OpenAI Releases GPT-5.6 to the Public After 13 Days of U.S. Government Review

From Restricted Preview to General Availability

OpenAI released the three variants of GPT-5.6 to any API customer, including ChatGPT and Codex, on Thursday, July 9. Sol, Terra, and Luna had been under coordinated review by the U.S. government since June 26, when the model was presented to around 20 partner organizations. The security review, related to the model's offensive capabilities in cybersecurity, was completed ahead of the initially estimated schedule. Sam Altman confirmed the timeline the night before; GitHub Copilot enabled support for all three variants on the same day.

The critical detail for corporate buyers lies in the pricing table. Sol, the top tier, costs $5 per million input tokens and $30 per million output tokens. Terra is priced at $2.50 and $15. Luna, cost-optimized, costs $1 and $6. All three variants share a context window of 1.05 million tokens in the API and a maximum output of 128,000 tokens. OpenAI maintains a 90% discount on cached reads; writing is now charged at 1.25 times the rate of uncached input.

The Middle Tier Changes the Economics of Adoption

Terra is the piece that reorders the calculation for platform teams. OpenAI positions it as equivalent to GPT-5.5 in quality and effectively sells it at half the price. For operations running agent pipelines with millions of calls per day, the unit cost determines whether a project survives within the FinOps pipeline.

According to the numbers released by OpenAI itself, Sol achieves 92.2% on BrowseComp and 62.6% on OSWorld 2.0. According to the technical material, the result on OSWorld outperforms Claude Opus 4.8 while using 85% fewer output tokens. The important metric shifts to quality per unit spent, which is relevant for teams operating agents in large-scale RPA.

Cybersecurity Capabilities Were the Bottleneck for Release

It was performance in cybersecurity that prompted the U.S. government's review process. In ExploitBench2, which measures the path from vulnerable code to arbitrary execution, Sol reaches 73.5% compared to 47.9% for GPT-5.5 under comparable budgets. In ExploitGym3, which asks agents to produce exploits based on real vulnerabilities, the pass rate nearly doubles within a two-hour window, from 15.1% to 24.9%; over six hours, it reaches 33.7%. In SEC-Bench Pro, a proof of concept in complex software, it increases from 45.8% to 71.2%.

For the CISO, the result has two sides. On the operational side, automated red teaming transitions from demonstration to a viable product. On the regulatory side, models with such a significant leap in offensive capability will weigh heavily in the evaluation expected by the EU AI Act, which begins its oversight window for general-purpose models on August 2.

GPT-Live Enters the Same Window

OpenAI also introduced GPT-Live, a new generation of voice models that can listen and speak simultaneously. The company targets contact centers and persistent conversational assistants, situations where full-turn latency diminishes perceived quality. No pricing was disclosed in the initial announcement.

What Changes for Enterprise Buyers in Three Geographies

In the U.S., the enterprise Copilot was the first major enabled channel, meaning that engineering teams from Fortune 500 companies are evaluating Terra against GPT-5.5 in the same July sprint. In Germany, the BSI has already indicated that audits of frontier models in corporate use will fall within its 2027 scope, and the number from ExploitBench2 will return to the technical notes of those reviews. In India, the hub for TCS, Infosys, Wipro, and Cognizant, Luna at $1 for input lowers the floor for what providers can charge for outsourced coding tasks: if a client requests legacy migration and Luna resolves 60% of the automated backlog, the offshoring markup shrinks.

Anthropic, whose Fable 5 was back on sale on July 1 after being suspended due to export controls, has not yet responded with a public price cut. The next move belongs to Dario Amodei.