# OpenAI Ships GPT-5.6, Trading Raw Scale for Tokens-Per-Answer Efficiency

The three-tier family uses fewer tokens by default, and OpenAI says two configuration settings alone raised its ARC-AGI-3 score from 13.3 to 38.3 percent.

- Published: 2026-07-30T05:37:30.825Z
- Canonical: https://polylog.news/ai/2026-07-30/openai-ships-gpt-5-6-trading-raw-scale-for-tokens-per-answer
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [OpenAI](https://openai.com/index/gpt-5-6-frontier-intelligence-efficiency/), [OpenAI (ARC-AGI-3)](https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores/), [Artificial Analysis](https://artificialanalysis.ai/articles/gpt-5-6-has-landed)

OpenAI has begun rolling out [GPT-5.6](https://openai.com/index/gpt-5-6-frontier-intelligence-efficiency/), a family split into three tiers. Sol handles the most demanding reasoning, Terra handles everyday work, and Luna handles high-volume, latency-sensitive calls. The selling point is not a new peak in capability but more capability per dollar. OpenAI says Terra matches the earlier GPT-5.5 on its intelligence benchmarks at roughly half the price. It also says Sol at maximum reasoning uses fewer output tokens than most models of comparable intelligence, about 15,000 tokens per Intelligence Index task against 16,000 for GPT-5.5. The independent evaluator [Artificial Analysis](https://artificialanalysis.ai/articles/gpt-5-6-has-landed) reports that Sol at maximum effort leads its Coding Agent Index at under half the cost of the previous leader. Because that figure comes from an outside tester rather than OpenAI, it carries more weight than the company's own claims.

A companion post on the [ARC-AGI-3](https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores/) puzzle-game benchmark makes a more important point. Under the official test setup, Sol scored 13.3 percent. Turning on two Application Programming Interface (API) settings raised that to 38.3 percent while cutting output tokens by roughly six times. The two settings were retained reasoning, which passes the previous response identifier so the model keeps its private chain of steps across tool calls, and compaction. The default setup was discarding the model's own reasoning between moves, so it had to work out each game again from the beginning. This is a measurement artifact, not a new capability. It also has a wider implication. Many published agent scores understate models whose strength depends on keeping state between steps.

On security-related capability, OpenAI reports that GPT-5.6 nearly doubled GPT-5.5's peak pass rate on ExploitGym, its vulnerability-discovery benchmark, rising from 15.1 to 24.9 percent under a two-hour limit and reaching 33.7 percent given six hours. Those are OpenAI's own figures on OpenAI's own benchmark, and they come the same week the company is dealing with an autonomous agent that used exactly that kind of capability to breach live infrastructure.

## What this means

The basis of competition is shifting from peak benchmark scores to intelligence per token, because serving reasoning models at scale means labs now spend most on inference rather than training. GPT-5.6's tiered pricing and token cuts put pressure on every rival that charges for raw capability. The ARC-AGI-3 result shows how much reported agent performance depends on test configuration rather than the model weights, which favors labs whose serving systems manage context state well.

## What to watch

- Whether independent evaluators reproduce the Terra "half the price of GPT-5.5" and Sol coding-index claims on their own test setups, which would confirm the efficiency gain is real rather than dependent on configuration.
- How quickly competitors release their own versions of retained reasoning and compaction, because if standard benchmarks understate models that keep state, rankings will change as test setups are updated.