Short answer: Choose GPT-5.6 Sol for complex professional work, GPT-5.6 Terra for a balance of intelligence and cost, and GPT-5.6 Luna for cost-sensitive, high-volume tasks. OpenAI’s current model pages document a promotional Sol rate of $4 input / $0.40 cached input / $20 output per million tokens, available at least through November 21, 2026; Terra and Luna remain current documented rates of $2 / $0.20 / $12 and $0.20 / $0.02 / $1.20. Product availability depends on the surface: ChatGPT Chat, ChatGPT Work, Codex and the API do not share every model or billing rule.
A completed surface change is important: OpenAI says GPT-5.4 and GPT-5.4 mini stopped being available in Codex for users signed in with a ChatGPT account on August 31, 2026. The documented replacement direction is GPT-5.6 Terra for GPT-5.4 and GPT-5.6 Luna for GPT-5.4 mini; check saved defaults, workspace-managed settings and automations. The retirement does not affect the OpenAI API or Codex use with your own API key.
Current model and availability references are the OpenAI model catalog, rate card and Codex Help Center, accessed September 5, 2026. The examples separate product facts from practical workflow advice.
What changed on August 31, 2026?
OpenAI’s current Codex Help Center guide says that GPT-5.4 and GPT-5.4 mini stopped being available in Codex for ChatGPT-account sign-ins on August 31. Its migration advice is direct:
- Replace GPT-5.4 with GPT-5.6 Terra.
- Replace GPT-5.4 mini with GPT-5.6 Luna.
- Review workspace defaults, saved model choices, managed configuration and automations before the deadline.
- Do not apply this retirement to the OpenAI API or a Codex workflow authenticated with your own API key.
This is a model-availability change, not proof that every GPT-5.4 workflow will produce the same result after a replacement. Keep the task, inputs, reasoning setting and acceptance check fixed while you compare the replacement. If the task is consequential, require a human review step.
Sol, Terra and Luna at a glance
OpenAI’s current API model catalog positions the family as three operating tiers. The labels are useful routing guidance, not an independent ranking of every task.
| Model | OpenAI’s stated role | Current documented rate per 1M tokens | Good starting point for |
|---|---|---|---|
| GPT-5.6 Sol | Flagship model for complex professional work | $4 input / $0.40 cached / $20 output (promotional; available at least through Nov. 21, 2026) | Ambiguous research, difficult coding, strategy and high-cost decisions |
| GPT-5.6 Terra | Balance of intelligence and cost | $2 input / $0.20 cached / $12 output | Everyday analysis, drafting, implementation and structured research |
| GPT-5.6 Luna | Cost-sensitive, high-volume workloads | $0.20 input / $0.02 cached / $1.20 output | Classification, extraction, repetitive transformations and bounded checks |
The current model pages list a 1.05-million-token context window and up to 128,000 output tokens for the three models. Capacity is not a reason to send every file or every previous response. Large context, long tool outputs and repeated history can increase usage and make review harder.
OpenAI’s current Sol, Terra and Luna model pages list reasoning options from none through max. Use the lowest setting that repeatedly passes your acceptance test. More reasoning can help with interacting constraints, but it can also increase time and usage; it is not a truth guarantee.
Current GPT-5.6 pricing: keep the surfaces separate
The current ChatGPT rate card and the live API model pages show the following token-based rates for supported Work and Codex activity:
| Model | Input | Cached input | Output | Stated boundary |
|---|---|---|---|---|
| GPT-5.6 Sol | $4.00 | $0.40 | $20.00 | Promotional pricing available at least through Nov. 21, 2026 |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 | Current rate shown in model and rate-card pages |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | Current rate shown in model and rate-card pages |
These are token-based rates for the documented billing surfaces, not a promise that a ChatGPT subscription is priced per token. Included plan usage, 5-hour and weekly limits, legacy credit meters and enterprise billing arrangements can follow different rules. For the detailed economics, long-context treatment and Fast mode, see the GPT-5.6 pricing guide.
OpenAI’s July 30 announcement said Sol pricing was unchanged at that time. An August 21 update to the GPT-5.6 page and the current rate card now show a temporary Sol price reduction to $4 input and $20 output. Keep both dates in the historical record; use the current Sol model page and rate card for a current budget.
For requests above 272,000 input tokens, the current Sol model page says a higher long-context multiplier applies to the full request. Check the applicable model and billing page before estimating a large job. A lower per-token price does not automatically make a workflow cheaper if it sends duplicated context, retries or unreviewed output.
Should you also consider GPT-6 Astra?
GPT-6 Astra is another option for difficult reasoning, coding, research and computer-use work. GPT-5.6 Sol, Terra and Luna remain practical choices for complex, balanced and high-volume work. OpenAI’s current Astra model page documents a 1.05-million-token context window, up to 128,000 output tokens, and reasoning levels from low through max. OpenAI says rollout starts with enterprises in its Trusted Access Program, with API and paid ChatGPT availability coming in the following days; access is not a promise of universal availability.
| Use case | Starting model | Decision cue |
|---|---|---|
| Complex research, coding or computer-use work | GPT-6 Astra | Test its additional capability against its higher price and access boundary |
| Ambiguous professional reasoning | GPT-5.6 Sol | Use the established Sol routing and acceptance test on this page |
| Routine synthesis and implementation | GPT-5.6 Terra | Balance quality and cost with a representative sample |
| High-volume, mechanically testable work | GPT-5.6 Luna | Use validation, rejection and escalation controls |
Astra’s standard API price is $10 input / $1 cached input / $12.50 cache write / $50 output per million tokens; above 272,000 input tokens OpenAI documents 2× input/cache and 1.5× output pricing for the full request, with Batch/Flex at 50% and Fast at 2× where available. See the dedicated Astra pricing guide for Astra estimates and the GPT-5.6 pricing guide for Sol, Terra and Luna economics. OpenAI’s safety overview reports internal evaluation results and a higher cyber capability classification; treat those as attributed company evidence, not an independent benchmark. Keep the same prompt contract, sources, tools, effort setting and acceptance test when comparing Astra with a GPT-5.6 route.
ChatGPT Chat is not Work, Codex or the API
OpenAI’s August 6 product update changed the ChatGPT Chat experience: Plus and Pro users received an updated GPT-5.6 Sol and a slider for how much thought a response receives, while Free and Go users receive GPT-5.6 Luna as the default with a Think control. OpenAI explicitly said that the release applies to ChatGPT Chat and does not change Work or Codex.
Use this surface map when documenting a workflow:
| Surface | Use it for | Do not infer |
|---|---|---|
| ChatGPT Chat | Conversation, quick research, writing, planning and the ChatGPT reasoning control | That a Chat rollout changed Codex or API behavior |
| ChatGPT Work | Connected tools, longer deliverables and workspace-controlled work | That every model or rate is available in every workspace |
| Codex | Repository work, local files, tools, tests and agentic coding | That ChatGPT Chat’s model picker or limit rules apply unchanged |
| OpenAI API | Programmable routing, explicit token accounting and repeatable evaluations | That an API rate or model retirement in ChatGPT-account Codex affects API-key traffic |
OpenAI’s current Codex guidance says Codex, ChatGPT Work, ChatGPT for Excel and Workspace Agents can share an agentic allowance and credit pool when those features are available on a plan. Usage varies with model, surface, task complexity, context, reasoning, speed and tools. Check Settings or the usage dashboard for the allowance and reset information that applies to your account.
Which GPT-5.6 model should marketers use?
Start with the cost of an incorrect result and the shape of the work. Then route only the step that needs more intelligence upward.
Choose Luna for repeatable, testable work
Luna is a sensible starting point for high-volume transformations with a clear validation rule:
- Classify search queries, pages, leads or creative variants.
- Extract fields from a known document format.
- Rewrite approved copy into channel-length variants without changing claims.
- Compare metadata, links or structured fields and flag exceptions.
- Run bounded implementation or test-writing steps after the design is settled.
Do not call a result “safe” merely because Luna is inexpensive. Add schema validation, a sample review, a rejection path and escalation for ambiguous or policy-sensitive cases.
Choose Terra for ordinary professional synthesis
Terra fits briefs, routine research synthesis, implementation plans, first drafts and analysis where the facts are available but several pieces must be combined. Give it the source set, the audience, the output shape and the checks that define acceptable work. If the evidence conflicts or the downside of a wrong conclusion is high, escalate the judgment step to Sol or a human.
Choose Sol for the reasoning bottleneck
Sol is the appropriate starting point when the task is ambiguous, cross-system, architecture-heavy, security-sensitive or expensive to get wrong. Examples include reconciling conflicting Search Console and analytics evidence, designing a migration with rollback conditions, evaluating a policy-sensitive campaign decision, or synthesizing a large source set into one defensible recommendation.
Sol is not automatically the best choice for every paragraph, classification or routine tool call. A useful pattern is Sol for the argument or decision, Terra for supporting synthesis, and Luna for deterministic checks. The cost-per-accepted-result framework is a useful companion when comparing those routes. The right split depends on a representative evaluation, not a prestige label.
How to choose reasoning effort
Use effort as a budget tied to risk and verification:
| Effort starting point | Use when | Required check |
|---|---|---|
| None or low | Direct transformations, short rewrites and simple extraction | Schema, format or spot-check validation |
| Medium | Normal research, briefs, drafting and bounded implementation | Source links, requirements and targeted review |
| High | Several interacting constraints, difficult debugging or strategy | Evidence reconciliation and explicit acceptance criteria |
| Xhigh or max | High-value work where additional exploration can be evaluated | Adversarial review, calculations and human sign-off where needed |
If an answer is weak because the goal, sources or completion condition is unclear, increasing effort may only produce a longer weak answer. Improve the work contract first:
- Goal: state the decision or artifact the reader needs.
- Context: provide the source set, current page, constraints and prior decisions that matter.
- Output: specify format, audience, depth and required links or tables.
- Boundaries: name actions that require approval and claims that must not be invented.
- Completion check: define the test, source audit or read-back that proves the work is finished.
After the August 31 change: verification checklist
- Inventory model references. Search workspace defaults, saved choices, managed configuration, automation prompts, scripts and documentation for GPT-5.4 and GPT-5.4 mini.
- Map each reference to a replacement. Start with Terra for GPT-5.4 and Luna for GPT-5.4 mini, then consider Sol if the task’s error cost requires it.
- Preserve the acceptance test. Use the same representative inputs, tools, context and review rubric before and after the change.
- Check the surface. Confirm whether the workflow is ChatGPT Chat, Work, Codex or API-key traffic. The August 31 retirement applies to Codex sign-ins with a ChatGPT account, not the API or own-key Codex.
- Review managed controls. Workspace policies can limit which models members can select. A saved default cannot make an unavailable model available.
- Check the usage panel. After migration, review the account or workspace usage page, credits and reset information that applies to the plan.
- Record the change. Note the date, replacement, reasoning effort, representative test and any quality or latency difference. Do not call a migration successful without evidence.
Marketing workflows that fit the model tiers
SEO research and content production
Use Luna to normalize Search Console exports, classify queries and check metadata. Use Terra to extract claims and create a source-backed brief. Use Sol to adjudicate search intent, duplicate ownership, conflicting evidence and the final argument. Keep the publication gate human-controlled, and link the final article to its sources. The Codex usage guide covers related context and allowance controls.
Paid-media operations
Luna can label search terms, creative variants and recurring report fields. Terra can summarize a stable performance export. Sol is better suited to a diagnosis where attribution, conversion lag, budget constraints and tracking quality point in different directions. Keep exclusions, budget changes and launches behind explicit approval.
Analytics and reporting
Start with Luna for mechanical cleanup and Terra for metric definitions and narrative structure. Escalate to Sol when definitions conflict or several causal explanations compete. Report missing data, lag and uncertainty instead of turning a partial export into a confident causal story.
Connected Work and scheduled workflows
For connected tools and recurring tasks, model choice is only one control. Review permissions, source freshness, deduplication, read-back, notification scope and the kill switch. The ChatGPT Work scheduled-tasks guide covers the workflow boundary, while the ChatGPT Work and Codex workflow guide covers the adjacent tool and workspace architecture; do not assume a model replacement changes those permissions.
Common GPT-5.6 prompting mistakes
- Using Sol for every clerical step because it is the flagship.
- Routing to Luna without a labelled sample, rejection path or drift check.
- Assuming the ChatGPT August 6 release changed Work, Codex or the API.
- Carrying the old GPT-5.4 model name into saved defaults after August 31.
- Using the old $5/$30 Sol rate after OpenAI’s August 21 price update.
- Converting a token price into a fixed number of ChatGPT messages.
- Uploading a large history when a compact source extract would work.
- Publishing fluent output without checking dates, links, calculations and claims.
Frequently asked questions
Which GPT-5.6 model is best overall?
OpenAI positions Sol as the flagship for complex professional work, Terra as the balance of intelligence and cost, and Luna for cost-sensitive high-volume work. “Best” still depends on the task, error cost, latency and acceptance test.
What is the current GPT-5.6 Sol price?
The current model page and ChatGPT rate card show $4 per million input tokens, $0.40 per million cached input tokens and $20 per million output tokens for the stated promotional period. OpenAI says that promotion is available at least through November 21, 2026. Check the applicable billing surface before budgeting.
Are GPT-5.4 and GPT-5.4 mini being retired everywhere?
No. OpenAI’s current Help Center says they leave Codex for users signed in with a ChatGPT account on August 31, 2026. It says the change does not affect the OpenAI API or Codex use with your own API key.
What should replace GPT-5.4 in Codex?
OpenAI’s documented replacement is GPT-5.6 Terra. It documents GPT-5.6 Luna as the replacement for GPT-5.4 mini. Run a representative comparison before changing a high-stakes workflow.
Does a higher reasoning setting guarantee a better answer?
No. Higher effort can give a difficult task more room to explore, but it can add latency and usage. Define the acceptance test and verify the output against the source or system of record.
Do ChatGPT, Work and Codex use the same model picker?
No. Availability, workspace controls, limits and billing can differ by surface and plan. Check the current product documentation and the picker or usage panel for the account you are actually using.
Bottom line
GPT-5.6 model selection is a routing decision. Use Sol where ambiguity and error cost justify it, Terra for ordinary professional synthesis, and Luna for repeatable volume with a measurable check. Keep ChatGPT Chat, Work, Codex and API facts separate. After August 31, use the documented GPT-5.6 replacement routes for GPT-5.4 and GPT-5.4 mini in ChatGPT-account Codex workflows, then run the same acceptance test and record what changed. Current pricing and availability can move again, so date your documentation and check the official model and Help Center pages before making a production decision.
Current GPT-5.6 model and availability references: Sol, Terra, Luna and the Codex Help Center, accessed September 5, 2026.
