benchr
OpenAI's newest models are GPT-6 Astra, released September 3, 2026, and the GPT-5.6 family: Sol, Terra, and Luna. GPT-5.5, GPT-5.4, GPT-5, and GPT-5 Mini remain on OpenAI's price list as older models. The table below reads current rates from the shared benchr model data, which carries GPT-6 Astra and GPT-5.6 Sol (listed as GPT-5.6) but not yet Terra or Luna; their rates are in the GPT-5.6 pricing breakdown.
| Model ID | Input / 1M | Output / 1M | Cached Input / 1M | Batch Input / 1M |
|---|---|---|---|---|
| gpt-5.5 | $5.00 | $30.00 | $0.500 | $2.50 |
| gpt-5.4 | $2.50 | $15.00 | $0.250 | — |
| gpt-5 | $1.25 | $10.00 | — | — |
| gpt-5-mini | $0.250 | $2.00 | $0.025 | — |
Pricing Strategy: Optimizing OpenAI Costs
If your application relies heavily on OpenAI models, you can employ several tactics to cut your monthly bills:
- Use Batch API for async work: OpenAI lists a 50% discount for eligible Batch API input and output tokens. Jobs use a documented 24-hour completion window, so they suit work that can wait; handle failed, expired, or canceled batches and verify current eligibility.
- Enable Prompt Caching: Dynamic system prompts and repetitive contexts are cached automatically. On the models the table lists with a cached rate, a cache hit bills at a tenth of standard input: $1 against $10 per 1M for GPT-6 Astra, $0.50 against $5 for GPT-5.5, and $0.025 against $0.25 for GPT-5 Mini.
- Use Structured Outputs wisely: A strict JSON schema counts toward input tokens, and first-request schema processing can affect latency. Reuse stable schemas where appropriate, then measure latency and valid-output rate on your workload.
For cross-provider comparisons, read the global AI Model Pricing Comparison, or calculate custom usage via the Cost Calculator.
Where GPT-5.4 fits
The ladder gained a rung in March 2026. GPT-5.4 sits at exactly double GPT-5 and exactly half of GPT-5.5: $2.50/1M input, $15/1M output, cached input at $0.25. What the middle rate buys is context up to 1M tokens (GPT-5 stops at 400K), built-in computer use, and the finance tuning that launched ChatGPT for Excel. It carries the same 272K long-context surcharge structure as GPT-5.5, so the headline rate only holds below that line. The GPT-5.4 pricing breakdown has the full math. Since the GPT-5.6 release, GPT-5.6 Terra ($2 input / $12 output per 1M) sits below GPT-5.4 on both sides with the same 1.05M window, so GPT-5.4 now makes sense mainly for workloads already validated on it.
Budgeting OpenAI workloads
OpenAI pricing decisions usually come down to routing. As of September 2026, benchr's editorial routing is GPT-5.6 Luna as the volume tier ($0.20 / $1.20 per 1M), GPT-5.6 Terra as the general default ($2 / $12), GPT-5.6 Sol for the hardest coding and agent work ($4 / $20; SWE-bench Verified 89.8, OpenAI-reported), and GPT-6 Astra only when the ceiling matters more than paying 2.5× Sol's rate and your stack already uses the Responses API, which Astra requires for tool calling. GPT-5 Mini, GPT-5.4, and GPT-5.5 each have a GPT-5.6 model that costs less on both input and output (Luna, Terra, and Sol); GPT-5 ($1.25 / $10) remains the cheapest full GPT-5-class rate, with a 400K window. Treat those as routing lanes, not as interchangeable chat models.
Before standardizing on one model, separate calls by risk. Use the smaller tier for extraction, classification, and simple drafting; use the middle tier for normal user-facing answers; reserve the flagship tier for tasks where a wrong answer creates enough review cost or product risk to justify the premium.
What to measure after launch
After deployment, track model cost by feature rather than by account total. The important lines are average input tokens, average output tokens, retry rate, and fallback rate per feature. A chat feature, an extraction feature, and a coding feature can have completely different economics even when they use the same OpenAI model.
Review the routing rules monthly. If GPT-5.6 Luna starts passing the same checks as GPT-5.6 Terra for a narrow workflow, move that workflow down. If Terra produces expensive review failures in a high-risk area, move only that area up to Sol instead of upgrading everything.
For teams already using OpenAI, the cheapest optimization is often prompt discipline rather than a model switch: shorter system messages, structured outputs that prevent verbose answers, and caching repeated instructions where the platform supports it. Clean prompts reduce cost and make evals more stable.
Finally, keep a simple monthly price audit. Provider pricing, batch discounts, and model aliases can change faster than product teams notice. A five-minute check against the official pricing page and your own usage dashboard prevents a quiet model change from becoming a surprise bill.