If GPT-5.4 disappeared from the ChatGPT model picker, it is tempting to assume that OpenAI shut it down everywhere. That is not the case. At the time of writing, gpt-5.4 remains available in the API, is listed in the official model catalog, and has no published API retirement date.
The practical question is still worth asking: which GPT-5.6 model should replace it for coding? The answer changes depending on whether you pay for ChatGPT Plus or call the API directly. GPT-5.6 is a family, not one model, and Sol, Terra, and Luna make different trade-offs between capability and cost.
This post compares the whole GPT-5.6 family with GPT-5.4, including the available benchmark results, current API prices, and what Plus subscribers can actually select.
GPT-5.4 Is Still an API Model
OpenAI released GPT-5.4 in March 2026 as a general-purpose reasoning model with strong coding, computer-use, and tool-use capabilities. It supports text and image input, a 1.05M-token context window, 128K output tokens, reasoning, structured outputs, function calling, web search, MCP, hosted shell, and computer use.
The current API model page lists the gpt-5.4-2026-03-05 snapshot. OpenAI’s deprecation page does not list GPT-5.4, so there is no official basis to say that API support has ended.
That distinction matters. ChatGPT model availability, Codex availability, and API availability are separate product decisions. A model that is no longer prominent in the ChatGPT interface can remain a supported API model for existing applications.
The GPT-5.6 Family Explained
GPT-5.6 has three API model tiers. All three support text and image input, a 1.05M-token context window, up to 128K output tokens, reasoning, structured outputs, function calling, prompt caching, web search, MCP, computer use, hosted shell, and other Responses API tools.
For applications currently using the unsuffixed gpt-5.6 alias, OpenAI routes that alias to GPT-5.6 Sol. Use a named model ID when you need an explicit cost and capability tier.
| Model | OpenAI positioning | Best fit for coding |
|---|---|---|
| GPT-5.6 Sol | Flagship model for complex professional work | Ambiguous bugs, architecture, difficult refactors, agentic work |
| GPT-5.6 Terra | Balance of intelligence and cost | Everyday production coding and the closest GPT-5.4 cost replacement |
| GPT-5.6 Luna | Cost-sensitive, high-volume model | Well-specified changes, tests, extraction, routing, and background tasks |
Each API model supports none, low, medium, high, xhigh, and max reasoning effort. More reasoning can improve difficult tasks, but it also increases latency and token consumption. Model tier and reasoning effort should be chosen together rather than treating every coding task as a flagship-model task.
What about GPT-5.6 Sol Pro?
GPT-5.6 Sol Pro is a ChatGPT option for difficult, longer-running work on eligible Pro, Business, and Enterprise plans. It is not one of the three current API model tiers listed in OpenAI’s API model catalog. For an API migration from GPT-5.4, the concrete choices are Sol, Terra, and Luna.
For ChatGPT Plus users, Sol powers Instant and is available at Medium and High reasoning. Plus does not include Sol Pro or Extra High. In Codex and ChatGPT Work, Plus users can choose Sol, Terra, and Luna. Terra and Luna are not selectable in standard ChatGPT conversations, so this distinction is easy to miss.
Coding Benchmarks: GPT-5.4 vs GPT-5.6
Both launch pages publish results for SWE-Bench Pro, which evaluates issue resolution in real software repositories. They are useful directional evidence, not a controlled head-to-head: release-time reasoning settings, agent harnesses, tools, and benchmark revisions can affect scores.
| Model | SWE-Bench Pro |
|---|---|
| GPT-5.4 | 57.7% |
| GPT-5.6 Sol | 64.6% |
| GPT-5.6 Terra | 63.4% |
| GPT-5.6 Luna | 62.7% |
On the published figures, every GPT-5.6 tier exceeds GPT-5.4 on SWE-Bench Pro. Terra is 5.7 percentage points above GPT-5.4 while costing less per API token, which makes it the most interesting replacement for many production coding workloads.
OpenAI also reports the following GPT-5.6 results for coding-agent evaluations:
| Benchmark | Sol | Terra | Luna |
|---|---|---|---|
| Artificial Analysis Coding Agent Index v1.1 | 80.0 | 77.4 | 74.6 |
| DeepSWE v1.1 | 72.7% | 69.6% | 67.2% |
| Terminal-Bench 2.1 | 88.8% | 87.4% | 84.7% |
GPT-5.4’s launch reported 75.1% on Terminal-Bench 2.0, rather than Terminal-Bench 2.1, and did not report DeepSWE v1.1 or the Artificial Analysis Coding Agent Index v1.1. Those numbers should not be presented as a controlled head-to-head comparison.
The safe conclusion is narrower: GPT-5.6 has stronger published coding-agent results, and it beats GPT-5.4 on the shared SWE-Bench Pro metric. Validate the result on your own repository before treating any public benchmark as a migration guarantee.
API Pricing: Which Model Matches GPT-5.4’s Cost?
These are current Standard API prices per 1M tokens for requests under 272K input tokens.
| Model | Input | Cached input | Output | Compared with GPT-5.4 |
|---|---|---|---|---|
| GPT-5.4 | $2.50 | $0.25 | $15.00 | Baseline |
| GPT-5.6 Sol | $4.00 | $0.40 | $20.00 | 60% more input, 33% more output |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 | 20% less across input, cache reads, and output |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | 92% less across input, cache reads, and output |
Sol’s current price is promotional and OpenAI says it is available at least through November 21, 2026. Do not build a long-term budget that assumes a temporary price without monitoring the pricing page.
Cache Writes and Long Context Change the Calculation
GPT-5.4 charges for uncached input, cached input, and output. GPT-5.6 adds a cache-write charge equal to 1.25 times its uncached input price, while cache reads retain the 90% discount.
For one-off requests, cache writes do not create value. For multi-turn coding agents that reuse a large stable prefix, prompt caching can still reduce the effective input cost dramatically. Measure cache-hit rates from real traffic before concluding that a nominal per-token price is the total cost of a workflow.
All four models support around 1M tokens of context. For GPT-5.4, Sol, Terra, and Luna, requests above 272K input tokens use long-context pricing: 2x input and 1.5x output for the full request. Feeding an entire repository into every call can therefore erase a model’s apparent price advantage. Retrieval, compaction, and smaller task scopes matter as much as model selection.
ChatGPT Plus Is Not API Billing
ChatGPT Plus costs $20 per month and includes Sol, Terra, and Luna in Codex and ChatGPT Work. It does not include Sol Pro, unlimited model usage, or API calls. API keys are billed separately at API rates.
Usage limits depend on task complexity, context length, tool use, model, and current capacity. OpenAI publishes estimates rather than fixed guarantees, and those estimates can change. A large coding task with many files, tool calls, and long-lived context can consume much more allowance than a short local edit.
For Plus subscribers, a practical routing strategy is:
- Use Sol in ChatGPT for ambiguous debugging, system design, and difficult multi-file work.
- Use Terra in Codex or ChatGPT Work for routine implementation, code review, test writing, and bounded refactors.
- Use Luna in Codex or ChatGPT Work for repetitive, well-specified tasks such as generating tests, formatting structured output, or processing high-volume background work.
This is not about making every task cheaper. It is about reserving the most constrained capability for work where additional reasoning materially changes the outcome.
Which GPT-5.6 Model Should Replace GPT-5.4?
You used GPT-5.4 in ChatGPT Plus
Choose GPT-5.6 Sol for the closest capability-oriented replacement. It is the paid ChatGPT model for complex work. When you work in Codex or ChatGPT Work, add Terra and Luna to your workflow instead of spending Sol usage on every small task.
You used GPT-5.4 through the API and want a similar bill
Choose GPT-5.6 Terra first. Its Standard price is 20% lower than GPT-5.4 for input, cached input, and output, while its published SWE-Bench Pro score is higher. It also retains the same broad class of API tools and long-context capacity.
Run an evaluation before changing production traffic. Test your actual prompts, tools, repositories, structured-output schemas, latency requirements, and cost per successfully completed task.
You used GPT-5.4 for the hardest coding tasks
Choose GPT-5.6 Sol. It costs more per token than GPT-5.4, but it is the flagship tier and has the strongest published coding results in the GPT-5.6 family. It is the better candidate for failures that are expensive: security-sensitive changes, architectural migrations, vague bug reports, and long agentic runs.
You used GPT-5.4 for high-volume, well-defined automation
Evaluate GPT-5.6 Luna. It is not a like-for-like capability replacement for GPT-5.4, but its price makes new routing strategies possible. Use a stronger model to plan or resolve uncertainty, then let Luna implement clearly specified changes, run standard checks, or classify the results.
A Safe Migration Checklist
- Keep the GPT-5.4 snapshot as your control during evaluation rather than switching everything at once.
- Create a representative test set: easy changes, regressions, multi-file refactors, tool calls, and failures that matter to your team.
- Measure task success, human rework, latency, total input and output tokens, cache-hit rate, and tool-call failures.
- Test Terra first if GPT-5.4 cost was acceptable but budget matters; test Sol first if reliability matters more than the token rate.
- Roll out gradually and keep a fallback until the new model meets your quality threshold.
Model migrations are engineering changes, not simple string replacements. Public benchmarks are useful signals, but the only benchmark that determines whether GPT-5.6 replaces GPT-5.4 in your workflow is your own workload.
The Bottom Line
GPT-5.4 remains available in the OpenAI API, but GPT-5.6 offers clearer choices for developers who want to move forward.
- ChatGPT Plus: use Sol as the main replacement, then route routine work to Terra and Luna in Codex or ChatGPT Work.
- API, similar or lower cost: start with Terra.
- API, maximum coding capability: use Sol.
- API, high-volume and well-defined work: evaluate Luna as part of a routed workflow.
If you are also comparing prior GPT-5.6 pricing with GPT-5.5, see GPT-5.5 vs GPT-5.6 for Coding. For a cost comparison with an alternative coding model, see Kimi K2.7 Code vs GPT-5.4 vs GPT-5.5.