Contents
- 1. What changed — from one flagship to three models
- 2. Spec at a glance
- 3. The Terra shock — "5.5-class quality at half the price"
- 4. Benchmarks — how much did the generation improve
- 5. New features added in 5.6
- 6. Real cost — how your bill changes after upgrading
- 7. Should you migrate — which model to switch to
- Summary
- FAQ
Just two and a half months after GPT-5.5 shipped on April 23, 2026, OpenAI made GPT-5.6 generally available on July 9. But this update is not "just a performance bump." The biggest change is a reorganization from "a single flagship model" into a "three-model lineup: Luna / Terra / Sol."
And for most GPT-5.5 users, what matters most is that the mid-tier model Terra delivers "GPT-5.5-class quality at roughly half the price." For those who want more performance there is the top-tier Sol (upgraded at the same price as GPT-5.5); for those who want to cut costs there is Terra — a real choice has appeared. In this article we lay out, based on official announcements and independent analysis, what changed from GPT-5.5 to 5.6 and which model you should switch to.
From one flagship to a three-model lineup
— so you can choose to "raise performance" or "lower the bill"
Terra delivers "5.5-class quality at half the price" ／ Sol delivers "more performance at the same price"
1. What changed — from one flagship to three models
GPT-5.5 was a two-tier setup: "the base model ($5/$30) ＋ the pricey GPT-5.5 Pro ($30/$180)." GPT-5.6 reshapes this into three models you pick by use case and cost.
For high-volume processing, classification, and light chat. A newly added tier one notch cheaper than GPT-5.5.
GPT-5.5-class quality at about half the price. The de facto successor for "people for whom 5.5 was enough."
The flagship that raises performance at the same price as GPT-5.5. The pure-performance successor.
In other words, where GPT-5.5 users migrate depends on their use case. Want the same quality for less? Terra. Want more performance for the same price? Sol. Want to crunch volume even cheaper? Luna. Compared with the single-model 5.5, the freedom of choice has grown enormously.
2. Spec at a glance
| Item | GPT-5.5 | GPT-5.6 Sol | GPT-5.6 Terra |
|---|---|---|---|
| Release | April 23, 2026 | July 9, 2026 | |
| Positioning | Single flagship (＋ Pro) | Top tier | Balanced (5.5-class at half price) |
| API price | $5 / $30 | $5 / $30 | $2.50 / $15 |
| Context | 1,050,000 | 1,050,000 | 1,050,000 |
| Max output | 128,000 | 128,000 | 128,000 |
| Knowledge cutoff | December 1, 2025 | February 16, 2026 (updated) | |
| SWE-Bench Pro | 58.6% | 64.6% (est.) | Not disclosed (said to be 5.5-class) |
| Reasoning efficiency | Baseline | About 10–15% better on comparable tasks (up to 54% for coding) | |
| Reasoning control | effort (none–xhigh) | effort (none–max, max added) | |
※ Prices and specs are based on each company's official announcements. Sol's SWE-Bench Pro is an independently compiled estimate (OpenAI has not published it officially). TerminalBench is version 2.0 for 5.5 and 2.1 for 5.6, so the versions differ and cannot be compared directly (see Section 4).
3. The Terra shock — "5.5-class quality at half the price"
The biggest impact of this update is not Sol's performance gain at the top, but rather Terra's price-to-performance in the middle. OpenAI positions Terra as "delivering GPT-5.5-class quality at roughly half the price."
Same quality, half the bill
Terra: $2.50
→ half
Terra: $15.00
→ half
→ effective cost drops another notch
※ OpenAI states that "GPT-5.6 is about 10–15% more token-efficient than GPT-5.5 on comparable tasks," and this stacks on top of Terra's half-price rates.
Terra is, in effect, the model for users who found "GPT-5.5 was enough, but I just want to lower the bill." If you can halve costs without dropping quality, it becomes the first migration target for many existing workloads. For most practical work that doesn't chase peak performance, Terra — not Sol — is enough.
4. Benchmarks — how much did the generation improve
One caveat when looking at generational gains: anything whose benchmark version has changed cannot be compared directly. Here we separate "what can be compared at the same version" from "what cannot."
Real coding is +6pt across the generation
Fix rate on real GitHub issues. About 6 points higher across the generation (Sol is an estimate)
On the same-metric SWE-Bench Pro, the score rises about 6 points from GPT-5.5's 58.6% to Sol (estimated 64.6%). Fix ability on real, production-grade coding has steadily improved. Note, however, that Sol's figure is not officially published by OpenAI and is an independently compiled estimate.
⚠️ Metrics that cannot be compared directly
- TerminalBench: GPT-5.5 scores 82.7% on 2.0, and GPT-5.6 Sol scores 88.8% on 2.1. Because the benchmark versions differ, you can't just subtract the numbers.
- GPQA, AIME, MMLU, etc.: Many are undisclosed for GPT-5.6 ("private benchmark questions"). The generational gain in reasoning is hard to compare directly.
This time OpenAI released new agentic metrics (Agents' Last Exam 53.6, Coding Agent Index 80, and so on) while withholding many direct comparison values for general reasoning and math. The reality is that measuring "how much smarter the generation got" rigorously with legacy benchmarks is difficult.
5. New features added in 5.6
- Three-model lineup (Luna/Terra/Sol) — choose by use case and cost. The biggest change.
- Programmatic Tool Calling — a new feature where the model generates JavaScript to orchestrate tool calls (Responses API). Runs complex agents in fewer round trips.
- "max" added to reasoning effort — six levels now: none/low/medium/high/xhigh plus max.
- ChatGPT Work — an agent for business use. A new desktop app bundled with Codex ships at the same time.
- GPT-Live (full-duplex voice) — natural voice conversation that can be interrupted rather than turn-based.
- GitHub Copilot support — Sol/Terra/Luna are now selectable in Copilot.
- Knowledge cutoff updated — December 2025 → February 16, 2026.
6. Real cost — how your bill changes after upgrading
Migrating to Terra has a large impact on cost. Comparing, say, a workload of 100M input tokens / 20M output tokens per month:
100M input × $5 ＋ 20M output × $30
100M input × $2.50 ＋ 20M output × $15. Token efficiency effectively trims it further.
On unit price alone, the bill is about halved. On top of that, the "about 10–15% better token efficiency on comparable tasks" stacks on, so the actual reduction can exceed the unit-price gap. For any workload where GPT-5.5-class quality is enough, migrating to Terra is almost pure cost savings.
※ The above is a rough estimate based on unit prices. Actual billing varies with prompt length, cache usage, and output volume. We recommend measuring with your own representative requests.
7. Should you migrate — which model to switch to
| How you use GPT-5.5 today | Recommended target | Reason |
|---|---|---|
| Quality is enough, want lower cost | Terra | 5.5-class quality at half price. Pure cost savings |
| Want more performance on the same budget | Sol | SWE-Bench Pro and more improve at the same $5/$30 |
| High-volume, light work, cheapest possible | Luna | New lowest tier at $1/$6 |
| You were using GPT-5.5 Pro ($30/$180) | Sol | Far cheaper than Pro and often enough for daily high-reasoning work |
| Code-fix quality is the top priority | Sol ＋ external comparison | For real coding, the Claude models are also candidates |
Migration itself is just switching the model ID, and the API is largely compatible. For most users, the optimal answer is a two-stage setup: "first drop down to Terra to halve the bill, then bump only the work that needs performance up to Sol." Compared with the single-model 5.5 era, tuning cost and quality has become easier.
Summary
- The biggest change is going to three models: from GPT-5.5's single flagship to the use-case-specific trio Luna/Terra/Sol.
- Terra is the star: GPT-5.5-class quality at about half price ($2.50/$15). The de facto successor for existing workloads.
- Sol raises performance at the same price: $5/$30 held steady while SWE-Bench Pro goes 58.6→64.6% (estimated, +6pt).
- Efficiency also improved: about 10–15% better token efficiency on comparable tasks, so real cost drops further.
- Caveats: TerminalBench can't be compared directly due to a version difference, and many general-reasoning benchmarks are undisclosed.
- The optimal migration: first halve the bill with Terra, then move only performance-critical work to Sol.
FAQ
Q1. If I'm switching from GPT-5.5, which model?
It depends on your use case. Want to keep the quality but lower cost? Terra (5.5-class at half price). Want more performance at the same price? Sol. Want to crunch volume as cheaply as possible? Luna. Most practical work is fine on Terra, and the best approach is a two-stage setup: bump only the performance-critical work up to Sol.
Q2. Is Terra really equal in quality to GPT-5.5?
OpenAI positions it as "GPT-5.5-class quality at about half price." Fully like-for-like benchmark comparisons are limited, but positionally it's "the successor for use cases where 5.5 was enough." If you're concerned, the safe move is to measure and compare Terra's and 5.5's output on your own representative tasks before migrating.
Q3. How much smarter did the generation get?
On the same-metric SWE-Bench Pro it rises about 6 points, from 58.6% to 64.6% (Sol, estimated). However, TerminalBench changed version from 2.0 to 2.1 and can't be compared directly, and much of the general reasoning like GPQA and AIME is undisclosed. The reality is "what clearly improved is real coding; rigorously comparing the rest is difficult."
Q4. How much does cost actually drop?
Migrating to Terra roughly halves the unit price. For 100M input / 20M output per month, that's roughly $1,100→$550. On top of that, the "about 10–15% better token efficiency on comparable tasks" stacks on, so the real reduction can exceed the unit-price gap.
Q5. I was using GPT-5.5 Pro. What's the successor?
For most use cases, Sol ($5/$30) is enough. GPT-5.5 Pro was pricey at $30/$180, but Sol handles high-reasoning work far more cheaply. Judge from your real tasks whether only ultra-hard reasoning still needs Pro-class capability.
Q6. Will my existing GPT-5.5 apps still work?
The API is largely compatible, and you can migrate by switching the model ID. But since there are now three models, redesigning which work you assign to Luna/Terra/Sol lets you optimize cost and quality. For many, starting from Terra is effective.
Q7. How does GPT-5.6 differ from other vendors' models (Claude)?
On real coding (SWE-Bench Pro), the Claude models tend to be strong. For details, see the comparisons GPT-5.6 Sol vs Claude Opus 4.8 and vs Claude Fable 5.
Related articles
- GPT-5.6 release: full breakdown — details on the three models Luna/Terra/Sol
- GPT-5.5 release: full breakdown — details on the previous-generation model
- GPT-5.6 Sol vs Claude Opus 4.8: in-depth comparison — comparison with a rival flagship
- GPT-5.6 Sol vs Claude Fable 5: in-depth comparison — top-tier showdown
