GPT-5.6 vs GPT-5.5
OpenAI's new flagship against the model it replaces — the headline number is that Sol costs exactly the same as GPT-5.5 (Thinking) while posting higher scores on every benchmark suite both have published.
GPT-5.6 vs GPT-5.5: side-by-side
| GPT-5.6 (Sol) | GPT-5.5 | |
|---|---|---|
| Maker | OpenAI | OpenAI |
| Released | Preview 26 Jun 2026 · GA 9 Jul 2026 | 23 April 2026 |
| Status | Generally available — flagship | Previous-generation flagship |
| Context window | 1,050,000 tokens | 1,000,000 tokens (400K in Codex CLI) |
| Max output | 128,000 tokens | Not specified on release materials |
| Knowledge cutoff | 16 February 2026 | Not published |
| API price (per 1M) | $5 / $30 (Sol) | $5 / $30 (Thinking) |
| Cheaper tiers | Terra $2.50/$15 · Luna $1/$6 | None below Thinking tier |
| Reasoning controls | Standard · max · ultra (subagent decomposition) | Thinking · Pro variant |
| Access | ChatGPT, ChatGPT Work, Codex, API | ChatGPT & Codex (Plus, Pro, Business, Enterprise) + API |
Price comparison
| Model | Input (per 1M) | Output (per 1M) |
|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 |
| GPT-5.6 Terra | $2.50 | $15.00 |
| GPT-5.6 Luna | $1.00 | $6.00 |
| GPT-5.5 (Thinking) | $5.00 | $30.00 |
| GPT-5.5 Pro | $30.00 | $180.00 |
Estimate your monthly API cost for GPT-5.6 vs GPT-5.5 with the LLM API pricing calculator →
Published benchmark figures
Each figure is the model's own published score as listed on its AI Model Hub page. Note the suite versions: GPT-5.5's materials report Terminal-Bench 2.0, while GPT-5.6's GA materials use Terminal-Bench 2.1 — different suite versions are not directly comparable in absolute terms, though both are shown here for context.
| Model | Benchmark | Published score |
|---|---|---|
| GPT-5.6 Sol | Terminal-Bench 2.1 | 91.9% |
| GPT-5.6 Sol | SWE-bench Pro | 64.6% |
| GPT-5.6 Sol | Agents' Last Exam | 53.6 |
| GPT-5.6 Sol | Artificial Analysis Coding Agent Index (max) | 80 |
| GPT-5.5 | Terminal-Bench 2.0 | 82.7% |
| GPT-5.5 | GDPval | 84.9% |
| GPT-5.5 | Expert-SWE | 73.1% |
| GPT-5.5 | SWE-bench Pro | 58.6% |
| GPT-5.5 | FrontierMath Tier 4 | 35.4% |
What the table says
- Pricing. Sol matches GPT-5.5 (Thinking) exactly at $5/$30 per 1M tokens — the first time in this OpenAI generation that a flagship replacement hasn't repriced. Terra and Luna are new tiers with no direct GPT-5.5 analogue, undercutting it by half and by 5-6x respectively.
- Benchmarks. On the one suite both report cleanly — SWE-bench Pro — Sol's 64.6% beats GPT-5.5's 58.6% by 6 points. On agentic terminal tasks Sol's 91.9% (TB 2.1) is well ahead of GPT-5.5's 82.7% (TB 2.0), though the suite-version difference means this isn't a clean apples-to-apples gain.
- Efficiency. OpenAI's own materials claim Sol is 54% more token-efficient on agentic coding than the next-highest-scoring model on the Artificial Analysis Coding Agent Index — a claim GPT-5.5 never made at launch.
- New capability, not just new numbers. GPT-5.6 adds ultra mode (internal subagent decomposition), a max reasoning tier, programmatic tool calling, and explicit prompt-cache breakpoints — features GPT-5.5 shipped without.
Key takeaway
Same price, higher published scores, and new reasoning/caching features — GPT-5.6 Sol is a straightforward upgrade path for anyone already paying GPT-5.5 (Thinking) rates. The more interesting decision is whether Terra's half-price tier covers what used to require full GPT-5.5. Run your own workload through the pricing calculator to compare all five rate cards at once.
Full GPT-5.6 overview Full GPT-5.5 overview
Frequently asked questions
Is GPT-5.6 more expensive than GPT-5.5?
No — Sol's $5/$30 rate card is identical to GPT-5.5 (Thinking). Terra ($2.50/$15) and Luna ($1/$6) are new, cheaper tiers with no GPT-5.5 equivalent.
Is GPT-5.6 better than GPT-5.5 on benchmarks?
On the suites both have published, yes — 91.9% vs 82.7% on Terminal-Bench (2.1 vs 2.0, different versions) and 64.6% vs 58.6% on SWE-bench Pro.
Should I switch from GPT-5.5 to GPT-5.6?
For most workloads, yes — Sol matches GPT-5.5's price with higher published scores and new reasoning/caching features. GPT-5.5 remains available but is now the previous-generation flagship.
What's new in GPT-5.6 that GPT-5.5 doesn't have?
Ultra mode (internal subagent decomposition), a max reasoning-effort tier, programmatic tool calling, native multi-agent API support, and explicit prompt-cache breakpoints with a 30-minute minimum cache life.