OpenAI 2026 API Pricing: GPT-5.6, Sora-2, and Service Updates

This guide outlines the 2026 pricing structure for OpenAI's API, featuring the new GPT-5.6 model series and Sora-2 video generation. It details costs across various modalities and processing modes, including Batch and Fast mode options. The documentation also highlights the phase-out of the fine-tuning platform and new surcharges for regional data residency.

Key Points

• Introduction of the GPT-5.6 model family (Sol, Terra, Luna) with distinct pricing for input, output, and cached tokens across four processing tiers.
• Launch of multimodal API pricing for Sora-2 video generation (per second) and GPT-realtime-2.1 audio and text processing.
• Implementation of a 10% pricing surcharge for regional data residency endpoints for models released on or after March 5, 2026.
• Announcement that the fine-tuning platform is being phased out, remaining accessible only to existing users for a limited time.
• Standardized tool pricing for web search ($10 per 1k calls) and hosted container sessions (billed per 20-minute session).

Sentiment

Analytical and skeptical, with a focus on the economic implications of AI commoditization and the sustainability of the current market.

In Agreement

• The price reductions make OpenAI's flagship models significantly more competitive against Anthropic and other providers.
• AI intelligence is becoming a commodity, leading to a 'race to the bottom' where pricing will eventually be as cheap as electricity.
• Increased competition is preventing a stagnant duopoly between OpenAI and Anthropic.
• The planetary naming scheme (Sol, Terra, Luna) is a creative and relatively intuitive way to represent model scales.

Opposed

• Temporary price reductions are of little use for production workloads that require long-term cost predictability.
• The planetary naming scheme is confusing and less effective than standard 'Small/Medium/Large' labels.
• Current 'open weight' models are not truly open source because they remain opaque binary blobs.
• Frontier labs are not 'greedy' but are actually losing massive amounts of money, suggesting the current pricing is unsustainable.
• Model naming and reasoning effort combinations are becoming too complex for users to easily navigate the Pareto frontier.