DeepSeek and OpenAI launch new models with divergent pricing models

The new model lineups

DeepSeek recently refreshed its lineup with the V4.1-Flash-1017 model, which serves as the foundation for its current flash offerings. This model was designed to be the smallest in a new architecture family, featuring native visual understanding and a 1 million token context window. It was released under the MIT license, which allows for specific types of open-weights usage.

Simultaneously, OpenAI rolled out the GPT-6 family in October 2026. This new generation includes different models tailored for various user levels, such as GPT-6 Sol for high-tier subscribers and GPT-6 Luna for free and lower-cost users. These models replaced the previous GPT-5.6 series across the ChatGPT platform.

The two companies are moving in different directions regarding how they deliver these technologies to the public. DeepSeek primarily focuses on metered API access, which charges users based on exact usage. In contrast, OpenAI continues to lean heavily on a subscription model for its chat interface while offering a separate API for developers.

Comparing the costs

Pricing for these models varies significantly based on the specific model chosen and the timing of the request. DeepSeek uses a peak and off-peak structure where off-peak rates are set at half the cost of peak rates. For example, DeepSeek Flash costs $0.15 per million input tokens during off-peak windows but rises to $0.30 during peak times.

OpenAI's pricing is more complex because it depends on the specific tier of the GPT-6 family. The GPT-6 Luna model actually lists a lower price of $0.10 per million input tokens than DeepSeek's cheapest route. However, the high-end GPT-6 Astra model sits in a different league entirely, with an output rate of $50.00 per million tokens.

The cost of data retrieval also plays a major role in the final bill for heavy users. DeepSeek offers a very low off-peak cached input rate of $0.003 per million tokens. OpenAI's cached rate for the Terra model is $0.05, which is significantly higher but still offers a discount over fresh inputs.

Workload impact

The difference in pricing becomes most apparent when looking at specific types of automated workloads. A small support-triage bot using DeepSeek Flash at peak rates might cost around $5.40 a month. That same workload on OpenAI's Terra model would cost roughly $20.00, which is the price of a single ChatGPT Plus subscription.

For high-volume tasks like agent loops, the impact of caching becomes the deciding factor for the budget. In a scenario with 40 million cached input tokens, DeepSeek's off-peak costs are nearly negligible. On the Terra schedule, that same cached portion would cost $2.00, showing how different architectures handle repeated data.

Heavy output generation, such as creating long reports or complex code diffs, tends to narrow the benefits of cheap input prices. Because output tokens are much more expensive across both platforms, the total cost of a task is often dominated by the final generation rather than the initial prompt. This means teams must balance the quality of the output against the volume of the input.

Privacy and hosting

A major distinction between the two providers is where the data actually resides during processing. Because DeepSeek's V4.1-Flash weights are shipped under the MIT license, a team with the necessary hardware can host the model themselves. This allows them to keep all prompts and data entirely off third-party servers.

OpenAI does not offer a self-hosting option for its GPT-6 models, meaning users must rely on their infrastructure. This creates a trade-off between the ease of using a managed service and the privacy of a self-hosted solution. For many enterprises, the ability to maintain complete control over data flow is a primary requirement for adoption.

Furthermore, the terms of service and billing changes can affect how these models are used in production. DeepSeek recently implemented a billing change that takes effect on November 1. Users are encouraged to check their specific dashboards to confirm the current rates, as some early snapshots of the platform showed different figures than the current live pricing.

Evaluating performance

While both companies provide their own benchmarks, there is currently no verified head-to-head table for the newest models. DeepSeek reported that its 1017 update achieved a 77% Pass@1 rate on the DeepSWE v1.1 test. This represents a slight improvement over the previous September version, which scored around 74.2%.

Experts suggest that public benchmarks can be misleading because they often depend on the specific framework used to run the tests. A model might perform well on one set of tasks but struggle with others, such as structured extraction or image reading. The quality of a model is often better measured by how well it handles a specific, real-world task rather than a general score.

To find the best fit, teams are encouraged to run private tests using their own anonymized data. By running a set of tasks through both APIs, users can record the actual pass rate, token usage, and latency. This method provides a clear cost-per-passed-task metric that can be used to justify budget decisions to stakeholders.