Pricing and budget · 2026-09-22

GPT-5.6 Terra: the 30 July 2026 price drop and the full reasoning_effort range

Covers that openai/gpt-5.6-terra's pricing card has two periods, which figure is current today, and how to choose among its six reasoning_effort levels from none to max.

Timeline showing GPT-5.6 Terra moving from an old price to a new, lower price on 30 July 2026, alongside six reasoning_effort levels from none to max.

The pricing card has two periods — which one is current

GPT-5.6 Terra's price history has two periods: before 30 July 2026, input was $2.5 per million tokens and output was $15; from that date onward (effectiveFrom: 2026-07-30, no effectiveTo), input dropped to $2 and output to $12, and that's the price current as of today. An integration referencing old documentation or a prior budget calculation may still be using the old figures; base your budget planning on the current pricing card on the model page.

There's also a separate, higher 'long_context' pricing tier for requests exceeding the standard 1.05M-token context range; account for that tier separately if you're sending very large context.

reasoning_effort carries six levels

Terra supports six reasoning_effort levels: none, low, medium, high, xhigh, and max. None skips the reasoning step entirely for the lowest latency; it fits simple, direct questions. Low and medium offer a reasonable balance for everyday assistant work and medium-complexity analysis. High, xhigh, and max provide deeper evaluation for multi-step coding, complex analysis, and agent loops; the catalog description specifically emphasizes 'true max reasoning effort,' meaning the max level genuinely uses the model's top reasoning capacity rather than being a nominal ceiling.

Both latency and reasoning-token cost rise with each level; rather than running a task at the highest level by default, starting from a low level and raising it only when answer quality falls short keeps cost under control.

  • none: lowest latency, simple questions.
  • low/medium: everyday assistant work and medium-complexity analysis.
  • high/xhigh/max: multi-step coding and complex agent tasks; max is a genuine ceiling.

Image input and prompt caching are also supported

The model accepts image input alongside text, and repeated context prefixes are served from a prompt cache; the cache-read price is far below the normal input price. In a production flow that resends the same system instruction or a fixed document prefix repeatedly, this can noticeably lower total cost.

Frequently asked questions

Is the $2.5/$15 figure I saw in older documentation still current?

No, that figure applied before 30 July 2026. The current price is $2 input / $12 output per million tokens; check the model page for the exact figure, since prices can be updated again over time.

What's the difference between max and xhigh?

Both provide high-depth reasoning; max is the top level, specifically emphasized in the catalog description as 'true max reasoning effort.' Compare the difference with your own tasks to confirm the extra cost is justified.

Related posts