Update, September 29, 2026: OpenAI released the successor GPT-6 Astra on September 3, 2026, and GPT-6 Sol and GPT-6 Luna on September 22 (US time). This article compares the generation before them: GPT-5.6 and GPT-5.5. Where it calls Sol the "flagship" or "top tier," that means within the GPT-5.6 generation. If you are choosing where to move from GPT-5.5 today, start with the GPT-6 generation's Astra, Sol, and Luna: as of September 30, 2026, the API model list recommends Astra if you are not sure where to start, GPT-6.1 Sol (released September 29; in ChatGPT it is available in Work and Codex, not in Chat) to balance intelligence and cost, and GPT-6 Luna for cost-sensitive, high-volume work. GPT-5.5 retires from ChatGPT, ChatGPT Work, and Codex on October 14, 2026 (the OpenAI API is not affected), and OpenAI names GPT-6 Sol as the replacement on paid plans and GPT-6 Luna in the desktop app on Free and Go (GPT-5.6 Sol, Terra, and Luna remain available during the transition period). For cost, see our measured comparison of Astra at low and GPT-5.6 Sol at high.
Contents
- 1. What changed — from one flagship to three models
- 2. Spec at a glance
- 3. The Terra shock — "5.5-class quality at 40% of 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 a far lower price." It launched at half the price, and the July 30, 2026 price cut brought it down to 40% of GPT-5.5's unit 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 40% of 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 40% of the price. The de facto successor for "people for whom 5.5 was enough."
GPT-5.6's flagship: the model 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 40% of the price) |
| API price | $5 / $30 | $5 / $30 | $2 / $12 |
| 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% | 63.4% |
| Reasoning efficiency | Baseline | OpenAI says it gets equal or better results with fewer tokens (no figure published for the reduction versus GPT-5.5) | |
| Reasoning control | effort (none–xhigh) | effort (none–max, max added) | |
* Prices and specs are based on OpenAI's model pages (GPT-5.5, GPT-5.6 Sol, GPT-5.6 Terra). Sol has a three-month promotional price of $4/$20 starting August 21, 2026. On SWE-Bench Pro, GPT-5.5's 58.6% is OpenAI's figure from the GPT-5.5 launch (April 2026), and Sol's 64.6% is from the evaluation table OpenAI published with GPT-5.6 (the same table lists GPT-5.5 at 59.4%). On TerminalBench, GPT-5.5's 82.7% at its launch was on version 2.0; the GPT-5.6 table also measures GPT-5.5 on 2.1, at 85.6% versus Sol's 88.8% (see Section 4).
3. The Terra shock — "5.5-class quality at 40% of 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." (OpenAI)
Same quality, 60% off the bill
Terra: $2.00
→ 60% off
Terra: $12.00
→ 60% off
→ no reduction figure published; if it holds, effective cost drops another notch
* OpenAI says it trained GPT-5.6 "to get more useful work from every token," but it has not published a figure for the reduction versus GPT-5.5. Terra launched at half the price, and the July 30, 2026 price cut took it to 40% of GPT-5.5's rate.
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 cut costs by 60% 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 (source: OpenAI)
On the same-metric SWE-Bench Pro, the score rises about 6 points from GPT-5.5's 58.6% to Sol (64.6%). Fix ability on real, production-grade coding has steadily improved. Note, however, that the two figures come from separate OpenAI announcements: in the evaluation table published with GPT-5.6, GPT-5.5 is listed at 59.4% (a gap of about 5 points).
⚠️ Caveats when comparing
- TerminalBench: GPT-5.5's 82.7% at launch is on 2.0, and GPT-5.6 Sol's 88.8% is on 2.1, so you can't just subtract them. In the GPT-5.6 evaluation table, which measures both on 2.1, GPT-5.5 scores 85.6% (a gap of about 3 points).
- GPQA, FrontierMath: These are in the GPT-5.6 evaluation table. On GPQA Diamond, Sol scores 94.6% versus GPT-5.5's 93.6%; on FrontierMath Tier 1–3, 89% versus 85.3%. Both are near the ceiling, so the gains are small. AIME and MMLU are not listed.
This time OpenAI put new agentic metrics up front (Agents' Last Exam 53.6, Coding Agent Index 80, and so on). On general reasoning, GPQA Diamond moves by just 1 point near the ceiling, so measuring "how much smarter the generation got" 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. The Codex app merged into the new ChatGPT desktop app.
- 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 + 20M output × $12. Better token efficiency could trim it further.
On unit price alone, the bill drops by about 60%. On top of that, OpenAI says GPT-5.6 finishes work with fewer tokens; if so, the actual reduction can exceed the unit-price gap (no reduction figure has been published). 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 40% of the 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 $0.20/$1.20 |
| 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 cut the bill by 60%, 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 40% of the price ($2/$12). 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% (OpenAI figures, +6pt).
- Efficiency also improved: OpenAI says it gets equal or better results with fewer tokens (no figure published for the reduction versus GPT-5.5).
- Caveats: TerminalBench versions differ between announcements (on the same 2.1, GPT-5.5 scores 85.6% and Sol 88.8%). Gains in general reasoning such as GPQA are small.
- The optimal migration: first cut the bill by 60% 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 40% of the 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." In the GPT-5.6 evaluation table, Terra beats GPT-5.5 on SWE-Bench Pro (63.4% vs 59.4%) but trails slightly on some items, such as GPQA Diamond (92.9% vs 93.6%). 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; both are OpenAI figures). TerminalBench changed version from 2.0 to 2.1 between announcements, but in the table that measures both on 2.1, GPT-5.5 scores 85.6% and Sol 88.8%. GPQA Diamond goes from 93.6% to 94.6%, near the ceiling, so the gain is small. The reality is "what clearly improved is real coding and agentic work; general reasoning is nearly flat."
Q4. How much does cost actually drop?
Migrating to Terra brings the unit price to about 40% of GPT-5.5's. For 100M input / 20M output per month, that's roughly $1,100→$440 (after the July 30, 2026 price cut). OpenAI also says GPT-5.6 finishes work with fewer tokens; if so, the real reduction can exceed the unit-price gap (no reduction figure has been published).
Q5. I was using GPT-5.5 Pro. What's the successor?
For most use cases, Sol ($5/$30, or $4/$20 on promotional pricing for three months from August 21, 2026) 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 (even in OpenAI's evaluation table, Claude Fable 5 scores 80% versus Sol's 64.6%). 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 of the time
- GPT-5.6 Sol vs Claude Fable 5: in-depth comparison — top-tier showdown
