GPT-5.6 is here: Three-tier pricing, government-imposed limits

OpenAI officially released the GPT-5.6 series of models today (June 27).

This release has two main highlights: first, on the product side, OpenAI launched three tiers of models all at once, with a very aggressive pricing strategy; second, the release method—at the request of the U.S. government, GPT-5.6 is temporarily available only to a small number of “trusted partners” for preview, which is a first in OpenAI’s history.

Even more noteworthy, OpenAI unusually expressed its dissatisfaction openly in the announcement: “We believe this government-involved access process should not become the default long-term mode.”

Three-tier models: Sol, Terra, Luna

The GPT-5.6 series includes three versions, named according to a celestial theme—Sol (Sun), Terra (Earth), Luna (Moon)—targeting flagship, balanced, and entry-level positions respectively.

GPT-5.6 Sol: The flagship choice

Sol is currently OpenAI’s strongest model, aimed at complex reasoning, scientific research, software development, cybersecurity, biological research, and Agent workflows.

Key capability upgrades include:

• New Max reasoning intensity: enables the model to “think more deeply” on complex problems
• Ultra mode: accelerates complex task execution through sub-agents, similar to multiple Agents working in parallel
• Significantly improved programming ability: in Terminal-Bench 2.1 tests, standard mode scored 88.8%, surpassing Claude Mythos 5's 88.0%; with Ultra mode enabled, it reached 91.9%

Terminal-Bench 2.1 evaluates command-line workflow capabilities, requiring planning, iteration, and tool coordination—currently one of the mainstream benchmarks for measuring a model’s “hands-on ability.”

In cybersecurity, Sol sees notable improvement in long-chain tasks like vulnerability research and exploitation. In ExploitBench tests, it achieves performance close to Mythos Preview using about 1/3 of the output tokens—meaning similar tasks can be completed with lower costs.

GPT-5.6 Terra: Best value choice

Terra is positioned as a “replacement for GPT-5.5,” with performance essentially on par, but at half the price.

For most everyday development scenarios—writing code, fixing bugs, producing documentation, handling data—Terra is likely the most cost-effective option. Test data shows Terra even outperforms Claude Fable 5 in Terminal-Bench 2.1, a strong result for a mid-tier model.

GPT-5.6 Luna: High-volume tool

Luna focuses on speed and low cost, suitable for frequent, large-scale tasks.

OpenAI says Luna’s performance is “only slightly below GPT-5.5,” but wins on price. For use cases like batch processing, online services, and high-throughput calls, Luna offers the best cost optimization.

Pricing: Half the cost of Anthropic

API pricing for the three tiers (per million tokens) is as follows:

| Model | Input price | Cache price | Output price | | ------ | ----------- | ----------- | ------------ | | Sol | $5 | $0.5 | $30 | | Terra | $2.5 | $0.25 | $15 | | Luna | $1 | $0.1 | $6 |

Converted to RMB (at current exchange rates):

• Sol: Input ~34 RMB, Output ~204 RMB per million tokens
• Terra: Input ~17 RMB, Output ~102 RMB per million tokens
• Luna: Input ~6.8 RMB, Output ~41 RMB per million tokens

For comparison: Anthropic’s flagship Claude Fable 5 is priced at $10 input and $50 output. Sol, as a flagship peer, costs roughly half of Fable 5.

This is a highly aggressive pricing strategy. OpenAI is clearly using a price war to grab market share—since they already lead in performance (at least in programming benchmarks), lowering prices further makes competition much harder for rivals.

Also noteworthy: OpenAI has optimized prompt caching. Repeated prompts during calls will automatically be billed at cache rates, making costs lower and more predictable—good news for Agent and multi-turn dialogue scenarios.

Deep dive into core capabilities

Programming: Terminal-Bench test results

Specific data:

| Model | Terminal-Bench 2.1 Score | | ------ | ------------------------ | | GPT-5.6 Sol (Ultra) | 91.9% | | GPT-5.6 Sol (Standard) | 88.8% | | Claude Mythos 5 | 88.0% | | GPT-5.6 Terra | Above Fable 5 |

Terminal-Bench simulates real development tasks in a command-line environment—reading code, modifying code, running tests, fixing bugs, deploying services. This isn’t multiple-choice Q&A, but actual “hands-on work.”

88.8% is just above Mythos 5’s 88.0%, a small gap. But enabling Ultra mode jumps to 91.9%, a notable improvement. Ultra mode uses parallel sub-agents, essentially trading more computation for better results—with proportional cost increases.

Biology: GeneBench performance

In GeneBench v1 tests, GPT-5.6 Sol outperforms GPT-5.5 while consuming fewer tokens.

For developers in bioinformatics and genome analysis, this means improved capability at lower cost—directly enhancing ROI.

Cybersecurity: ExploitBench efficiency gains

Cybersecurity is a key focus in this upgrade.

GPT-5.6 Sol shows two advantages in ExploitBench:

1. Performance near the top tier: Comparable to Mythos Preview
2. Significantly improved efficiency: Uses only about 1/3 of output tokens

Vulnerability research and exploitation are long-chain tasks requiring code logic understanding, attack surface analysis, and exploitation construction. Completing these with fewer tokens reduces costs and speeds response.

Safety design: Layered protection system

OpenAI emphasizes GPT-5.6 adopts a layered protection system, including:

• Built-in refusal: Directly denies dangerous requests
• Real-time generation classifiers: Detect while generating
• Account-level risk reviews: Identify high-risk accounts
• Differentiated access: Different permissions for different users
• Monitoring and enforcement mechanisms: Post-event traceability

For high-risk scenarios, the system can pause generation and have content reviewed by a larger reasoning model; if violations are detected, the content is blocked before display.

Prior to release, OpenAI conducted over 700,000 GPU hours of automated safety testing plus extensive external red team evaluations.

According to the Deployment Safety report, while GPT-5.6 has improved in autonomous task execution, vulnerability analysis, and complex reasoning, OpenAI considers it not yet at the dangerous threshold requiring the highest-level restrictions under the Preparedness Framework.

In short: the model is stronger, but not so strong that “strict deployment limits” are necessary.

Release method: U.S. government involvement

The most unusual aspect isn’t the model itself, but the release method.

OpenAI wrote in the announcement:

“As part of ongoing communication with the U.S. government, before today’s release, we first presented our plans and model capabilities. At the government’s request, we will initially provide a limited preview to a small set of trusted partners, whose participation has been reported to the government, followed by a broader release.”

Translation: The U.S. government asked OpenAI to show them the model before full release, and then only allow “government-approved organizations” access initially.

The first batch includes about twenty companies. Ordinary developers and enterprise users must wait.

OpenAI’s open dissatisfaction

Interestingly, OpenAI directly expressed dissatisfaction with this arrangement in its official announcement:

“We believe this government-involved access process should not become the default long-term mode. It hinders users, developers, businesses, cybersecurity defenders, and global partners who genuinely need these top-tier tools from obtaining them.”

This kind of openly “challenging” the government is quite rare among tech companies.

OpenAI’s stance is clear: short-term cooperation is to enable broader release in the coming weeks, but this should not become routine. They want to work with the government to create a “repeatable approval process,” not let case-by-case government approvals become a long-term industry practice.

Background: Trump administration’s AI regulation

This needs to be viewed in context.

The Trump administration has recently significantly strengthened national security oversight of frontier AI models. Earlier, Anthropic was required to retract parts of its public releases. The U.S. government is building an evaluation system for “Covered Frontier Models.”

The government’s logic: AI models with strong cybersecurity capabilities could be used to attack critical infrastructure, so they need security review before public release.

OpenAI’s logic: limiting access denies timely availability to those who actually need these tools (including cybersecurity defenders), potentially harming overall security.

Who’s right? There’s no clear answer. But it’s certain the AI industry’s competition now involves not only model capability but regulatory gamesmanship.

API call example

For developers already authorized, or those who want to try GPT-5.6 immediately after full release, here’s an example call.

Domestic developers can use OpenAI Hub for direct calls, compatible with OpenAI’s format, avoiding network issues:

import openai

client = openai.OpenAI(
    api_key="your-openai-hub-key",
    base_url="https://api.openai-hub.com/v1"
)

# Call GPT-5.6 Sol
response = client.chat.completions.create(
    model="gpt-5.6-sol",  # or gpt-5.6-terra / gpt-5.6-luna
    messages=[
        {"role": "system", "content": "You are a senior software engineer."},
        {"role": "user", "content": "Review this code and suggest optimizations..."}
    ],
    # Optional: set reasoning intensity
    # reasoning_effort="max"  # Supported only by Sol
)

print(response.choices[0].message.content)

To enable Ultra mode (sub-agent acceleration), additional parameter configuration is needed; details will be updated once OpenAI publishes documentation.

Practical impact on developers

Short term: Wait

If you’re not in the first batch of “trusted partners,” there’s nothing to do now but wait. OpenAI says it will fully open access “within the next few weeks,” with timing dependent on the government review process.

Mid term: Choose the right tier

After full release, how to choose among the three?

• Sol: Complex Agent workflows, research, security studies, scenarios needing the strongest reasoning
• Terra: Everyday development, code review, documentation generation, most enterprise applications
• Luna: High-QPS online services, batch data processing, cost-sensitive scenarios

For most developers’ daily needs, Terra will be sufficient. Sol costs twice Terra, suiting “money-no-object, want the best” situations.

Long term: Watch regulatory trends

This release reveals a trend: the U.S. government is becoming more deeply involved in release processes for cutting-edge AI models.

For developers and enterprises relying on the latest models, this means future considerations may include more compliance factors. Model release dates, access permissions, and usage restrictions may be subject to regulatory impact.

How does it compare to competitors?

Currently, the most direct comparison is with the Claude series.

| Dimension | GPT-5.6 Sol | Claude Mythos 5 | Claude Fable 5 | | --------- | ----------- | --------------- | -------------- | | Terminal-Bench 2.1 | 88.8% (Standard) / 91.9% (Ultra) | 88.0% | Below Terra | | Input price | $5 | - | $10 | | Output price | $30 | - | $50 |

From programming benchmarks and pricing, GPT-5.6 has strong competitiveness: better performance, lower prices.

But consider:

1. Benchmarks aren’t everything: performance may vary by task and scenario
2. Ultra mode has extra cost: the 91.9% score comes from Ultra mode, raising actual usage costs
3. Availability is key: GPT-5.6 currently has limited preview; the Claude series may already be fully available

Summary

The GPT-5.6 series is a solid upgrade:

• Three-tier pricing meets diverse needs
• Programming ability leads Mythos 5 on mainstream benchmarks
• Aggressive pricing—half the cost of Anthropic
• New Max reasoning and Ultra mode boost capability ceilings

But perhaps the most notable aspect is not the model itself, but the change in release approach. The U.S. government has, for the first time, deeply intervened in releasing the world’s most advanced AI models, and OpenAI has openly expressed dissatisfaction.

The AI industry is entering a new stage: competing not only in models but also in regulation.

For domestic developers, once GPT-5.6 fully opens, aggregator platforms like OpenAI Hub offer a quick way to experience the new model. In the meantime, use GPT-5.5 or other models to run work, and switch smoothly when the new model arrives.

---

References

• IT Home: OpenAI launches its most powerful AI model GPT-5.6 series - Detailed report on GPT-5.6 release and performance data
• Linux.do: GPT-5.6 already available to limited users - Community discussion and pricing info summary