OpenAI's current API model lineup
OpenAI's current generation is GPT-5.6, sold across three tiers rather than a single product. At the top sits Sol, priced as the most expensive model in the lineup on both input and output tokens (the bare "gpt-5.6" model id routes here), the choice when you need the strongest available reasoning and quality. Below it, Terra offers a lower rate for general-purpose work, with Luna stepping down further for high-volume, low-complexity tasks. GPT-5.5, the prior-generation flagship, stays priced and available for teams that haven't re-validated against 5.6 yet. Alongside all of that, OpenAI keeps GPT-4.1 available specifically as a long-context option; it supports up to a 1-million-token context window and is positioned as the cheaper route for very large prompts, even though it isn't the newest release.
Where reasoning-model pricing went
If you're used to pricing a dedicated reasoning model like o3 separately from a general-purpose one, that split no longer exists at OpenAI. The standalone o-series, o1, o3, o4-mini, was retired from ChatGPT in February 2026 and folded into the GPT-5.x generation as a "Thinking" tier: turn up reasoning effort on GPT-5.6 Sol and you get the multi-step reasoning behavior the o-series used to provide, billed at that model's normal input/output rate rather than a separate SKU. Practically, this means there's one fewer axis to shop on. You're choosing a model tier (Sol down to Luna) and a reasoning-effort setting, not a model tier and a separate reasoning-versus-general-purpose model family.
How OpenAI's pricing structure works
Every model bills input tokens (what you send) and output tokens (what the model generates) at separate per-million-token rates, and output is consistently priced higher than input across the whole lineup. Generation is more compute-intensive than reading a prompt. The spread between input and output rate is fairly wide on the flagship models, which means a task that produces long answers (drafting, summarizing at length, code generation) is disproportionately sensitive to the output rate, while a task that mostly reads a large prompt and returns a short answer (classification, extraction, routing) is dominated by the input rate. Reasoning-heavy calls with effort turned up tend to generate substantially more output tokens before the final answer, so they lean further toward output-rate-sensitive than a typical direct-answer call on the same model, worth accounting for separately from the base per-token rate. GPT-5.6 also prices cache reads at 90% off the standard input rate and cache writes at 1.25x input, on top of the base rate and the 272K-token long-context surcharge (2x input / 1.5x output above that threshold) it inherits from GPT-5.5's structure. Use the calculator above with your actual average prompt and completion lengths, not round numbers, since the ratio of input to output tokens in your real traffic decides which rate matters more to your bill.
Model-lineup quirks worth knowing
Two things stand out in how OpenAI has structured this lineup. First, the Sol/Terra/Luna naming replaces the previous generation's "mini"/"nano" suffixes, but the price gap between tiers is still large, so if your product routes a mix of simple and complex requests through Sol by default, you're very likely overpaying on the simple half of that mix; route it to Terra or Luna instead. Second, GPT-4.1's role as a dedicated long-context model means it can beat newer, nominally more advanced models on cost for large-document tasks specifically, even though it isn't the newest release; pricing and capability don't move in lockstep with model generation, so it's worth checking GPT-4.1 against GPT-5.6 Terra or Luna whenever your prompts run long.
A worked example on GPT-5.6 Sol
At Sol's standard rate ($5 input / $30 output per million tokens, below the 272K-token threshold), a 1,000-token prompt with a 500-token answer costs 1,000 ÷ 1,000,000 × $5 = $0.005 for input, plus 500 ÷ 1,000,000 × $30 = $0.015 for output, $0.02 per call, or $2,000 a month at 100,000 calls. Route the same volume to Luna instead ($1 input / $6 output) and the bill drops to 1,000 ÷ 1,000,000 × $1 = $0.001 plus 500 ÷ 1,000,000 × $6 = $0.003, $0.004 per call, $400 a month, a 5x reduction for a task that doesn't need Sol's full capability.
Cutting your OpenAI bill
The single biggest lever is model routing: send simple, high-volume requests to GPT-5.6 Luna instead of a flagship model by default, and reserve Sol (with reasoning effort turned up, if the task needs it) for the requests that actually need that level of capability. Beyond routing, cap max output tokens so a verbose response doesn't silently inflate cost (this matters even more once reasoning effort is on, since the model generates more before it answers), trim system prompts and few-shot examples down to what's actually improving output quality, and structure a stable prefix to repeat byte-identical across calls so it qualifies for the 90%-off cache-read rate instead of the full input rate every time. If your workload regularly sends very large prompts, compare GPT-4.1's long-context rate directly against your default model. For that specific shape of traffic it's often the cheaper option even without switching to a smaller model.
Who the OpenAI API suits
OpenAI's spread across five models in active pricing gives you unusually fine-grained control over the cost/capability trade-off within a single provider, which suits teams that want to standardize on one API surface but still tune cost per feature, running chat on Terra, background classification on Luna, reasoning-heavy analysis on Sol with effort turned up, and long-document work on GPT-4.1, all from the same lineup. If you'd rather compare OpenAI's rates directly against Anthropic, Google, or the smaller open-model providers before committing, the all-provider hub calculator puts every vendor's models in the same view.
Compare OpenAI against Anthropic's Claude API pricing, Google's Gemini API pricing, or the budget-focused DeepSeek API pricing calculator.
