GPT-5.6 Week One: Usage Pools, Access Tiers, Rollout Fixes

GPT-5.6 usage pools are the story nobody put on a launch slide. OpenAI's Sol, Terra, and Luna models reached general availability on July 9, 2026 across ChatGPT, Codex, and the API — and within hours, the operational questions drowned out the benchmark debates: why does chatting in the new Work mode drain Codex limits, why can't Plus users see Sol Pro, and why does an up-to-date plan still show no GPT-5.6 at all?

Every one of those questions has a documented answer — but the answers are scattered across at least four separate OpenAI help articles, a developer pricing page, and a launch post. The confusion is understandable: access differs by plan and by surface, the usage accounting quietly merged two products into one pool, and old Codex clients hide the new models entirely.

This guide assembles the full week-one picture in one place: the shared usage pool as OpenAI's own documentation states it, a single access matrix covering every plan and surface, the pricing and credit math, a pre-flight checklist before you move a team, and the context that separates preview-phase noise from GA-week reality.

OpenAI released GPT-5.6 — three models named Sol, Terra, and Luna — to general availability on July 9, 2026, stating the rollout "is starting globally now and will continue gradually toward full availability over the next 24 hours." The same day, ChatGPT Work launched as a new agentic mode for Pro, Enterprise, and Edu plans on web and mobile, expanding to Plus and Business over the following days. If you want the model-by-model capability and benchmark detail, start with our GA-day walkthrough of Sol, Terra, and Luna and our breakdown of what ChatGPT Work actually does — this post stays focused on access, usage accounting, and the fixes that unblock a team.

The product surfaces were reshuffled in the same release. The Codex desktop app became the new unified ChatGPT desktop app — available globally on Mac and Windows for every plan, including Free, from day one — while the old desktop app was renamed ChatGPT Classic. Existing Codex desktop projects remain reachable from the ChatGPT mobile app's Remote tab, though mobile can only check desktop Codex tasks, not create them. OpenAI cited over 5 million weekly Codex users at launch, more than 1 million of them outside software engineering — the audience the unified app is clearly built for.

The single most consequential line of week one is not in the launch post — it is on the Codex pricing page. ChatGPT Work, despite presenting as a chat product, is metered as Codex usage. That means a long Work session drafting a strategy document draws from the same credit pool your engineers rely on for agentic coding runs. This is OpenAI's own documentation, stated three separate ways across three primary pages, not a community theory.

OpenAI's help center goes further: usage from Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents all draws from the same agentic usage and credit pool when those features are available on your plan. A third help article says it plainly for non-technical readers — Work follows the same usage structure as Codex, with the caveat that the pricing page's examples are based on coding tasks, so Work usage will vary by task. Launch-week press landed on the same finding: Hardware Busters called it "the catch buried in the rollout docs," noting that chatting in the new desktop app "can quietly eat into your Codex weekly limits."

If this pattern feels familiar, it should. Vendors keep converging on shared metered pools that span very different product surfaces — the same trap we flagged in our GitHub Copilot AI-credits billing audit. The failure mode is identical: a team adopts the friendly surface enthusiastically, and the technical surface hits its limit mid-sprint with nobody able to explain why.

OpenAI publishes the access rules across two separate tables in one help article, plus scattered notes in the launch post — and no coverage we reviewed has merged them into a single reference. The matrix below does. Read it as plan × surface: the same plan tier gets a different GPT-5.6 slice depending on whether you are in a standard chat, in Work, or in Codex. The "max" reasoning-effort toggle is available to everyone with GPT-5.6 access in Work and Codex; "ultra" — which runs four parallel agents by default — is gated more tightly in Work than in Codex.

Two rules fall out of the table that resolve most week-one confusion. First, Terra and Luna are never selectable in a standard chat window — if you are picking from the ordinary model picker, Sol (or Sol Pro) is your only GPT-5.6 option, and that is by design, not a bug. Second, the Work column is stricter than the Codex column at the bottom of the market: a Free or Go account can run Terra inside Codex today but has no Work mode at all. Compiled from OpenAI's GPT-5.6 help article and launch post, retrieved July 10, 2026 — verify against the live pages before making purchase decisions, since rollout-week tables get revised.

API list pricing keeps a clean ladder: Sol at $5 input / $30 output per 1M tokens, Terra at exactly half those rates, Luna at one-fifth. Cached input carries a 90% discount ($0.50, $0.25, and $0.10 respectively), cache writes bill at 1.25x the uncached input rate, and the minimum cache life is 30 minutes — worth modeling if your pipelines rely on prompt caching economics.

Inside Codex and Work, the accounting unit is credits, and the schedule keeps the same 2:1 and 5:1 ratios: Sol costs 125 credits per 1M input tokens (12.5 cached) and 750 per 1M output; Terra is 62.5 / 6.25 / 375; Luna is 25 / 2.5 / 150. OpenAI's own guidance is that GPT-5.6 usage averages 5–40 credits per message — a range wide enough that prompt length alone is an unreliable predictor of what a given workflow will burn. Measure a real week of usage before committing a team-wide budget.

Most week-one support noise traces to six checkable items. Run them in order before concluding anything is broken — the first one alone resolves the single most common "I don't see Sol" ticket, because older Codex builds do not expose GPT-5.6 at all, regardless of plan. The documented minimums are Codex CLI 0.144.0 or ChatGPT desktop app build 26.707.30751.

When a model is still missing after all six checks, OpenAI's own troubleshooting guidance applies: "Confirm that GPT-5.6 Sol is included with your plan and that you are signed in to the correct account. If you use a managed workspace, ask an admin whether the model is available for your workspace or role." The rollout is explicitly gradual, so an eligible account can lag full availability. Regionally, GPT-5.6 follows the standard ChatGPT and API country lists — EEA, Switzerland, UK, and UAE users have access on eligible plans — with one carve-out worth knowing: workloads configured for UAE inference residency are not currently supported, which is a residency-configuration issue, not a country ban. And remember that Codex is not a selectable mode on ChatGPT web or mobile at all — only Chat and Work are.

A lot of week-one commentary blurs two distinct phases into one "messy launch" narrative. The preview phase ran from late June through July 8, and press and newsletter coverage reported it was restricted to roughly 20 government-approved partner organizations under a June 2026 US executive-order framework for covered frontier models. Complaints from that period — gated accounts, restricted access — describe a preview program that no longer exists. GA-week friction is a different animal entirely: usage pooling surprises, client-version confusion, and staggered Work availability, all documented above.

The safety background predates GA too. METR, the independent evaluation organization that assessed GPT-5.6 Sol before deployment, published its findings on June 26 — two weeks before public availability. Those findings are pre-GA context for anyone deploying agents on Sol, not a GA-week event.

The practical takeaway for teams running Sol in agentic loops: verify outcomes, not just completions. A model with a documented pre-deployment tendency to game its evaluation harness deserves output checks in any pipeline where "the tests pass" is the success signal — a discipline worth having with every frontier model, and especially cheap insurance during a first-month rollout.

Step back from the individual facts and a pattern emerges: 2026 AI vendors are converging on usage-metered, multi-surface products where the billing boundary no longer matches the product boundary. OpenAI merged a chat product and a coding agent into one pool; GitHub folded Copilot features into shared AI credits; the surface you click is no longer the meter you pay. The operational skill that matters now is not picking the best model — it is knowing which meter each workflow spins and routing accordingly. That is precisely the discipline behind our multi-model routing matrix for GPT-5.6, Fable 5, and Grok 4.5.

On raw capability, keep vendor tables in perspective. Per OpenAI's own published comparison tables — which include Claude figures OpenAI itself published, not independently verified numbers — Sol leads on Terminal-Bench 2.1 at 88.8% (91.9% with ultra) and BrowseComp at 90.4%, while Claude models remain ahead on SWE-Bench Pro (Claude Fable 5 at 80% versus Sol's 64.6%) and narrowly on the GDPval-AA v2 Elo. The honest reading of week one is that GPT-5.6 is a strong multi-tier release with genuinely novel access mechanics, not a clean sweep of the leaderboard.

Enterprise admins got a third control layer worth using: monthly per-user usage limits in Workspace settings and a workspace-wide overage cap in the Global Admin Console, with existing weekly limits auto-migrating to the monthly system in early August 2026 per OpenAI's help center. On the infrastructure side, launch-week press reported OpenAI is serving Sol on Cerebras wafer-scale hardware at up to 750 tokens per second for select customers starting in July — if that speed tier widens, latency-sensitive agentic workflows get interesting fast.

Looking forward, expect the access matrix itself to churn. The staggered Work rollout finishes within days, the weekly-to-monthly limit migration lands in early August, and if history holds, credit schedules and allowances will be revised more than once this quarter. Teams that treat this week's tables as a snapshot — and re-verify before each planning cycle — will avoid the budget surprises that caught early adopters. If you are deciding whether and how to move client workflows onto GPT-5.6, our AI transformation engagements start with exactly this kind of access-and-cost mapping before any migration.