November 2. Nobody in the Sol threads has published the same per-call comparison for November. Without it, the reports can't tell a metering change apart from heavier sessions.

There is a second, sobering yardstick. OpenAI's own estimate for Plus is 15 to 160 local GPT-6.1 Sol messages per five hours, and the page says these "are not fixed message limits". A range that wide leaves room for a five-hour window to end after a dozen heavy agentic turns without anything being broken.

What's still unexplained

Some reports don't fit "bigger sessions". In #56672, a freshly reset Plus allowance ended after 4m21s. In #61471, one Sol task at ultra used the whole 5-hour window in 7m33s, and the reporter's local capture tool showed three entries attributed to GPT-6 Astra, about $16 of the tool's $20 estimate, even though Sol was the model selected. On the rate card, Astra costs five times as much as Sol per fresh input or output token and ten times as much per cached token. A subagent quietly running on Astra would explain a lot. The reporter is careful to say the capture tool hasn't been checked against OpenAI's records, and so are we. #61493 describes a related oddity: a "By model" usage view that puts 40.9% on GPT-6 Astra for a user who says they never selected it.

The data that would answer this already exists in Codex's own protocol. The app-server schema defines a ThreadUsage response with ThreadUsageBreakdownGroup entries split by model, reasoningEffort and speed, each with estimatedUsageCreditsMicros and cached and uncached input counts:

export type ThreadUsageBreakdownGroup = { model: string | null, reasoningEffort: string | null, speed: string | null, estimatedUsageCreditsMicros: bigint, netNewInputTokens: bigint | null, cachedInputTokens: bigint | null, inputTokens: bigint | null, outputTokens: bigint | null, totalTokens: bigint | null, };

None of the 15 issues includes one. It is the cleanest way to show whether a "Sol" task was really billed as Sol.

If your Sol allowance is going fast

All of these are documented settings or methods already used in the issues. None of them is a confirmed fix.

  • Check speed first. On included usage, Fast mode costs 2.5x Standard and GPT-6.1 Sol Ultrafast costs 8x, per OpenAI's Speed page. The #60704 reporter's 43 minutes of Ultrafast cost 7 weekly points. service_tier in config.toml sets the default.
  • Watch the window. /status in the CLI shows the context in use and both limits. If long sessions are what's expensive, model_context_window and model_auto_compact_token_limit are documented config keys. Lowering them trades the larger October default for more compactions, which is the trade-off Tibo described when he raised it.
  • Measure before and after. The token_count method from #59690 needs nothing but your own session files. It is what turned the Terra complaint into a number.
  • Look at the subagents. If the usage view shows a model you didn't pick, note which subagent ran it.

OpenAI has spent the past month handing out resets, most recently the banked one on November 12. It has said nothing about Sol metering. Tibo's only post since then about a Codex problem, on November 16, says the "model at capacity" errors may finally be fixed. A reset gives back allowance, but it doesn't tell anyone where the allowance went.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.