OpenAI slashes GPT-5.6 prices up to 80% and adds high-speed Sol mode

Running real-world enterprise AI agents has always come down to a brutal math problem: multi-step workflows burn through hundreds of calls, and using top-tier models everywhere quickly blows through budgets.

OpenAI introduced major pricing and latency updates to the GPT-5.6 family. The lightweight GPT-5.6 Luna had its API prices slashed by 80%, everyday workhorse GPT-5.6 Terra dropped by 20%, and frontier model GPT-5.6 Sol gained an ultra-low latency Fast mode that boosts speeds up to 2.5x.

Huge cuts at the bottom, raw speed at the top

The pricing reduction on Luna fundamentally changes the economics of high-volume tasks.

Starting today, Luna drops to $0.20 per million input tokens and $1.20 per million output tokens. Terra falls to $2 per million input and $12 per million output tokens. According to OpenAI, Luna matches frontier performance from a year ago at roughly 6 cents on the dollar, running at nearly nine times the speed. On the Agents' Last Exam benchmark, Luna actually beat Fable 5 while running at an estimated per-task cost nearly 99% lower.

These price cuts also carry over to ChatGPT Work and Codex. Subscription prices and quotas stay the same, but running Luna and Terra now burns fewer usage credits.

Meanwhile, developers who care more about speed than saving cash get Fast mode in the API for GPT-5.6 Sol. Replacing the older Priority Processing tier, Fast mode charges twice the standard rate to deliver up to 2.5x faster throughput with identical intelligence. Any existing priority requests automatically forward to Fast mode.

AI rewrote its own inference stack

The secret behind the sudden discounts isn't just hardware buildouts; OpenAI had its own AI find the savings.

OpenAI shared that Sol autonomously analyzed production workloads, rewrote low-level serving kernels, and designed hundreds of optimization experiments under human guidance. That autonomous kernel overhaul shaved 20% off end-to-end serving costs and boosted token generation efficiency by more than 15%.

Paired with smarter context caching and routing, the GPT-5.6 models take cleaner execution paths, meaning fewer tokens wasted on repetitive agent loops.

That makes split architecture far more practical. Teams can deploy Sol to map out architecture and break down complex problems, then hand off code execution, unit testing, and tool calling to a fleet of dirt-cheap Luna instances.

Why it matters

For builders running multi-agent workflows or massive document ingestion pipelines, this pricing drop turns previously non-viable ideas into practical products. Luna at $0.20 per million input tokens makes brute-force background analysis remarkably cheap without forcing developers to drop down to poorly supported open-weight setups.

Across the competitive landscape, the cost floor is shifting. Price cuts are no longer just loss-leading marketing moves; they reflect structural efficiency gains generated by frontier models optimizing their own runtime software. Rivals will have to match not just raw benchmark reasoning, but this self-reinforcing engineering feedback loop.

Still, we think developers should stay cautious with Sol Fast mode. Paying a 100% price premium for 2.5x speed is fantastic for live voice or user-facing copilots, but running unchecked background tasks on Fast mode will drain API balances fast.