GPT-5.6 — LAUNCHING THURSDAY JULY 10
● Sol (flagship): $5/$30 per million tokens. 88.8% Terminal-Bench 2.1 (Sol Ultra: 91.9%). Advanced coding, biology, cybersecurity. "Max reasoning effort" and "Ultra mode".
● Terra (mid-range): $2.50/$15 per million tokens — 2x cheaper than GPT-5.5. GPT-5.5-competitive performance. Directly competes with Claude Sonnet 5's $2/$10 intro pricing.
● Luna (fastest): $1/$6 per million tokens. Fastest and cheapest of the three. Optimised for high-volume, cost-sensitive applications.
● Sol Fast mode: Up to 750 tokens/second at $12.50/$75 per million tokens.
● Access: ChatGPT, Codex, and API — Thursday July 10.
● Government review: Commerce Dept CAISI cleared all three after 13 days. 700,000+ A100-equivalent hours of automated red-teaming. Sol stays below "Cyber Critical" threshold in OpenAI's Preparedness Framework.
● Grok 4.5: Also launching publicly Thursday — same day, direct competition.
Full Pricing and Benchmark Table
| Model |
Input / 1M |
Output / 1M |
Terminal-Bench 2.1 |
Best for |
| GPT-5.6 Sol |
$5 |
$30 |
88.8% |
Frontier reasoning, long-horizon agentic work, hard science |
| GPT-5.6 Sol Ultra |
$12.50 |
$75 |
91.9% ✓ SOTA |
750 tok/s fast mode, max reasoning effort |
| GPT-5.6 Terra ★ |
$2.50 |
$15 |
GPT-5.5-competitive |
Enterprise workloads — the competitive pricing story |
| GPT-5.6 Luna |
$1 |
$6 |
Fastest of three |
High-volume, cost-sensitive, speed-first applications |
| Claude Sonnet 5 (comparison) |
$2 intro / $3 |
$10 intro / $15 |
80.4% |
Agentic coding, 1M context, 3,000+ MCP integrations |
| GPT-5.5 (comparison) |
$5 |
$30 |
78.2% |
Being superseded by Sol at same price |
★ Terra is the most significant competitive development: GPT-5.5-competitive performance at $2.50/$15/M directly contests Claude Sonnet 5's $2/$10 intro pricing (through Aug 31) and $3/$15 standard pricing.
Sol's Benchmarks — What OpenAI Is Claiming
GPT-5.6 Sol Ultra took the top spot in the TerminalBench 2.1 coding benchmark with 91.9 percent. Claude Mythos 5 hit 88.0 percent, while Google's Gemini 3.1 Pro Preview trailed the field at 70.7 percent. On TerminalBench 2.1, Sol scored 88.8 percent, Sol Ultra hit 91.9 percent, and Mythos 5 landed at 88 percent. That makes Sol the strongest publicly available coding model by Terminal-Bench 2.1 — beating Claude Mythos 5 (88.0%) and Anthropic's Fable 5 (80.3% SWE-bench Pro).
The company says Sol stays below the "Cyber Critical" risk threshold set out in its Preparedness Framework. The model relies on several layers of protection: built-in refusals at the model level and real-time filters that flag potential misuse related to cyberattacks or biological threats. The company also says it put in more than 700,000 hours of A100-equivalent computing power toward automated red-teaming. The cybersecurity framing is not incidental — Sol's advanced coding and cybersecurity capabilities were the exact reason the government held it for 13 days. OpenAI is pre-empting the next review cycle by publishing the safety evidence now.
Terra vs Claude Sonnet 5 — The API Pricing War Begins
Terra at $2.50/$15/M is the pricing event that reshapes the mid-tier API market. Claude Sonnet 5's introductory pricing of $2/$10/M runs through August 31 — cheaper than Terra on input by $0.50, but Terra's $15/M output matches Sonnet 5's standard post-intro price exactly. From September 1, when Sonnet 5 steps up to $3/$15/M, Terra and Sonnet 5 will be priced identically on output ($15/M) with Terra cheaper on input ($2.50 vs $3).
The quality comparison between Terra and Sonnet 5 is the question developers will need to benchmark for their specific workloads. OpenAI claims Terra delivers GPT-5.5-competitive performance. Sonnet 5 beats GPT-5.5 on SWE-bench Pro (63.2% vs 58.6%) and Terminal-Bench 2.1 (80.4% vs 78.2%). If Terra truly matches GPT-5.5, Sonnet 5 leads on coding benchmarks — but the margin narrows and the price is now equal. This is the most competitive mid-tier API landscape since the market began. Run your own benchmarks on your actual production workloads before committing. Do not choose based on the headline claim.
Luna — Why $1/$6/M Changes High-Volume Economics
Luna at $1/$6/M is the cheapest major-lab frontier model output price ever offered. For context: Grok 4.3 on Amazon Bedrock costs $2.50/M output — 2.4x more than Luna. Claude Haiku 4.5 costs $4/M output — 2.3x more than Luna. For high-volume routing, classification, and simple agentic tasks where you currently use Haiku or a small model, Luna offers an alternative at a price previously only available from open-source self-hosted models. The quality ceiling matters: Luna is the smallest of the GPT-5.6 family. But for tasks where GPT-5.5-level capability is overkill, Luna makes a compelling cost case that developers will need to evaluate.
Developer Action Items — What to Do Before Thursday
Benchmark Terra vs Sonnet 5 on your production workloads. Same price post-September 1. Sonnet 5 leads on current benchmarks. Terra may close the gap. Run actual tests on your data — not synthetic benchmarks.
Evaluate Luna for high-volume routing tasks. If you currently use Claude Haiku 4.5 ($4/M output) or Grok 4.3 ($2.50/M) for classification and routing, Luna at $1/$6/M deserves a test run. Quality ceiling applies — evaluate on your specific task before switching production traffic.
Benchmark Sonnet 5 tokenizer costs before Sept 1. Claude Sonnet 5's intro pricing ($2/$10) runs until August 31. The tokenizer adds 1.0-1.35x more tokens — real costs are $2.60-$3.90/M input. With Terra now available at $2.50 input, the actual cost gap between Terra and Sonnet 5 post-intro may be smaller than the headline rates suggest. Benchmark now.
If you need the highest coding ceiling: Sol at $5/$30 with 88.8% Terminal-Bench 2.1 is the new frontier for publicly accessible coding models. Fable 5 (80.3% SWE-bench Pro) is still stronger on agentic coding by SWE-bench metric — but Sol's Terminal-Bench leadership is legitimate. The comparison will depend on what type of coding work you do.
Sources: Axios (July 8), MLQ News, The Decoder, Outlook Business, Android Headlines, QZ · Related: Grok 4 vs Claude Sonnet 5 vs GPT-5.5 benchmarks → · Best AI tools July 2026 rankings → · GPT-5.6 restricted preview explained →