OpenAI Cuts GPT-5.6 API Prices — Luna Down 80%, Now Undercuts DeepSeek on Long-Context Workloads

OpenAI cut GPT-5.6 Luna API pricing by 80% to $0.20/$1.20 per million tokens, and Terra by 20% to $2/$12. Luna now undercuts DeepSeek V4 Pro on input-heavy long-context workloads. Fast mode for Sol replaces Priority Processing. Sol-driven kernel work cut serving cost 20%.

GPT-5.6 Luna is now $0.20 per million input tokens and $1.20 per million output tokens — an 80% cut from its launch price. GPT-5.6 Terra dropped 20% to $2 input and $12 output per million tokens. Sol pricing stayed put, but Sol now ships with a Fast mode that runs up to 2.5× faster than Standard at 2× the price, and Sol-driven kernel work already cut OpenAI’s serving cost by 20%. Effective July 30, 2026. (OpenAI announcement)

The new price table

Model Old input / M Old output / M New input / M New output / M Cut
GPT-5.6 Luna $1.00 $6.00 $0.20 $1.20 80%
GPT-5.6 Terra $2.50 $15.00 $2.00 $12.00 20%
GPT-5.6 Sol unchanged unchanged unchanged unchanged —

Sources: OpenAI (primary), VentureBeat, BasicTutorials, Axios, and aipricing.guru’s tracker. Old Sol numbers stayed flat; Sol’s own price wasn’t touched, but its economics changed anyway (see below).

Fast mode replaces Priority Processing

The Priority Processing tier in the API is gone. In its place: Fast mode for GPT-5.6 Sol. Up to 2.5× faster than Standard processing, at roughly 2× the price, with no change in model intelligence. Requests tagged priority auto-route to Fast mode, so existing code doesn’t break — backward-compatible. (OpenAI)

The Fast mode lines up with /fast in Codex, so the routing story is consistent across products. If you were using Priority Processing before, your bill roughly doubles per token and your latency roughly halves. That’s the trade, spelled out.

Luna vs DeepSeek V4 Pro on long-context workloads

This is where the price war gets interesting. Luna at $0.20/M input now beats DeepSeek V4 Pro on input-heavy long-context workloads. The math from XenoSpectrum for a 1M-input, 100K-output monthly bill on requests under 272K tokens:

  • Luna: 1,000,000 × $0.20 + 100,000 × $1.20 = $0.32
  • DeepSeek V4 Pro: 1,000,000 × $0.40 + 100,000 × $1.22 = $0.522

Luna comes in at roughly 61% of DeepSeek’s cost on that workload. Above 272K tokens the comparison gets murkier, and DeepSeek still wins on some price tiers — but the headline is real: on input-heavy, sub-272K-token work, Luna is now the cheaper choice.

The competitive response was immediate. DeepSeek launched V4 within hours of the Luna cut. (36kr)

The honest caveat — Luna is not the cheapest pure-token model

VentureBeat’s coverage is worth quoting directly: Luna is not the cheapest model in the broader market on a pure token basis. Xiaomi’s MiMo-V2.5 Flash, DeepSeek’s flash tier, and several other APIs remain less expensive on raw per-token pricing. (VentureBeat)

Luna’s win is in the combination of price and capability tier. It’s still a GPT-5.6-class model at the price of a flash model, which is what the 80% cut was about — not putting Luna below every flash tier, but making the price-to-capability ratio structurally different from what DeepSeek charges for comparable intelligence.

OpenAI’s own benchmark numbers

Three claims from OpenAI’s announcement worth flagging because they come from OpenAI itself, not independent third parties:

  • “Luna delivers performance comparable to models that were frontier-class a year ago at roughly 6 cents on the dollar per task, and at nearly 9× the speed.”
  • “On professional work, as measured by Agents’ Last Exam, Luna outperforms Fable 5 at an estimated cost per task nearly 99% lower.”
  • “GPT-5.6 Sol autonomously rewrote and optimized production kernels — kernel work helped reduce end-to-end cost of serving the model by 20%. Sol experiments increased token-generation efficiency by more than 15%.” (OpenAI)

The Sol story is the underrated part of this announcement. Sol didn’t get a price cut; it got a 20% serving-cost reduction from its own kernel work. The model is now cheaper to run, which is what funds Luna’s 80% cut without cratering margin.

Where the models land

All three (Sol, Terra, Luna) stay available in ChatGPT Work, Codex, and the OpenAI API. Free and Go users in ChatGPT Work and Codex get Terra; Plus, Pro, Business, and Enterprise users can choose Terra and Luna. ChatGPT and Codex subscription prices and quota budgets are unchanged — Terra and Luna usage just consumes fewer credits per request. AWS rollout started the same day (July 30). (OpenAI)

The price war is live, and the cuts came from OpenAI, not the open-weights camp. Luna beating DeepSeek on input-heavy workloads is the line that matters for builders picking an API today. The MiMo and DeepSeek flash caveat matters for builders who only need cheap tokens and don’t need GPT-5.6-class capability. Both are true. Pick accordingly.

Sources