senn-techOpenAI GPT-5.6: Speed as a Separate Product
On 30 July 2026 OpenAI repriced GPT-5.6 and introduced a Fast Mode for the strongest model, Sol. The new message: if you want speed, you pay extra.
| Model | Input / 1M tokens | Output / 1M tokens | Change | Fast Mode surcharge |
|---|---|---|---|---|
| GPT-5.6 Luna | $0.20 | $1.20 | −80% | — |
| GPT-5.6 Terra | $2.00 | $12.00 | −20% | — |
| GPT-5.6 Sol | $5.00 | $30.00 | unchanged | +$5.00 input / +$30.00 output, 2.5× faster |
Luna is the clear winner: it now costs one fifth of the previous input price and one fifth of the previous output price. Terra gets a moderate discount, while Sol stays the same — speed becomes the upsell there.
What is changing in practice
- Luna −80%: input drops from roughly $1.00 to $0.20, output from roughly $6.00 to $1.20 per million tokens.
- Terra −20%: input at $2.00, output at $12.00 — a small discount for the broad mid-tier.
- Sol unchanged: input $5.00, output $30.00, but now with an optional Fast Mode.
- Fast Mode: 2.5× faster responses at 2× the price — so $10.00 input / $60.00 output.
- Marketing claim: A task that cost about $1 a year ago should now cost about $0.06 with Luna.
Speed as a premium feature
Until now the most expensive model was usually also the fastest. With GPT-5.6 OpenAI deliberately separates capability from latency: Sol remains the most capable model, but if you want it fast you book Fast Mode on top. This is clear price segmentation — comparable to express shipping or reserved cloud capacity.
For customers it means: do not automatically order the top model. Check whether the application is actually latency-critical. Batch jobs, document analysis, and overnight reports do not need Fast Mode. Interactive agents, live chat, and tool calls with many rounds do.
The agent workflow calculation
Agentic systems are chains of short calls. Each step costs input tokens, often with a long context. This is exactly where the Luna price cut hits hardest: a planning call, a classification call, a tool call — together they could previously add up to several dollars quickly. With Luna even a hundred calls stay in the low double-digit cent range.
Sol Fast Mode, by contrast, is a latency upgrade for the rare critical path: when a single call determines the response time of an entire chain, the 2.5× speed-up is worth it.
Our take
OpenAI is making the business model more transparent but also more complex. The cheap Luna tier drastically lowers the barrier to entry for agentic applications. Fast Mode, however, is not a technical must-have but a deliberate price feature — useful for real latency cases, unnecessary for everything else. Workloads that are analysed before ordering will save more than the headline price drop alone.
Further reading
What does the 80% Luna price drop mean for existing workloads?+
Luna falls to $0.20 input and $1.20 output per million tokens. Workloads previously running on Terra or Sol because Luna was too expensive can now migrate much more cheaply. The biggest savings come from short classification, summarization, or retrieval steps.
When does Sol Fast Mode pay off?+
Fast Mode costs twice the standard Sol rate and promises 2.5× faster responses. It pays off when latency directly costs money — customer-service chat, live classification, coding agents with many sequential calls. For overnight batch processing or report generation it is wasted budget.
How does this change the cost picture for agentic workflows?+
Agentic flows consist of many small calls, not one long prompt. That is exactly where the per-token input price matters. OpenAI claims a task that cost roughly $1 a year ago now costs about $0.06 with Luna. The exact figure depends on the benchmark, but the direction is clear: agents are getting cheaper, not more expensive.