GPT-5.6: Sol, Terra, and Luna Available Without Subscription

GPT-5.6: Sol, Terra, and Luna Available Without Subscription

GPT-5.6 ends the naming chaos at OpenAI. The number indicates the generation, making it easier to categorize, thus avoiding the need to rename the entire family during the next upgrade of Luna.

Sol is the flagship model, designed for the 10% of the hardest tasks: long-term coding agents, security research, in-depth scientific analysis. New reasoning controls are integrated here.

Terra is the workhorse model, the obvious migration target. It offers GPT-5.5 class quality at half the price, aimed at production volumes: support, internal tools, document pipelines.

Luna is the fast model, and discreetly the surprise model of the launch. The cheapest model in the family ranks close to GPT-5.5 on several tests.

The respective names in the API are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. This may seem like a small change on paper, but it is significant for any developer who has tried to keep track of o3, o4-mini, GPT-4 Turbo, and 4o all at once.

Pricing: Four Ways to Pay

Three models, but four prices, as the launch week revealed a peculiarity.

• Input / 1M tokens
• Output / 1M tokens
• Flagship, deepest reasoning
• Same model at up to 750 tokens/sec
• GPT-5.5 class quality at half price
• Fast and high-volume workloads

Sol Fast is the new offering here: the same flagship brain served from Cerebras hardware at up to 750 tokens per second, for 2.5 times the standard rate. Speed as an explicit paid category, rather than a queue lottery, is something OpenAI has never sold before. If your product is limited by latency, this element changes what is viable.

The more discreet pricing story concerns the cache, and agent builders should be more concerned about this than the main rates:

• Explicit cache breakpoints, you control what is cached instead of guessing
• Minimum cache duration of 30 minutes
• Cache writes charged at 1.25 times the uncached input rate
• Cache reads retaining the 90% discount

For long-term agents that re-read the same context hundreds of times, this discount accumulates into an order-of-magnitude reduction in input costs. Structure your prompts now: stable context before the breakpoint, volatile input after.

Capabilities: Maximum Effort, Ultra Mode, and a Surprise Model

OpenAI is withholding the expanded evaluation suite for the GA system card, but preliminary figures already sketch the picture. Two new controls highlight Sol:

• Maximum reasoning effort, a new ceiling that gives Sol the most time to think through a problem.
• Ultra mode, which completely transcends the single-agent paradigm. Sol activates sub-agents and coordinates them to parallelize complex work.

On benchmarks, remarkable claims:

• Terminal-Bench 2.1: Sol sets a new state-of-the-art on command-line workflow requiring planning, iteration, and tool coordination.
• GeneBench v1: Sol outperforms GPT-5.5 on long-term genomic analysis and quantitative biology, using fewer tokens to achieve it.
• ExploitBench: Sol is competitive with Mythos Preview at about a third of the output tokens.

The family effect: Sol and Terra set new highs across the board, while Luna performs close to GPT-5.5 on several tests despite its lower price.

This last point is the surprise model. The flagship quality of the latest generation is now available at $1 per million input tokens. The model across the family is not just "smarter," it is smarter per token and per dollar. Efficiency is the true title.

The Capability No One Expected in the Budget Category

Here is the detail of the system card that was buried under the drama of availability, and it deserves its own section.

All three models, not just Sol, are rated at OpenAI's "High" risk level for cyber and biological capability. During internal capture-the-flag security tests:

• Internal CTF results across the family

To give you perspective, these models are comparable to Fable 5 category of Claude.

"GPT-5.6 Sol is better at helping people find and fix vulnerabilities than at reliably executing end-to-end attacks."

This is the company's framework, and the strategy follows: empower defenders, make offensive use difficult, uncertain, and detectable.

Five Layers Deep: The Protection Stack

The security architecture delivered with 5.6 is the most elaborate OpenAI has publicly described, with configurations tailored to the capability of each level. The design assumption is straightforward: no single protection survives a determined and adaptive attacker.

Here’s how the process unfolded:

• Trained refusals. The model itself refuses prohibited cyber assistance, including disguised or jailbroken requests.

• Real-time classifiers. Cyber and biological misuse detectors evaluate output as it is generated.

• Reasoning model review. High-risk generations are paused mid-way while a larger model examines the full context. Unauthorized outputs never reach the user.

• Account-level signals. Reported activity triggers a review across conversations, allowing OpenAI to distinguish a security researcher from a persistent malicious actor.

• Differentiated access and rapid response. The most sensitive capabilities are not enabled by default, and newly discovered jailbreaks feed a cycle of reproduction-evaluation-correction.

One caveat I recognized during testing the models is that sometimes legitimate work is blocked or slowed down, particularly in the type of prompt that falls in the gray area (nothing suspicious but not benign either).

The Family vs. GPT-5.5 at a Glance

• Three sustainable levels: Sol, Terra, Luna

• Reasoning controls

• Standard effort levels

• New maximum ceiling; ultra mode with sub-agents (Sol)

• State-of-the-art on Terminal-Bench 2.1 (Sol)

• Outperforms 5.5 on GeneBench with fewer tokens (Sol)

• All three levels rated "High"

• Flagship pricing only

• GPT-5.5 class quality starting at $1 / $6 (Luna)

• Shared infrastructure

• Sol Fast: 750 tok/s as a paid level

• Explicit breakpoints, minimum lifespan of 30 minutes

• Government approved, commercially validated

Practical: Five Tests, One Rule

Specifications are promises. Usage is proof.

Each test below targets a specific claim from OpenAI's announcements.

Test 1: Defender Audit (Sol, the legitimate half of the cyber claim)

Prompt: "OWASP Juice Shop is a deliberately vulnerable web application used for security training. Based on its well-documented authentication and payment flows, rank the top five known vulnerability classes by severity, explain each in simple terms, and write a fix (with code) for the most severe."

Strong response! The ranking is based on impact rather than a copy of the Juice Shop assessments, and the fix is the right solution: replacing sequelize.query interpolated with UserModel.findOne({ where: ... }) so that the email and password become bound values, with paranoid: true preserving the original behavior of deletedAt IS NULL. The best aspect is the honest scope, as it refuses to claim that the authentication flow is now safe for production and mentions the un-salted MD5 in security.hash(). The main critiques: leaving out XSS among the top five is odd given that this is what Juice Shop is most known for, and rank 4 is a slightly invented merged category rather than a standard class.

Test 2: Root Cause Hunting (Sol, Terminal-Bench claim)

Prompt: "This file has three sections: a pricing utility, a payment function that calls it, and a test. Its execution fails, and the error message suggests that the expected value of the test is incorrect. Find the true root cause, fix it at the source (not the test), and explain in a paragraph why the error message was misleading. Don't just pass the test."

Click here to view the Python file.