OpenAI cut standard application programming interface (API) prices for GPT-5.6 Luna by 80 percent and GPT-5.6 Terra by 20 percent on July 30, 2026. The company left GPT-5.6 Sol's standard token rates unchanged.

OpenAI attributed the reductions to lower serving costs and more efficient token generation. Its announcement did not name Chinese competitors, although lower-cost and open-weight models from Chinese developers have widened the choices available to buyers.

Rows of server racks and network cables inside a data center (illustrative image)

Luna now starts at US$0.20 per million input tokens

OpenAI's July 30 announcement set Luna at US$0.20 per million input tokens and US$1.20 per million output tokens, while Terra moved to US$2 and US$12 respectively. Sol remained at US$5 for input and US$30 for output under the comparable standard rate.

OpenAI's current pricing table identifies those figures as standard short-context rates and lists separate prices for cached input, cache writes and long-context use. The page also shows lower Batch and Flex rates, higher Fast mode rates and a 10 percent surcharge for eligible regional-processing endpoints.

For a monthly workload totaling 100 million uncached input tokens and 20 million output tokens, the listed short-context rates produce an estimated token charge of US$44 on Luna, US$440 on Terra or US$1,100 on Sol. That calculation excludes tools, cache writes, regional processing and any requests billed under another service tier.

A computer monitor displaying data charts on an office desk (illustrative image)

OpenAI points to serving efficiency, not Chinese competition

OpenAI said GPT-5.6 Sol helped optimize production kernels that reduced end-to-end serving costs by 20 percent, while related experiments improved token-generation efficiency by more than 15 percent. The company presented those gains as the reason it could lower Luna and Terra prices.

The official announcement did not say the cuts were a response to Moonshot AI, MiniMax, Z.ai or another Chinese company. The South China Morning Post framed the move as a defense against lower-cost Chinese rivals and reported that Luna moved ahead of Z.ai's GLM-5.2 and MiniMax's M3 in Artificial Analysis's intelligence-per-dollar ranking.

Chinese vendors are competing through price and access

Token rates are only one part of the competitive pressure. Some Chinese developers also publish model weights, allowing customers or hosting companies to operate a model without sending every request to the developer's own API.

Moonshot AI's Kimi K3 paper describes a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters and a one-million-token context window, and says the full weights were released. Access to weights does not supply the hardware or engineering needed to run a model of that size, so “free weights” and “free inference” are different propositions.

MiniMax introduced M3 on June 1, 2026, with a one-million-token context window and API access, and said it planned to release the corresponding weights. The company also published its own coding and agent benchmarks, but vendor-run results need comparison with independent tests and the buyer's actual workload.

A computer chip and circuit board shown in close-up (illustrative image)

Price per token is not price per completed task

A cheaper token lowers the bill only if the model completes the work at an acceptable quality level. Longer outputs, retries, failed tool calls and extra review can offset a lower posted rate, while prompt caching can reduce the cost of repeated context.

Comparisons therefore need a fixed test set and a common success threshold. Useful measurements include completion rate, input and output volume, response time, retry frequency, human review time and total cost for each accepted result.

Model routing can reduce spending when routine work goes to a lower-cost model and difficult cases move to a more capable one. The saving depends on the routing error rate: an inexpensive first attempt adds cost if most requests must be repeated elsewhere.

API and open-weight deployments carry different costs

A hosted API turns model operation into a usage charge and leaves capacity planning to the provider. A self-hosted model replaces some of that charge with servers or cloud accelerators, deployment engineers, monitoring, security controls, upgrades and idle capacity.

Data handling also differs by provider, hosting location and customer configuration. Open weights can offer more control over where requests are processed, but the operator then assumes responsibility for access controls, logs, model updates and the rules that apply in each country where the service runs.

A team discussing laptops around an office meeting table (illustrative image)

A four-part check before changing models

  • Token profile: Forecast uncached input, cached input, cache writes and output separately, then apply the correct context length and service tier.
  • Task quality: Run the same representative cases against each candidate and count accepted results rather than relying on a composite benchmark alone.
  • Operating conditions: Measure latency, rate limits, outages, tool charges, regional processing and the staff time needed to review failures.
  • Switching path: Keep prompts, evaluations and application interfaces portable enough to test another provider without rebuilding the product around one model.

The July price cut changes the baseline for high-volume GPT-5.6 workloads, especially where Luna meets the required quality threshold. It does not establish that one model is cheapest for every task or that an open-weight alternative will cost less after infrastructure and operations are included.

Frequently asked questions

Did OpenAI say Chinese rivals caused the price cuts?
No. OpenAI's July 30 announcement credited serving and token-generation efficiencies; the link to Chinese competition is an interpretation reported by the South China Morning Post.

Did ChatGPT or Codex subscription prices fall?
No. OpenAI said subscription prices and quota budgets stayed unchanged, while Terra and Luna began consuming fewer credits in ChatGPT Work and Codex.

Does open-weight mean a model is free to run?
No. Weight access can remove dependence on the developer's hosted API, but computing, electricity, deployment, monitoring and security still carry costs.

Which figure matters most when comparing models?
No single figure settles the choice. Total cost per accepted result combines token rates with output length, retries, latency, review work and the share of tasks that meet the required quality level.